0002006003

EURASIP Journal on Wireless Communications and Networking
Radio Resource Management

in 3G+ Systems
Guest Editors: Alagan Anpalagan, Rath Vannithamby,

Weihua Zhuang, and Sonia Aïssa
EURASIP Journal on
Wireless Communications and Networking

in 3G+ Systems
EURASIP Journal on

in 3G+ Systems
Guest Editors: Alagan Anpalagan, Rath Vannithamby,
Weihua Zhuang, and Sonia Aissa
Copyright © 2006 Hindawi Publishing Corporation. All rights reserved.
This is a special issue published in volume 2006 of “EURASIP Journal on Wireless Communications and Networking.” All articles are
open access articles distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and
reproduction in any medium, provided the original work is properly cited.
Editor-in-Chief
Phillip Regalia, Institut National des Telecommunications, France
Associate Editors
Thushara Abhayapala, Australia Fary Z. Ghassemlooy, UK Marc Moonen, Belgium
Farid Ahmed, USA Christian Hartmann, Germany Eric Moulines, France
Alagan Anpalagan, Canada Stefan Kaiser, Germany Sayandev Mukherjee, USA
N. Arumugam, Singapore George K. Karagiannidis, Greece Kameswara Rao Namuduri, USA
Anthony Boucouvalas, UK Hyung-Myung Kim, Korea A. Pandharipande, The Netherlands
Lin Cai, Canada C. C. Ko, Singapore Athina Petropulu, USA
Biao Chen, USA Visa Koivunen, Finland H. Vincent Poor, USA
Pascal Chevalier, France Richard Kozick, USA Brian Sadler, USA
Chia-Chin Chong, USA Bhaskar Krishnamachari, USA Ivan Stojmenovic, Canada
Huaiyu Dai, USA Sangarapillai Lambotharan, UK Lee Swindlehurst, USA
Soura Dasgupta, USA Vincent Lau, Hong Kong Sergios Theodoridis, Greece
Petar Djuric, USA David I. Laurenson, UK Lang Tong, USA
Mischa Dohler, France Tho Le-Ngoc, Canada Luc Vandendorpe, Belgium
Abraham O. Fapojuwo, Canada Tongtong Li, USA Yang Xiao, USA
Michael Gastpar, USA Wei Li, USA Xueshi Yang, USA
Alex Gershman, Germany Yonghui Li, Australia Lawrence Yeung, Hong Kong
Wolfgang Gerstacker, Germany Zhiqiang Liu, USA Dongmei Zhao, Canada
David Gesbert, France Stephen McLaughlin, UK Weihua Zhuang, Canada
Contents
Radio Resource Management in 3G+ Systems, Alagan Anpalagan, Rath Vannithamby,
Weihua Zhuang, and Sonia Aïssa
Volume 2006 (2006), Article ID 70210, 3 pages
Space-Time Water-Filling for Composite MIMO Fading Channels, Zukang Shen, Robert W. Heath Jr.,
Jeffrey G. Andrews, and Brian L. Evans
Capacity Planning for Group-Mobility Users in OFDMA Wireless Networks, Ki-Dong Lee
and Victor C. M. Leung
Joint Downlink Power Control and Multicode Receivers for Downlink Transmissions in High Speed
UMTS, Bessem Sayadi, Stefan Ataman, and Inbar Fijalkow
Adaptive Downlink Resource Allocation Strategies for Real-Time Data Services in OFDM Cellular
Systems, Navid Damji and Tho Le-Ngoc
Opportunistic Nonorthogonal Packet Scheduling in Fixed Broadband Wireless Access Networks,

Mahmudur Rahman, Halim Yanikomeroglu, Mohamed H. Ahmed, and Samy Mahmoud
Rate-Optimal Multiuser Scheduling with Reduced Feedback Load and Analysis of Delay Effects,
Vegard Hassel, Mohamed-Slim Alouini, Geir E. Øien, and David Gesbert
Adaptive Rate-Scheduling with Reactive Delay Control for Next Generation CDMA Wireless Mobile
Systems, Oliver Yu, Emir Saric, and Anfei Li
Effective Radio Resource Management for Multimedia Broadcast/Multicast Services in UMTS

Networks, Nuno Souto, Armando Soares, Patricia Eusébio, Américo Correia, and João C. Silva
Impact of Video Coding on Delay and Jitter in 3G Wireless Video Multicast Services, Kostas E. Psannis
and Yutaka Ishibashi
Cross-Layer Quality-of-Service Analysis and Call Admission Control in the Uplink of CDMA Cellular
Networks, Chun Nie, Yong Huat Chew, and David Tung Chong Wong
Cross-Layer Design and Analysis of Downlink Communications in Cellular CDMA Systems,

Jin Yuan Sun, Lian Zhao, and Alagan Anpalagan
On Cross-Layer Design for Streaming Video Delivery in Multiuser Wireless Environments,
Lai-U Choi, Wolfgang Kellerer, and Eckehard Steinbach
Multiservice Vertical Handoff Decision Algorithms, Fang Zhu and Janise McNair
Hindawi Publishing Corporation
Volume 2006, Article ID 70210, Pages 1–3
DOI 10.1155/WCN/2006/70210
Editorial
Radio Resource Management in 3G+ Systems
Alagan Anpalagan,1 Rath Vannithamby,2 Weihua Zhuang,3 and Sonia Aı̈ssa4

1 Department of Electrical and Computer Engineering, Ryerson University, Toronto, ON, Canada, M5B 2K3
2 Intel
Corporation, 2111 NE 25th Avenue, Hillsboro, OR 97229, USA
3 Department of Electrical and Computer Engineering, University of Waterloo, 200 University Avenue West, Waterloo,
ON, Canada N2L 3G1

4 INRS-EMT, Université du Québec, Place Bonaventure, 800 de la Gauchetiere Ouest, Suite 6900, Montreal, QC, Canada H5A 1K6
Received 9 July 2006; Accepted 9 July 2006

Copyright © 2006 Alagan Anpalagan et al. This is an open access article distributed under the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly
cited.
The 3G+ wireless systems can be characterized by aggre- In the first paper “Space-time water-filling for composite
gate bit rates in the range of Mbps, quality-of-service (QoS) MIMO fading channels,” Shen et al. analyze the ergodic ca-
support for interactive multimedia services, global mobility, pacity and outage probability of the MIMO fading channel.
service portability, enhanced ubiquity, and larger user ca- The above capacity and probability with space-time water-
pacity and coverage. All digital entirely packet-switched ra- filling are evaluated through numerical integration which is
dio networks involving hybrid networking and access tech- simplified by the approximation of the eigenvalue distribu-
nologies are envisioned in 3G+ systems. In such systems, tion of the composite MIMO fading channel. The authors
radio resource management (RRM) plays a major role in compare the performance of space-time water-filling with
the provision of QoS and efficient utilization of scarce ra- that of spatial water-filling and show that the former outper-
dio resources. With the required support for multimedia ser- forms the latter in terms of capacity per antenna in the pres-
vices to multiple users over diverse wireless networks and the ence of large shadowing effects at lower-SNR regions, how-
ever-increasing demand for high-quality wireless services, ever with the higher outage probability.
the need for effective and efficient RRM techniques becomes The second paper “Capacity planning for group-mobility
more important than ever. The addition of efficient packet users in OFDMA wireless networks” by Lee and Leung
data channels in both forward and reverse directions and proposes three admission capacity planning methods for
QoS support in 3G standards lead to a more flexible network, OFDMA cellular networks in which a significant fraction of
but at the same time increase the complexity of determin- users experience group-mobility. This is achieved first by de-
ing the optimal allocation of resources especially on the ra- riving the outage ratio and excess capacity ratio based on
dio interface. This special issue is devoted to addressing the the average channel gains, and then solving the optimiza-
urgent and important need for efficient RRM techniques in tion problems by maximizing the reduction of the outage ra-
the evolving next-generation wireless systems. tio, the excess capacity ratio, and the convex combination of
This special issue consists of thirteen papers that have both.
been selected following an extensive review process. Of those In the third paper “Joint downlink power control and
papers, two papers are on capacity prediction and outage multicode receivers for downlink transmissions in high speed
analysis, two are on downlink power minimization algo- UMTS,” Sayadi et al. investigate how to combine downlink
rithms, three are on rate scheduling based on real-time chan- power control and joint multicode detection for an HSDPA
nel conditions, two are on multicast multimedia service de- link. The authors present an iterative algorithm that controls
livery, three are on cross-layer design, and one is on vertical both the transmitted code powers and the joint multicode
handoff mechanism. All of the papers are within the frame- receiver filter coefficients at the base station for a high-speed
work of the radio resource management and are summarized multicode user to decrease intercode interference and to in-
as follows. crease the system capacity.
2 EURASIP Journal on Wireless Communications and Networking
The fourth paper “Adaptive downlink resource allocation QoS in supporting multimedia service. This paper proposes
strategies for real-time data services in OFDM cellular sys- a scheme that dynamically selects the video frame sequence
tems” by Damji and Le-Ngoc presents a framework to ana- based on storing multiple differently encoded versions of the
lyze the outage probability of different transmission band- video stream at the server in order to improve the delay and
widths and modulation schemes in fading and shadowing jitter seen by the user.
environments in cellular OFDM networks. It is shown that In the tenth paper “Cross-layer quality-of-service analy-
the power minimization scheme outperforms the proposed sis and call admission control in the uplink of CDMA cellular
bandwidth-constrained power minimization scheme in low networks,” Nie et al. present an analytical formulation for the
shadowing environment; however, in severe shadowing en- QoS performances of the four UMTS traffic classes jointly
vironment with both frequency-selective and flat fading, the at both the data link and network layers, study the effect of
proposed scheme significantly outperforms the power mini- lengthening the ON periods of the NRT services under Go-
mization scheme. Back-N (GBN) automatic retransmission request, and pro-
In the fifth paper “Opportunistic nonorthogonal packet pose a QoS-based call admission control (CAC) scheme to
scheduling in fixed broadband wireless-access networks,” achieve the maximum system capacity.
Rahman et al. consider packet scheduling as a means of in- The eleventh paper “Cross-layer design and analysis of
terference management in the downlinks of fixed broadband downlink communications in cellular CDMA systems” by
wireless-access networks. This is achieved by forming inter- Sun et al. focuses on cross-layer analysis and design for the
ferer groups of base stations and allowing more than one downlink of a cellular CDMA network with voice and data
base station to transmit packets at a time. To this end, the communications. The authors propose adaptive scheduling
authors propose a nonorthogonal transmission scheme that for link layer, priority-based handoff strategy for network
requires SINR estimation at the scheduler. The performance admission control, and an algorithm for the avoidance of
in terms of spectral efficiency, mean packet delay, and packet TCP spurious timeouts at the transport layer. They demon-
dropping rate is compared against the orthogonal scheme to strate that the system performance in terms of the capacity,
show the superior performance of the proposed nonorthog- throughput, dropping probability, outage, power efficiency,
onal scheme. delay, and fairness can be enhanced by jointly considering
In the sixth paper “Rate-optimal multiuser scheduling the interactions across layers.
with reduced feedback load and analysis of delay effects,” In the twelfth paper “On cross-layer design for stream-
Hassel et al. propose a feedback algorithm that always col- ing video delivery in multiuser wireless environments,” Choi
lects feedback from the user with the best channel conditions et al. discuss a cross-layer design mechanism that interacts
so that to reduce the load compared to full feedback. Closed- across application layer, data link layer, and physical layer
form expression for the CNR threshold that minimizes the to optimize the quality of wireless streaming video applica-
feedback load for this algorithm is derived and the impacts tion. Cross-layer optimization is realized via three concepts,
of scheduling delay and outdated channel estimates are ana- namely, parameter abstraction, cross-layer optimization, and
lytically and numerically evaluated. decision distribution. The paper also analyzes the perfor-
The seventh paper “Adaptive rate scheduling with reac- mance such as peak signal-to-noise ratio with and without
tive delay control for next generation CDMA wireless mo- joint optimization.
bile systems” by Yu et al. proposes an adaptive rate sched- The final paper “Multiservice vertical handoff decision
uler called feedback-enhanced target-tracking weighted fair algorithms” by Zhu and McNair addresses vertical handoff
queuing (FT-WFQ) rate scheduler that mitigates for the ar- for a mobile node to handoff between different types of net-
rival rate estimation errors and delay-bandwidth coupling. works. Several optimizations are proposed for the execution
Analytical and simulation results indicate that the FT-WFQ of vertical handoff decision algorithms, with the goal of pro-
scheduler can significantly reduce degradations caused by ar- viding the QoS experienced by each user. The optimizations
rival rate estimation errors and can minimize delay degrada- incorporate a network elimination feature to reduce the de-
tions during nonstationary loading conditions. lay and processing required in the evaluation of the cost func-
In the eighth paper “Effective radio resource manage- tion, and a multinetwork optimization is introduced to im-
ment for multimedia broadcast/multicast services in UMTS prove users satisfaction for mobile terminals with multiple
networks,” Souto et al. discuss a mechanism for multimedia active sessions.
broadcast and multicast delivery of layered video based on
the user location. Using this scheme, the users close to the ACKNOWLEDGMENTS
base station can enjoy richer video quality compared to when
layered video coding is not used. This paper also proposes a Many people deserve our gratitude for helping us to put to-
nonuniform constellation mechanism based on channel con- gether this special issue within a relatively short period of
dition. The capacity and coverage are improved using these time. First of all, we wish to thank the authors who submitted
advanced schemes. their valuable and timely contributions to this special issue.
The ninth paper “Impact of video coding on delay and Certainly, the many submissions resulted in high-quality pa-
jitter in 3G wireless video multicast services” by Psannis and pers being published in the special issue. Our sincere appre-
Ishibashi addresses the issues due to mobility in multime- ciation goes to many dedicated and hardworking reviewers
dia multicast services. It is important to provide the required without whom our decision making would have been very
Alagan Anpalagan et al. 3
difficult. Last but not the least, thanks to the Editor-in-Chief Weihua Zhuang received the B.S. and M.S.
degrees from Dalian Maritime University,
and the Editorial Board of EURASIP JWCN for giving us
China, and the Ph.D. degree from the Uni-
this opportunity and the support to make this special issue versity of New Brunswick, Canada, all in
a reality. electrical engineering. Since October 1993,
she has been with the Department of Elec-
trical and Computer Engineering, Univer-
Alagan Anpalagan sity of Waterloo, Canada, where she is a Pro-
Rath Vannithamby fessor. She is a coauthor of the textbook
Weihua Zhuang
(Prentice Hall, 2003). Her current research interests include multi-
Sonia Aı̈ssa media wireless communications, wireless networks, and radio po-
sitioning. She is a licensed Professional Engineer in the Province
of Ontario, Canada. She received the Distinguished Performance
Award in 2006 from the Faculty of Engineering, University of Wa-
terloo, and the Outstanding Performance Award in 2005 from the
Alagan Anpalagan received the B.A.S., University of Waterloo, for outstanding achievements in teaching,
M.A.S., and Ph.D. degrees in electrical en- research, and service, and the Premier’s Research Excellence Award
gineering from the University of Toronto, (PREA) in 2001 from the Ontario Government for demonstrated
Canada, in 1995, 1997, and 2001, respec- excellence of scientific and academic contributions. She is an Ed-
tively. Since August 2001, he has been with itor/Associate Editor of IEEE Transactions on Wireless Commu-
the Ryerson University, Toronto, Canada, nications, IEEE Transactions on Vehicular Technology, EURASIP
where he cofounded WINCORE Labora- Journal on Wireless Communications and Networking, and Inter-
tory in 2002 and leads the WAN (Wireless national Journal of Sensor Networks.
Access and Networking) R&D Group. Cur-
rently, he is an Associate Professor and Pro- Sonia Aı̈ssa received her Ph.D. degree from
gram Director for graduate studies. His research interests are in McGill University, Canada, in 1998. She is
general wireless communication, mobile networks, and system per- now an Associate Professor at INRS-EMT,
formance analysis; and in particular QoS-aware radio resource University of Québec, Montreal, Canada,
management, joint study of wireless physical/link layer character- where she holds the Québec Government
istics, cross-layer resource optimization, and wireless sensor net- FQRNT Fellowship “Strategic Program for
working. Prior to his academic career, he was a Technical Consul- Professors-Researchers,” and Adjunct Pro-
tant at Bell Mobility working on 1×RTT system deployment stud- fessor at Concordia University, Canada.
ies in 2001, and in 1997 he was with Nortel Networks working on From 1996 to 1997, she was a Visiting Re-
R&D projects in systems engineering. He currently serves as IEEE searcher at the Department of Electronics
Toronto Section Chair, previously he served as Chair, Communica- and Communications of Kyoto University, Japan. During that pe-
tions Chapter—IEEE Toronto Section (2004-2005) and Technical riod, she also conducted research at the Wireless Systems Labora-
Program Cochair, and IEEE Canadian Conference on Electrical and tories of NTT, Kanagawa, Japan. From 1998 to 2000, she was a Re-
Computer Engineering (2004). He also serves as an Associate Edi- search Associate at INRS-Telecommunications, Canada. From 2000
tor EURASIP Journal of Wireless Communications and Network- to 2002, she was a Principal Investigator in the major program of
ing. He is an IEEE Senior Member and a Registered Professional personal and mobile communications of the Canadian Institute for
Engineer in the province of Ontario, Canada. Telecommunications Research. In 2006, she was a Visiting Asso-
ciate Professor at the Graduate School of Informatics, Kyoto Uni-
Rath Vannithamby received his B.S., M.S., versity, Japan. Her research interest includes radio resource man-
and Ph.D. degrees in electrical and com- agement and cross-layer design for MIMO wireless networks. She
puter engineering from the University of serves as an Editor for the IEEE Transactions on Wireless Com-
Toronto, Ontario, Canada, in 1994, 1996, munications, and Associate Editor for the IEEE Communications
and 2001, respectively. He was a Research Magazine, and the IEEE Wireless Communications Magazine. She
Assistant in the Network Architecture Lab is the Chair of the Montreal Chapter IEEE Women In Engineer-
and Wireless Communications Lab at the ing Society, served as Technical Program Chair for the Wireless
University of Toronto from 1994 to 1996 Communications Symposium of IEEE ICC ’2006, and is acting as
and from 1996 to 2000, respectively. He was PHY/MAC Program Chair for the IEEE WCNC ’2007.
also a Teaching Assistant at the University of
Toronto from 1994 to 2000. He is currently a Research Scientist at
Intel Corporation, Hillsboro, Oregon, USA and leads a group re-
sponsible for the MAC and signaling layer standardization of 3G
systems. Prior to joining Intel, he was a Staff Engineer at Eric-
sson Inc., San Diego, California, USA. He is a Member of IEEE
and IEEE/TCPC. He has published over 15 papers, and has over 30
patents pending. He has served on technical program committee
for major wireless communication conferences. His research inter-
ests are in the area of radio resource management techniques and
MAC and signaling layer protocols for high-speed wireless-access
networks using OFDMA/CDMA technologies including 3G and
IEEE 802.16.
DOI 10.1155/WCN/2006/16281
Space-Time Water-Filling for Composite MIMO

Fading Channels
Zukang Shen, Robert W. Heath Jr., Jeffrey G. Andrews, and Brian L. Evans
Department of Electrical and Computer Engineering, University of Texas at Austin, Austin, TX 78712, USA
Received 1 September 2005; Revised 14 February 2006; Accepted 13 March 2006

We analyze the ergodic capacity and channel outage probability for a composite MIMO channel model, which includes both
fast fading and shadowing effects. The ergodic capacity and exact channel outage probability with space-time water-filling can
be evaluated through numerical integrations, which can be further simplified by using approximated empirical eigenvalue and
maximal eigenvalue distribution of MIMO fading channels. We also compare the performance of space-time water-filling with
spatial water-filling. For MIMO channels with small shadowing effects, spatial water-filling performs very close to space-time
water-filling in terms of ergodic capacity. For MIMO channels with large shadowing effects, however, space-time water-filling
achieves significantly higher capacity per antenna than spatial water-filling at low to moderate SNR regimes, but with a much
higher channel outage probability. We show that the analytical capacity and outage probability results agree very well with those
obtained from Monte Carlo simulations.
Copyright © 2006 Zukang Shen et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
1. INTRODUCTION capacity in MIMO Rayleigh fading channels with shadowing

effects has not been evaluated, and the exact channel outage
Multiple-input multiple-output (MIMO) communication probability calculation has not been discussed. Furthermore,
systems exploit the degrees of freedom introduced by mul- while [1–4] have studied either spatial or space-time water-
tiple transmit and receive antennas to offer high spectral filling, the capacity gain of space-time water-filling over spa-
efficiency. In narrowband channels, when channel state in- tial water-filling has not been quantified.
formation is available at the transmitter and instantaneous
adaptation is possible, the capacity achieving distribution is In this paper, we perform space-time water-filling for a
found by using the well-known water-filling algorithm [1, 2]. mixed MIMO channel model that includes both Rayleigh
With only average power constraints, a two-dimensional fading and shadowing effects. We show that the ergodic ca-
water-filling in both the temporal and spatial domains pacity and the exact channel outage probability can both be
has recently been shown to be optimal [3, 4]. By study- evaluated through numerical integrations. Hence, the time-
ing the empirical distribution of the eigenvalues of Gaus- consuming Monte Carlo simulations, that is, generating a
sian random matrices [1], two-dimensional water-filling for large number of channel realizations and then performing
Rayleigh MIMO channels [3, 4] can be transformed into averaging, can be avoided. We also show that for Rayleigh
one-dimensional water-filling for a time-varying SISO chan- channels without shadowing, space-time water-filling gains
nel [5]. With the freedom to optimize the transmit power little in capacity over spatial water-filling. For Rayleigh chan-
in both time and spatial domains, two-dimensional space- nels with shadowing, space-time water-filling achieves higher
time water-filling disables data transmission when all of spectral efficiency per antenna over spatial water-filling, with
the effective channel gains are not high enough to utilize a tradeoff of higher channel outage probability. In either
transmit power efficiently, thereby resulting in a larger er- case, space-time water-filling actually has lower computa-
godic capacity when compared to spatial-only water-filling. tional complexity than spatial water-filling.
In [3], a MIMO channel outage probability is defined to
quantify how often the transmission is blocked, and upper 2. SYSTEM MODEL
bounds in Rayleigh fading channels on this outage proba-
bility have been developed. Although the ergodic capacity in A point-to-point MIMO system is shown in Figure 1. Let Nt
i.i.d. MIMO Rayleigh fading channels is well understood, the and Nr denote the number of transmit and receive antennas,
User data Space- Space- User data

time time
transmitter receiver
Figure 1: Point-to-point MIMO systems.
respectively. The symbolwise discrete-time input-output re- 3. SPATIAL AND SPACE-TIME WATER-FILLINGS
lationship of a narrowband point-to-point MIMO system
can be simplified as 3.1. Spatial water-filling
y = Hx + v, (1) The problem of spatial water-filling for MIMO Rayleigh fad-

ing channels was presented in [1]. Channel state informa-
where H is the Nr × Nt MIMO channel matrix, x is the Nt × 1 tion is assumed to be available at the transmitter and power
transmitted symbol vector, y is the Nr × 1 received symbol adaption is performed with a total power constraint for each
vector, and v is the Nr × 1 additive white Gaussian noise vec- channel realization. The capacity maximization problem can
tor, with variance E[vv† ] = σ 2 I, where (·)† denotes the op- be represented as
eration of matrix complex conjugate transpose.
1
In this paper, the MIMO channel H is modeled as max log
I + 2
HQH†

Q σ (3)
√
H = sHw , (2) subject to tr(Q) ≤ P,
where Hw is an Nr × Nt Rayleigh fast fading MIMO chan- where H is the MIMO channel, Q is the autocorrelation ma-
nel whose entries are i.i.d. complex Gaussian random vari- trix of the input vector x, defined as Q = E[xx† ], P is the
ables [1], and s is a scalar log-normal random variable, that instantaneous power limit, |A| denotes the determinant of
is, 10 log10 s ∼ N (0, ρ2 ), representing the shadowing ef- A, and tr(A) denotes the trace of matrix A.
fect. Notice that log-normal shadowing models the channel Notice that H† H can be diagonalized as H† H = U† ΛU,
power variation from objects on large spatial scales; hence, where U is a unitary matrix, Λ = diag{λ1 , . . . , λM }, and
the square root of s is used in (2). Further, shadowing can λ1 ≥ λ2 ≥ · · · ≥ λM ≥ 0. It is pointed out in [1] that
the optimization in (3) can be carried out over Q = UQU†
be modeled as a multiplicative factor to fast fading [6, 7].
Since shadowing occurs on large spatial scales, it is assumed and the capacity-achieving Q is a diagonal matrix. Let Q
=
that the shadowing value s equally effects all elements of diag{q1 , q2 , . . . , qM }, then the optimal value for qi is qi =
(σ 2 ,M)
Hw . Furthermore, s is assumed to be independent of Hw . (Γ0 − σ 2 /λi )+ , where σ 2 is the noise variance, a+ denotes
As the shadowing effect varies slower relative to fast fad- (σ 2 ,M)
max{0, a}, and Γ0 is solved to satisfy Mi=1 qi = P.
ing, the channel model discussed in this paper is suitable for
transmissions over a long time period. Throughout this pa-
per, we assume perfect channel state information is known 3.2. Space-time water-filling
at the transmitter. The MIMO channel capacity with imper- The problem of two-dimensional space-time water-filling
fect channel state information can be found in [8]. Further, can be formulated as
we consider MIMO systems with equal numbers of transmit
1
and receive antennas, that is, Nt = Nr = M, since express- †
max E log I + 2 HQH
ing the channel eigenvalue distribution is simpler than for Q σ (4)
unequal numbers of transmit and receive antennas [1]. The
subject to E tr(Q) ≤ P,
same technique discussed in this paper, however, can be ap-
plied to MIMO systems with unequal numbers of transmit where P is the average power constraint; H and Q have the
and receive antennas. same meaning as in (3), that is, Q = E[xx† ] is the covariance
Zukang Shen et al. 3
matrix of the transmitted signal for a particular channel re- Since 10 log10 s ∼ N (0, ρ2 ), by a simple change of vari-
alization H. Hence, Q is a function of H. The expectation in ables, the pdf of s can be written as
E[tr(Q)] is carried over all MIMO channel realizations. This
notation can be understood as the symbol rate is much faster
10 1 −(10 log10 s)2 /2ρ2
than the MIMO channel variation and Q is evaluated from r(s) = √ e . (9)
all symbols within one channel realization. ρ log 10 2π s
Notice that

M
Furthermore, s is independent of Hw , hence s is independent
p(λk )λk of t. The cdf of λ is
1 †
E log

I + 2 HQH

=E log 1 +
σ σ2 k=1
∞ λ/s

p(λ)λ F(λ) = r(s)g(t)dt ds. (10)

= ME log 1 + , 0 0
σ2
(5)
Differentiating F(λ) with respect to λ generates the pdf of λ:
where λk is the kth unordered eigenvalue of H† H, λ denotes ∞

any of them, and p(λ) denotes the power adaption as a func- 10 λ 1 −(10 log10 s)2 /2ρ2
f (λ) = √ g e ds. (11)
tion of λ. Hence, (4) can be rewritten as ρ log 10 2π 0 s s2

p(λ)λ With f (λ) available, the optimal cutoff value Γ0(σ
2 , M)
can be
max M log 1 + f (λ)dλ
p(λ) σ2 found by numerically solving
(6)
subject to M p(λ) f (λ)dλ = P,
∞
2 , M) σ2
M (σ 2 , M)
Γ0(σ − f (λ)dλ = P (12)
where f (λ) is the empirical eigenvalue probability density σ 2 /Γ0 λ
function. The problem in (6) is essentially the same as in [5].
2 , M)
The optimal power adaption is p(λ) = (Γ(σ 0 − σ 2 /λ)+ , and the ergodic capacity can be expressed as
(σ 2 , M)
where Γ0 is found numerically to satisfy the average
power constraint in (6). Notice that the power adaptation ∞ 2 , M)
is zero for the MIMO channel eigenvalue λ smaller than 1 Γ0(σ λ
2 E log †
I+ σ 2 HQH = M log f (λ)dλ.
σ 2 /Γ0(σ , M) , which means no transmission is allowed in this (σ 2 , M)
σ 2 /Γ0 σ2
MIMO eigenmode. (13)
2
To find Γ0(σ , M) , it is necessary to find f (λ) first. From (2),
H† H = sH†w Hw . Let {tk }M k=1 be the ordered eigenvalues for 4. CHANNEL OUTAGE PROBABILITY
H†w Hw , that is, t1 ≥ t2 ≥ · · · ≥ tM . Hence, λk = stk , where
λk is the kth largest eigenvalue of H† H. The ordered joint The capacity achieving power distribution from space-time
eigenvalue distribution of Gaussian random matrices H†w Hw water-filling blocks transmission when all eigenvalues of
has been given in [1, 9] as H† H are not high enough to utilize transmit power effi-
ciently. The channel outage probability defined in [3] is
2 equivalent to the probability that the largest eigenvalue of
gordered t1 , t2 , . . . , tM = KM e− i ti ti − t j , (7) 2
i< j
H† H is smaller than σ 2 /Γ0(σ , M) . Since the eigenvalues {λk }M
k=1
of H† H are in descending order, the channel outage proba-
bility can be expressed as
where KM is a normalizing factor.
In this paper, the empirical eigenvalue distribution for

H†w Hw is defined to be the probability density function for σ2
2
an eigenvalue t smaller than a certain threshold z. Telatar de- Pout σ , M = P λ1 ≤ 2 , M) . (14)
Γ0(σ
rived its pdf g(t) by integrating out all other eigenvalues in
the unordered joint eigenvalue distribution of Gaussian ran-
dom matrices [1] to obtain Although the channel outage probability is defined in [3],
only upper bounds in MIMO Rayleigh fading channels on
M −1
1 2 this outage probability are derived. In this paper, the exact
g(t) = L (t)e−t , (8) channel outage probability is expressed in terms of the max-
M i=0 i
imal eigenvalue distribution, denoted as fmax (λ1 ).
Recall that λ1 = st1 , where s is the shadowing random
where Lk (t) = (1/k!)et (dk /dt k )(e−t t k ). variable and t1 is the maximal eigenvalue of H†w Hw . The
2
distribution of t1 is denoted as gmax (t1 ) and can be obtained Table 1: Cutoff value Γ0(σ , M) for 2 × 2 MIMO fading channels. The
from (7) by integrating out tM , tM −1 , . . . , t2 , that is, average power constraint is P = 1. The exact empirical eigenvalue
2
t1 tM−2 tM−1 distribution [8] is used in finding Γ0(σ , M) .

gmax t1 = ··· KM e − i ti
SNR ρ=0 ρ=8 ρ = 16
0 0 0
2 (σ 2 , M) (σ 2 , M) (σ 2 , M)
× ti − t j dtM dtM −1 · · · dt2 . (1/σ 2 ) (dB) Γ0 P sim Γ0 P sim Γ0 P sim
i< j
−5 2.0935 0.9998 1.8233 1.0000 1.5254 1.0181
(15)
0 1.2907 0.9998 1.2774 1.0005 1.2098 1.0146
Mathematica’s built-in function Integrate can be used to per- 5 0.9075 0.9999 0.9526 0.9999 0.9894 1.0116
form the symbolic integration in (15). For example, when 10 0.7005 0.9999 0.7576 1.0001 0.8345 1.0098
M = 2, gmax (t1 ) = e−t1 (2 − 2t1 + t12 − 2e−t1 ). 15 0.5918 0.9999 0.6411 0.9999 0.7255 1.0086
With gmax (t1 ) available, the same procedure in (9)–(11)
20 0.5392 0.9998 0.5732 1.0000 0.6491 1.0078
can be used to calculate fmax (λ1 ), with t and g(t) replaced by
t1 and gmax (t1 ), respectively. The channel outage probability 25 0.5158 0.9999 0.5356 1.0001 0.5963 1.0071
becomes 30 0.5061 0.9999 0.5161 0.9999 0.5606 1.0068

Pout σ 2 , M
σ 2 /Γ(σ0 2 , M) outage probability analysis still requires a certain amount of

= fmax λ1 dλ1 computation. In [10], Wong showed that√the distribution of
0
the largest singular value of Hw , that is, t1 , can be well ap-
10
= √ proximated with a Nakagami-m distribution. In other words,
ρ log 10 2π gmax (t1 ) can be approximated with
σ 2 /Γ(σ0 2 , M) ∞
λ1 1 −(10 log10 s)2 /2ρ2
× gmax e ds dλ1 . mm
0 0 s s2 gmax t1 = t m−1 e−mt1 /Ω , (19)
(16) Γ(m)Ωm 1
where m and Ω are coefficients dependent on the MIMO sys-

5. APPROXIMATED CAPACITY AND CHANNEL tem size M; Γ(m) is the Gamma function, which is imple-
OUTAGE ANALYSIS mented in Mathematica as Gamma[m]. Wong also showed
the values of m and Ω for different transmit and receive an-
Even for medium-sized MIMO systems, for example, M = 4
tenna numbers, up to the 6 × 6 MIMO case [10]. For ex-
or 6, the calculation of the empirical eigenvalue distribution
ample, for M = 4, (m, Ω) = (12.5216, 9.7758); for M = 6,
g(t) in (8) for H†w Hw is computationally intensive, and the re-
(m, Ω) = (24.0821, 16.5881). Substituting (19) into (16), the
sultant g(t) is too complicated to be handled in closed form.
outage probability can be calculated as
Therefore, an approximation to g(t) will be utilized to sim-
2
plify the calculation of Γ0(σ , M) . An interesting property of 10mm
Gaussian random matrices is that the distribution of t/M has Pout (σ 2 , M) = √
Γ(m)Ωm ρ log 10 2π
a limit as the number of antennas increases [1]. Hence, σ 2 /Γ(σ0 2 , M) ∞ m−1
λ1
× e−mλ1 /sΩ
1 4 1 0 0 s
g(t) ≈ − , t ∈ (0, 4M) (17)
2π tM M 2 1
× 2 e−(10 log10 s) /2ρ ds dλ1 .
2 2
as M → ∞. Simulations show that this approximation holds s

(20)
well even for medium-sized MIMO systems, for example,
M = 4 or 6. With (17), for Rayleigh fading channel with
2 , M) 6. NUMERICAL RESULTS AND DISCUSSION
shadowing variance ρ, the cutoff value Γ(σ
0 can be found
by numerically solving In this section, the achievable spectral efficiencies per an-
10M tenna of the following three cases are compared by Monte
Carlo simulations: (1) space-time water-filling, (2) water-
(2π)(3/2) ρ log 10
filling in space only, and (3) equal power distribution. We
∞ ∞ 2 also compare the results from numerical integrations with
2 σ 4s 1
× Γ0(σ , M) − − (18)
those obtained from Monte Carlo simulations.
σ 2 /Γ0
(σ 2 , M)
λ/4M λ λM M 2
1 −(10 log10 s)2 /2ρ2 In all simulations, the Rayleigh MIMO channel Hw has
× e ds dλ = P. variance of 1/2 for both real and imaginary components. The
s2
shadowing effect has a log-normal distribution with standard
Although the lengthy calculation of g(t) can be avoided deviation of ρ [11]. For the pure Rayleigh fading channel, s
with the approximation in (17), the method in (15) to find is a constant of 1. For notational simplicity, we denote the
the maximal eigenvalue distribution gmax (t1 ) for channel pure Rayleigh fading case as ρ = 0. We also study the cases
6 100
5
10−1
Capacity (bps/Hz/antenna)
Outage probability
4
10−2
3 ρ = 16
10−3
2 ρ=8
10−4
1 ρ=0
0 10−5
−5 0 5 10 15 −5 0 5 10 15 20 25 30
SNR (dB) SNR (dB)
Space-time WF, numerical Spatial WF, simulated ρ = 16, simulated ρ = 8, numerical

Space-time WF, simulated Equal power, simulated ρ = 16, numerical ρ = 0, simulated
ρ = 8, simulated ρ = 0, numerical
Figure 2: Capacity of 2 × 2 MIMO fading channels. The variance Figure 3: Channel outage probability for 2 × 2 MIMO fading chan-
of the log-normal random variable is denoted by ρ. The numerical nels. The numerical results are obtained from (16) with Mathemat-
results are obtained from (13) with Mathematica 5.0. ica 5.0. The variance of the log-normal random variable is denoted
by ρ.
where ρ = 8 and 16. Table 1 shows the cutoff values for a 2 × 2 presents the channel outages evaluated from (16) with Math-
MIMO system with different SNRs and log-normal shadow- ematica 5.0, and the results again agree very well with those
ing variances. These cutoff values are obtained from the nu- obtained from Monte Carlo simulations.
2
merical method NIntegrate in Mathematica 5.0. The average Table 2 shows the cutoff values Γ0(σ , M) and P sim for 4 × 4
power constraint is P = 1. In Table 1, the columns P sim show and 6 × 6 MIMO systems. The cutoff values are evaluated
the average power obtained in Monte Carlo simulations. If with the approximation in (17). Even with the approximated
2
the cutoff value Γ0(σ , M) is calculated exactly, then P sim will empirical eigenvalue distribution, the cutoff values are still
equal P. Table 1 shows that for ρ = 0 and 8, the cutoff values very accurate, which is partially shown by the fact that P sim
are very accurate. For ρ = 16, P sim has 1-2% relative error has a relative error not exceeding 2.5% compared to P.
compared to P, which is primarily caused by the limited ac- Figure 4 shows the capacity per antenna for a 4×4 MIMO
2
curacy in the process of numerically finding Γ0(σ , M) for high system. The capacity per antenna for the 6 × 6 case is very
shadowing variances. close to the 4 × 4 case. From Figures 2 and 4, the capacity per
Figure 2 shows the capacity per antenna versus SNR antenna is insensitive to the number of antennas in the sys-
under different shadowing variances. For Rayleigh chan- tem. Numerical results from (13) are also shown in Figure 4.
nels without shadowing, spatial water-filling achieves al- Figure 5 shows the channel outage probability for the
most the same capacity as space-time water-filling. However, 4 × 4 and 6 × 6 MIMO systems, with shadowing variance
for Rayleigh channels with shadowing variance ρ = 8, the ρ = 8. The outage probability is evaluated through (20).
space-time water-filling algorithm achieves approximately For the same shadowing variance, the outage probabilities for
0.15 bps/Hz/antenna over spatial water-filling at low SNRs, the 4 × 4 and 6 × 6 MIMO systems are very close, since the
and has a 1.7 dB SNR gain over equal power distribution at shadowing variable equally effects all eigenvalues of H†w Hw
a spectral efficiency of 2 bps/Hz/antenna. For Rayleigh fad- and therefore dominates the channel outage probability.
ing with shadowing variance ρ = 16, space-time water-filling Figure 5 shows that even with the approximated maximal
achieves 0.3 bps/Hz/antenna over spatial water-filling. Notice eigenvalue distribution, the results from (20) still agree with
that compared to the pure Rayleigh fading case, the average the Monte Carlo simulations very well.
channel power is increased with the introduction of shadow- We also compare the main advantages and disadvan-
ing, but this does not affect the comparison between 2D and tages of space-time water-filling versus spatial water-filling
1D water-fillings. Further, Figure 2 shows that the numerical in Table 3. For space-time water-filling, only the cutoff
results evaluated from (13) with Mathematica 5.0 agree with threshold needs to be precomputed, while for spatial water-
the Monte Carlo results. filling, the optimal power distribution needs to be com-
Figure 3 shows the channel outage probability for a 2 × 2 puted for each channel realization to achieve capacity. On the
MIMO system. With the increase of the shadowing variance, other hand, the two-dimensional algorithm requires a priori
higher channel outage probability is observed. Figure 3 also knowledge of the channel eigenvalue distribution in order
2
Table 2: Cutoff value Γ0(σ , M) for 4 × 4 and 6 × 6 MIMO fading channels. The average power constraint is P = 1. The approximated empirical
2
eigenvalue distribution [8] is used in finding Γ0(σ , M) .
SNR M = 4, ρ = 0 M = 4, ρ = 8 M = 6, ρ = 0 M = 6, ρ = 8
2 2 2 2
(1/σ 2 ) (dB) Γ0(σ , M) P sim Γ0(σ , M) P sim Γ0(σ , M) P sim Γ0(σ , M) P sim
−5 1.0532 1.0010 0.9185 1.0019 0.7021 0.9996 0.6123 1.0015
0 0.6443 0.9988 0.6468 1.0036 0.4295 1.0001 0.4312 1.0016
5 0.4532 1.0050 0.4854 1.0057 0.3021 0.9999 0.3236 1.0023
10 0.3583 0.9994 0.3888 1.0087 0.2389 1.0029 0.2592 1.0038
15 0.3090 1.0070 0.3310 1.0124 0.2060 0.9998 0.2206 1.0053
20 0.2826 1.0204 0.2967 1.0157 0.1884 1.0068 0.1978 1.0082
25 0.2681 1.0243 0.2767 1.0177 0.1787 1.0142 0.1844 1.0101
30 0.2601 1.0208 0.2651 1.0169 0.1734 1.0155 0.1767 1.0111
Table 3: Comparison of space-time and spatial water-fillings.
Space-time water-filling Spatial water-filling

Computational complexity Low High
Channel eigenvalue distribution Required Not required
Ergodic capacity High Low
Outage probability High Low
Transmission mode Block transmission Continuous transmission
4.5 100
4
10−1
3.5
Capacity (bps/Hz/antenna)
10−2
Outage probability
2.5
ρ=8 10−3
2
1.5 10−4
1
10−5
0.5 ρ=0
0 10−6
−5 0 5 10 15 −5 0 5 10 15 20 25 30
SNR (dB) SNR (dB)
Space-time WF, numerical Spatial WF, simulated 4 × 4 MIMO, simulated 6 × 6 MIMO, simulated
Space-time WF, simulated Equal power, simulated 4 × 4 MIMO, numerical 6 × 6 MIMO, numerical
Figure 4: Capacity of 4 × 4 MIMO fading channels. The variance Figure 5: Channel outage probability for 4 × 4 and 6 × 6 MIMO
of the log-normal random variable is denoted by ρ. The numerical fading channels. The numerical results are obtained from (20) with
results are obtained from (13) with Mathematica 5.0. Mathematica 5.0. The variance of the log-normal random variable
is ρ = 8.
to calculate the optimal cutoff threshold. Furthermore, the outage and hence is similar to block transmission. For spatial
higher capacity achieved by two-dimensional water-filling water-filling, the transmission mode is continuous since for
comes with a larger channel outage probability. Since shad- every channel realization, the transmitter always has power to
owing changes much slower than fast fading, the transmis- transmit. Further, the capacity gap between space-time and
sion of space-time water-filling is subject to long periods of spatial water-filling depends on the distributions of the fast
fading and shadowing gains. An analytical expression for the [10] K.-K. Wong, “Performance analysis of single and multiuser
gap, however, is difficult to obtain. MIMO diversity channels using Nakagami-m distribution,”
IEEE Transactions on Wireless Communications, vol. 3, no. 4,
pp. 1043–1047, 2004.
7. CONCLUSION [11] G. L. Stüber, Principles of Mobile Communication, Kluwer Aca-
In this paper, the ergodic capacity and channel outage prob- demic, Dordrecht, The Netherlands, 2nd edition, 2001.
ability in a composite MIMO channel model with both
fast fading and shadowing have been analyzed. With the Zukang Shen received his B.S.E.E. de-
eigenvalue distribution of MIMO fading channels, both the gree from Tsinghua University in 2001, his
capacity and the channel outage probability have been eval- M.S.E.E. and Ph.D. degrees from The Uni-
uated through numerical integration, which avoids time- versity of Texas at Austin in 2003 and 2006,
consuming Monte Carlo simulations and provides more di- respectively. He is currently with Texas
rect insight into the system. Furthermore, approximations Instruments, Dallas, Texas. Dr. Shen was
to the empirical eigenvalue distribution and the maximal awarded the David Bruton Jr. Graduate Fel-
eigenvalue distribution can greatly simplify the capacity and lowship for the 2004–2005 academic year by
the Office of Graduate Studies at The Uni-
outage probability analysis. Numerical results illustrate that
versity of Texas at Austin. He also received
while the capacity difference is negligible for Rayleigh fad- UT Austin Texas Telecommunications Engineering Consortium
ing channels, space-time water-filling has an advantage when Fellowships for the 2001–2002 and 2003–2004 academic years. His
large-scale fading is taken into account. In all cases, it is sim- research interests include multicarrier communication systems, re-
pler to compute the solution for space-time water-filling be- source allocation in multiuser environments, MIMO channel ca-
cause it avoids the cutoff value calculation for each channel pacity analysis, digital signal processing, and information theory.
realization, but it requires knowledge of the channel distribu-
tion. The spectral efficiency gain of space-time water-filling Robert W. Heath Jr. received the B.S.
over spatial water-filling is also shown to be associated with a and M.S. degrees from the University of
Virginia, Charlottesville, Va, in 1996 and
higher channel outage probability. Hence, space-time water-
1997, respectively, and the Ph.D. degree
filling is more suitable for burst mode transmission when from Stanford University, Stanford, Calif,
the channel gain distribution has a heavy tail, and spatial in 2002, all in electrical engineering. From
water-filling is preferred for continuous transmission when 1998 to 2001, he was a Senior Engineer
the channel gain distribution is close to Rayleigh or is un- then Senior Consultant with Iospan Wire-
known. less Inc., San Jose, Calif, where he played
a key role in the design and implementa-
REFERENCES tion of the physical and link layers of the first commercial MIMO-
OFDM communication system. In 2003, he founded MIMO Wire-
[1] I. E. Telatar, “Capacity of multi-antenna Gaussian channels,” less Inc., consulting company dedicated to the advancement of
European Transactions on Telecommunications, vol. 10, no. 6, MIMO technology. Since January 2002, he has been with the De-
pp. 585–595, 1999. partment of Electrical and Computer Engineering at the University
[2] A. Goldsmith, S. A. Jafar, N. Jindal, and S. Vishwanath, “Ca- of Texas at Austin where he is currently an Assistant Professor and a
pacity limits of MIMO channels,” IEEE Journal on Selected Ar- Member of the Wireless Networking and Communications Group.
eas in Communications, vol. 21, no. 5, pp. 684–702, 2003. His research interests include several aspects of MIMO commu-
[3] S. K. Jayaweera and H. V. Poor, “Capacity of multiple-antenna nication such as antenna design, practical receiver architectures,
systems with both receiver and transmitter channel state in- limited feedback techniques, ad hoc networking, scheduling al-
formation,” IEEE Transactions on Information Theory, vol. 49, gorithms, and more recently 60 GHz communication. Dr. Heath
no. 10, pp. 2697–2709, 2003. serves as an Editor for the IEEE Transactions on Communication
[4] E. Biglieri, G. Caire, and G. Taricco, “Limiting performance of and an Associate Editor for the IEEE Transactions on Vehicular
block-fading channels with multiple antennas,” IEEE Trans- Technology.
actions on Information Theory, vol. 47, no. 4, pp. 1273–1289,
2001. Jeffrey G. Andrews is an Assistant Professor
[5] A. J. Goldsmith and P. P. Varaiya, “Capacity of fading channels in the Department of Electrical and Com-
with channel side information,” IEEE Transactions on Informa- puter Engineering at the University of Texas
tion Theory, vol. 43, no. 6, pp. 1986–1992, 1997. at Austin, and an Associate Director of the
[6] B. M. Hochwald, T. L. Marzetta, and V. Tarokh, “Multiple- Wireless Networking and Communications
antenna channel hardening and its implications for rate feed- Group (WNCG). He received the B.S. de-
back and scheduling,” IEEE Transactions on Information The- gree in engineering with high distinction
ory, vol. 50, no. 9, pp. 1893–1909, 2004. from Harvey Mudd College in 1995, and
[7] T. S. Rappaport, Wireless Communications: Principles and Prac- the M.S. and Ph.D. degrees in electrical en-
tice, Prentice Hall PTR, Upper Saddle River, NJ, USA, 2002. gineering from Stanford University in 1999
[8] T. Yoo and A. J. Goldsmith, “Capacity and power allocation and 2002, respectively. He developed code-division multiple-access
for fading MIMO channels with channel estimation error,” to (CDMA) systems as an engineer at Qualcomm from 1995 to 1997,
appear in IEEE Transactions on Information Theory. and has served as a frequent consultant on communication systems
[9] A. Edelman, Eigenvalue and condition numbers of random ma- to numerous corporations, startups, and government agencies, in-
trices, Ph.D. thesis, MIT, Cambridge, Mass, USA, May 1989. cluding Microsoft, Palm, Ricoh, ADC, and NASA. Dr. Andrews
serves as an Associate Editor for the IEEE Transactions on Wire-

less Communications. He also is actively involved in IEEE confer-
ences, serving on the organizing committee of the 2006 Communi-
cation Theory Workshop as well as regularly serving as a Member
of the technical program committees for ICC and Globecom. He
is a coauthor of the forthcoming book from Prentice-Hall, Under-
standing WiMAX: Fundamentals of Wireless Broadband Networks.
Brian L. Evans is the Mitchell Professor

of electrical and computer engineering at
the University of Texas at Austin in Austin,
Texas, USA. His B.S.E.E.C.S. (1987) degree
is from the Rose-Hulman Institute of Tech-
nology in Terre Haute, Indiana, USA, and
his M.S.E.E. (1988) and Ph.D.E.E. (1993)
degrees are from the Georgia Institute of
Technology in Atlanta, Georgia, USA. From
1993 to 1996, he was a Postdoctoral Re-
searcher at the University of California, Berkeley, in design au-
tomation for embedded digital systems. At UT Austin, his research
group develops signal quality bounds, optimal algorithms, low-
complexity algorithms and real-time embedded software of high-
quality image halftoning for desktop printers, smart image acqui-
sition for digital still cameras, high-bitrate equalizers for multicar-
rier ADSL receivers, and resource allocation for multiuser OFDM
basestations. Dr. Evans is the architect of the Signals and Systems
Pack for Mathematica. He received a 1997 US National Science
Foundation CAREER Award.
DOI 10.1155/WCN/2006/75820
Capacity Planning for Group-Mobility Users in OFDMA

Wireless Networks
Ki-Dong Lee and Victor C. M. Leung
Department of Electrical and Computer Engineering (ECE), University of British Columbia (UBC), Vancouver, BC, Canada V6T 1Z4
Received 11 October 2005; Revised 28 April 2006; Accepted 26 May 2006

Because of the random nature of user mobility, the channel gain of each user in a cellular network changes over time causing
the signal-to-interference ratio (SNR) of the user to fluctuate continuously. Ongoing connections may experience outage events
during periods of low SNR. As the outage ratio depends on the SNR statistics and the number of connections admitted in the
system, admission capacity planning needs to take into account the SNR fluctuations. In this paper, we propose new methods
for admission capacity planning in orthogonal frequency-division multiple-access (OFMDA) cellular networks which consider
the randomness of the channel gain in formulating the outage ratio and the excess capacity ratio. Admission capacity planning is
solved by three optimization problems that maximize the reduction of the outage ratio, the excess capacity ratio, and the convex
combination of them. The simplicity of the problem formulations facilitates their solutions in real time. The proposed planning
method provides an attractive means for dimensioning OFDMA cellular networks in which a large fraction of users experience
group-mobility.
Copyright © 2006 K.-D. Lee and V. C. M. Leung. This is an open access article distributed under the Creative Commons
Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is
properly cited.
1. INTRODUCTION on each subcarrier. Also, they developed suboptimal solu-

tion algorithms by using the linear programming technique
Orthogonal frequency-division multiple-access (OFDMA) is and the Hungarian method. In [4], the authors formulated
one of the most promising solutions to provide a high- a combinatorial problem to jointly optimize the subcarrier
performance physical layer in emerging cellular networks. and power allocation. In their formulation they considered
OFDMA is based on OFDM and inherits immunity to inter- a constraint to allocate resources to users according to the
symbol interference and frequency selective fading. Recently, predetermined fractions with respect to the transmission
adaptive resource management for multiuser OFDMA sys- opportunity. By using the constraint, the resources can be
tems has attracted enormous research interest [1–7]. fairly allocated. A novel scheme to fairly allocate subcarri-
In [1], the authors studied how to minimize the total ers, rate, and power for multiuser OFDMA system was pro-
transmission power while satisfying a minimum rate con- posed [6], where a new generalized proportional fairness cri-
straint for each user. The problem was formulated as an in- terion, based on Nash bargaining solutions (NBS) and coali-
teger programming problem and a continuous-relaxation- tions, was used. The study in [6] is very different from the
based suboptimal solution method was studied. In [2], a class previous OFDMA scheduling studies in the sense that the re-
of computationally inexpensive methods for power alloca- source allocation is performed with a game-theoretic deci-
tion and subcarrier assignment were developed, and those sion rule. They proposed a very fast near-optimal algorithm
are shown to achieve comparable performance, but do not using the Hungarian method. They showed by simulations
require intensive computation. that their fair-scheduling scheme provides a similar overall
Several studies have considered providing a fair oppor- rate to that of the rate-maximizing scheme. In [7], they pro-
tunity for users to access a wireless system so that no user vided achievable rate formulations from the physical layer
may dominate in resource occupancy while others starve. In perspective and studied algorithms using Lagrangian multi-
[3], the authors proposed a fair-scheduling scheme to mini- plier theorem, and they showed that their algorithms can find
mize the total transmit power by allocating subcarriers to the the global optimum even though the problems have multiple
users and then to determine the number of bits transmitted local optima.
However, most previous studies on resource allocation planning approach can also be applied to situations with in-
in OFDMA systems did not consider the connection-level dividual mobility.
performance which is limited by the fluctuations in perfor- Recently, Niyato and Hossain [8] studied two call admis-
mance, for example, signal-to-interference ratio (SNR) in the sion schemes in OFDMA networks. However, they did not
lower layer. Because of the random user mobility, the aver- consider the nonstationary nature of SNR in determining the
age channel gain of a targeted group of users (referred sim- threshold value for admission control, which is the major dif-
ply as the average channel gain in the rest of the paper) in a ference between their contributions and ours. In this paper,
cellular network changes over time causing the average SNR we propose new methods for admission capacity planning
of the user group to continuously fluctuate. Since the maxi- in OFMDA cellular networks, which take into consideration
mum achievable transmit rate is bounded by the SNR, ongo- the random nature of the average channel gain. We derive
ing connections may experience outage events and, further- the outage ratio and the excess capacity ratio, and formu-
more, the outage ratio increases for any given number of con- late three optimization problems to maximize the reduction
nections admitted in the system. Therefore, it is necessary to of the outage ratio, the excess capacity ratio, and the con-
take the fluctuating nature of SNR into account when plan- vex combination of them. The simplicity of the problem for-
ning for the admission capacity. Several different optimiza- mulation enables the admission capacity planning problems
tion criteria have been used for admission capacity planning, to be solved in real time. Extensive simulation results show
such as the average call blocking probability, the average de- that (1) the outage ratio and the excess capacity ratio are
lay, and the utilization of bandwidth resources. small when the variance of the average channel gain is small;
More specifically, we consider admission capacity plan- (2) the desired bit-error rate (BER) and the minimum re-
ning for cellular networks in which a significant fraction of quired transmit rate per connection affect the optimal ad-
users experience “group-mobility,” which is commonly ob- mission capacity but have little affect on the Pareto efficiency
served in mass transportation systems (e.g., bus or train pas- between the outage ratio and the excess capacity ratio; and
sengers). In general, the mobility patterns of users experienc- (3) for relatively small (large) values of targeted outage ratio,
ing group-mobility are correlated causing their channel gains the admission capacity increases (decreases) when the vari-
to be correlated as well. From the perspective of queuing the- ance of the average channel gain is small. We believe that the
ory, group-mobility users arrive at a network according to proposed admission capacity planning method provides an
the “bulk arrival” process, which tends to degrade the tele- attractive means for dimensioning of OFDMA cellular net-
traffic performance (for more details, refer to Section 3.2). In works in which a large fraction of users experience group-
the case of a batch of users arriving at a new cell, for example, mobility.
during a handover event involving the mobile platform, there The remainder of this paper is organized as follows.
are bulk arrivals of calls in the cell. During the cell dwell time Section 2 gives the motivations of this work. Section 3 de-
of users within a mobility-group, new calls may arrive and scribes the model considered in this paper. In Section 4, we
ongoing calls may be completed. The system model based on derive the outage ratio and the excess capacity ratio. In Sec-
batch arrivals therefore gives pessimistic results. However, as tions 5 to 7, we formulate three optimization problems and
the cell size gets smaller, the number of handovers increases develop exact solution methods for maximizing the reduc-
and the results based on batch arrivals become closer to the tion in the outage ratio, the excess capacity ratio, and the con-
actual system performance. vex combination of them. We present simulation results in
Thus, on the one hand, evaluation of admission ca- Section 8 and discuss their implications. Section 9 concludes
pacity without considering the degrading effect of group- the paper.
mobility users may produce results that are too optimistic.
On the other hand, it is clear that the proposed admission 2. MOTIVATIONS AND SCOPE OF THIS WORK
capacity planning based on group-mobility analysis yields a
worst-case quality of service (QoS). However, from service 2.1. Motivations of this work
providers’ perspectives, to provide QoS has higher priority
than to improve bandwidth utilization. For example, even There are extensive studies on subcarrier and power alloca-
though one handover call and one new call will pay the same tions in OFDM (see [1–7] and the literature therein), where
cost per unit time, handover calls are usually given a higher the authors assume that the SNR is not variable during the
priority than new calls from the QoS satisfaction perspective. scheduling period. The results of these studies can be used in
This implies that service provider may prefer the degree of an adaptive manner in accordance with the frequent changes
bandwidth wastage caused by the proposed pessimistic plan- of SNR. Regardless of adaptations with respect to SNR vari-
ning approach compared to the QoS degradation caused by ations, outage events of ongoing real-time connections are
a more optimistic planning approach. Therefore, it stands to unavoidable in the cases where the instantaneous capacity
reason that while admission capacity planning in the pres- with respect to the locations of users residing in a cell be-
ence of group-mobility users gives pessimistic results when comes lower than the minimum capacity required to serve
group-mobility patterns are absent, the possibility of adverse those connections (see Figure 1). A simple solution to im-
impact of group-mobility users must be properly taken into prove the outage ratio of ongoing connections is to apply a
account. With the proposed method, by modifying the out- certain “bound” to the maximum number of connections.
age ratio and the excess capacity ratio, the admission capacity Because of simplicity of this type of solution, it is useful for
K.-D. Lee and V. C. M. Leung 3
where a ≈ −1.5/ log(5 BER) (BER denotes desired bit-error

rate), Gi j denotes the channel gain of user j at subcarrier i,
s
ct ion ns σ 2 is the thermal noise power, and pi j denotes the power allo-
ne tio
n nec cated to user j at subcarrier i [6]. Each connection has a min-
co on
10
0 8 0c imum rate requirement φ, such that an outage event occurs
P0 P1 if the assigned rate is smaller than the minimum required
transmit rate φ.
y
ctor Decrease in Since the users are generally mobile, we consider that the
raje
in t channel gain
channel gains Gi j ’s are random variables. Thus, the optimal
Tra
Base station
allocation of subcarrier and power is dependent upon the in-
stantaneous values of the random variables. Thus, it is not
possible to use a fixed allocation strategy.
Figure 1: An example of group-mobility users on board a train.
In such situations, we propose an alternative to approxi-
The maximum capacities are 100 connections at location P0 and mate the average rate per connection when y connections are
80 connections at P1 . For a planned admission capacity y = 100, ongoing as follows:
a small excess capacity exists and 20 connections are likely to be
dropped. For a planned admission capacity y = 80, a large excess C Ḡ · (y/C · p̄)
R(y) ≈ W log2 1 + a ·
capacity exists and 0 connections are likely to be dropped. y σ2

C a p̄y
= W log2 1 + 2 · Ḡ (2)
practical applications. However, it is necessary to investigate y σ C

how to find appropriate bounds for connection admission C a p̄y
= W log2 1 + ρ(y) · Ḡ ρ(y) = 2 ,
that take into account the particular characteristics of OFDM y σ C
systems, which differentiates this problem from similar prob-
lems in the other wireless systems. where C/ y denotes the average number of subcarriers allo-
y W is the bandwidth of a subcarrier,
cated to a connection,
Ḡ = (1/ yC) Ci=1 j =1 Gi j , and y/C · p̄ is the average power
2.2. Scope of this work
allocated to a subcarrier. There are practical reasons to use
The scope of this work is to find appropriate upper bounds Ḡ instead of the individual random variables Gi j ’s. First, the
of the number of ongoing connections. The objectives are variances of Gi j ’s with respect to indices i and j are small in
to minimize the number of outage events while keeping ca- the case of group-mobility users because the users are located
pacity wastage below a specific limit, or to minimize capac- at the nearly same position with respect to the base station.
ity wastage while keeping the number of outage events be- Second, the mean value Ḡ is an unbiased estimator that pro-
low a given tolerance level. In this paper, we call these up- vides sufficient statistical information on the targeted pop-
per bounds the “admission capacity.” We consider the case ulation. The probability density function (pdf) of random
where the channel gain of user j using subcarrier i, denoted variable Ḡ is denoted by fG (·). In the case of a system filled
by Gi j , is a random variable that varies over time. In this with individual mobility users, the approximation used in (2)
case, the optimal subcarrier and power allocations will vary may not be sufficiently accurate because the channel gains
over time as they are completely dependent on the values and allocated powers of individual mobility users are quite
of the random variables Gi j ’s. We assume the perfect con- different, which is beyond the scope of this work. In the case
dition that optimum power and subcarrier allocations are of group-mobility users, however, because of the first reason,
made given the values of Gi j ’s. This assumption is necessary the approximation is much more accurate.
and widely adopted in the literature to enable an analytical
evaluation of the achievable system capacity. For example, in 3.2. Connections of group-mobility users
capacity planning of CDMA systems with time-division du-
plex (TDD), it is commonly assumed to have perfect power Figure 1 gives an example of group-mobility users traveling
control and resource allocation [9, 10]. onboard a train. The real-time traffic performance of group-
mobility users is usually lower than that of individual mobil-
ity users. For example, consider two M/M/m/m queue mod-
3. MODEL DESCRIPTION
els with the same service rate: an M/M/2c/2c queue with the
3.1. System model arrival and departure rates λ and μ, respectively, where each
arrival requires two channels and M/M/2c/2c one with the
We consider an OFDMA cellular system. A cell has a total of arrival and departure rates 2λ and μ, respectively, where each
C subcarriers and each user has a transmission power limit arrival requires a single channel [11]. The former is the 2-
of p̄. The achievable rate of user j using subcarrier i, Ci j , is user group-mobility example. It can be simply verified that
given by the blocking probability in the former queue model is greater
than that in the latter queue model. This is because group-
Gi j p i j mobility users move in bulk, requesting the respective min-
Ci j = W log2 1 + a · , (1)
σ2 imum capacities almost at the same epoch, in the event of
handovers in the case of a cellular network. Here, note that 4.2. Excess capacity ratio
although each bulk arrival in the former queue model is a
Poisson process, the arrival process of each user is not gener- The average amount of excess capacity S(y) is given by
ally Poisson and, furthermore, it is not a stationary process. y ∞

In this case, the blocking probability of a customer is usually S(y) = (r − φ) · fR (r)dr
greater even when the utilization of bandwidth resources is k=1 φ
low. ∞
The other property of group-mobility users is that they =y (r − φ) · fR (r)dr (8)
have an approximately equal SNR ceteris paribus. This also φ
reduces the capacity that a base station can achieve, as it can- ∞ ∞
not take full advantage of multiuser diversity. =y r · fR (r)dr − φy fR (r)dr,
φ φ
The reason that we take group-mobility users into ac-
count is to examine worst-case performance for admission where f (x) = dF(x)/dx. Substituting G for R, that is, GR (y)
control planning, whereas a great number of previous stud- for R(y), we have
ies overestimated the performance by simplifying the arrival −1
model into a Poisson arrival process [12]. dr
fR (r) = fG (g) ·

, (9)
dg
4. OUTAGE RATIO AND EXCESS CAPACITY RATIO which gives
In this section, we derive the outage ratio and the excess ca- Rmax Gmax
CW R
pacity ratio. The outage ratio is defined as the average frac- r · fR (r)dr = log2 1 + ρ(y)g
r =φ y g =GR (y)
tion of the total number of connections suffering from out-
ages, whereas the excess capacity ratio is defined as the aver- −1
dr dr
age fraction of the achievable capacity that is not utilized for · fG (g) ·
·
dg
dg dg
real-time traffic delivery, even though used for non-real-time
traffic delivery, out of the total achievable capacity. Gmax
CW R
= log2 1 + ρ(y)g · fG (g)dg,
y GR (y)
4.1. Outage ratio Rmax Gmax
R
Let random variable KD (y) denote the number of outages fR (r)dr = fG (g)dg,
r =φ g =GR (y)
(or number of dropped connections) when y connections are (10)
ongoing. The probability that k users are dropped by outage
is given by where Rmax = max R(y) and Gmax
R = max GR (y). Thus, (8) is

rewritten as
y k y −k
Pr KD (y)=k = · Pr R(y)<φ · 1−Pr R(y)<φ Gmax
k R
S(y) = CW log2 1 + ρ(y)g · fG (g)dg

GR (y) (11)
y y −k
= · FR (φ)k · 1 − FR (φ) .
k − φy 1 − FG GR (y) .
(3)
When y ongoing connections have been admitted, the total
The average number of connections experiencing outages is amount of the achievable capacity is given by
given by y Rmax

y
ST (y) = r · fR (r)dr
E KD (y) = k · Pr KD (y) = k = yFR (φ). (4) k=1 r =0
k=1 (12)
Gmax
R
By substituting G for R, we have =y log2 1 + ρ(y)g · fG (g)dg.
g =0
E KD (y) = yFG GR (y) , (5)
Finally, the excess capacity ratio is given by
where GR (y) is the solution of (2) at R = φ with respect to
G, that is, S(y)
PS (y) = . (13)
2 yφ/(CW) − 1 ST (y)
GR (y) = . (6)
ρ(y)
5. MINIMIZATION OF OUTAGE RATIO OF
Thus, the outage ratio is expressed as

ONGOING CONNECTIONS
E KD (y)
PO (y) = We can find the optimal y that minimizes the outage ratio of
y (7)
ongoing connections by solving the following simple prob-
= FG GR (y) . lem (P1).
5.1. Problem formulation: outage ratio minimization The feasible region of y in problem (P1) is given by

(P1) F1 = y : PS (y) ≤ γS = y : y ≥ PS−1 γS . (19)
minimize PO (y), This is supported by Proposition 2, namely, PS−1 (·) exists

subject to PS (y) ≤ γS , (14) and, furthermore,
y : nonnegative integer. dPS−1 1
= < 0. (20)
dy dPS /d y
The role of problem (P1) is to find y that minimizes the
outage ratio of ongoing connections subject to the constraint Thus, there exists a unique optimal solution of (P1),
that the excess capacity ratio is not greater than γS . which is given by

yO∗ = PS−1 γS , (21)
5.2. Solution method of (P1)
where x is the smallest integer not less than x.
Proposition 1. PO (y) is strictly increasing.
Proof. 6. MINIMIZATION OF EXCESS CAPACITY RATIO
dPO dGR (y) Next, we consider the problem of minimizing the fraction of
= fG GR (y) · > 0. (15) excess capacity. The amount of excess capacity represents ca-
dy dy
pacity that is not used by any real-time traffic users and is
therefore wasted. The problem is formulated by (P2) as fol-
Proposition 2. PS (y) is strictly decreasing. lows.
Proof. We have 6.1. Problem formulation: excess capacity

dS(y) ratio minimization
= −CW log2 1 + ρ(y)GR (y)
dy (P2)

dGR (y)
· fG GR (y) · minimize PS (y),
dy
subject to PO (y) ≤ γO , (22)

dGR (y)
− φ 1 − FG GR (y) + yφ fG GR (y) · y : nonnegative integer.
dy
dGR (y) Problem (P2) is subject to the constraint that the outage ratio
= − yφ fG GR (y) · is not greater than γO .
dy
dGR (y)
− φ 1 − FG GR (y) + yφ fG GR (y) · 6.2. Solution method of (P2)
dy
The feasible region of y in problem (P2) is given by
= −φ 1 − FG GR (y) < 0,

(16) F2 = y : PO (y) ≤ γO
Gmax −1
(23)
dST (y) R = y : y ≤ PO (γO ) .
= log2 1 + ρ(y)g
dy 0 Similar to the case of (P1), this is supported by Proposition 1.
Gmax
d R Thus, there exists a unique optimal solution of (P2), which is
· fG (g)dg + y log2 1 + ρ(y)g · fG (g)dg given by
dy 0
Gmax
R yS∗ = PO−1 γO , (24)
= log2 1 + ρ(y)g
0
Gmax where x
is the largest integer not greater than x.
R a p̄/σ 2 C
· fG (g)dg + y · fG (g)dg > 0.
0 1 + ρ(y)g 7. JOINT MINIMIZATION OF OUTAGE RATIO
(17) AND CAPACITY WASTAGE
The inequality (18) can also be demonstrated by the property 7.1. Definition and formalism
of multiuser diversity, where the achievable capacity increases
as the number of users increases [6]. (P3)
From the above results, we have
minimize PC (y : α)
dPS dS(y)/d y · ST (y) − S(y) · dST (y)/d y = αPO (y) + (1 − α)PS (y), y : nonnegative integer.
= 2 <0. (18)
dy ST (y) (25)
Here, α is a constant between 0 and 1, which denotes the

relative marginal utility1 of the outage ratio with respect to
PS (y) (see Figures 13–15). The objective function is a con- 0.1
vex combination of outage ratio and capacity waste frac-
tion. Note that the objective function is not always strictly
α)PS
convex. The necessary and sufficient condition for the ob-
0.01
jective function (αPO (y) + (1 − α)PS (y)) to be strictly con-
αPO + (1
vex is that the second difference2 is positive for all integers
y = 1, . . . , C − 1. For the sake of tractability, we may con-
sider as a sufficient condition that the second derivative of 1E 3
{αPO (y) + (1 − α)PS (y)} is positive if

dfG d2 GR /d y 2 1
> − fG GR (y) · − −1 φ (26) 1E 4
dy dGR /d y α
1220 1240 1260 1280 1300
for 1 < y < C − 1. The nonconvexity of PC (y : α) with respect
Max. no. of connections, y
to y can be observed in the examples shown in Figure 2.
BER = 1E 4, α = 0.3 N (100, 5), α = 0.3
BER = 1E 4, α = 0.5 N (100, 5), α = 0.5
7.2. Is it useful? BER = 1E 5, α = 0.3 N (100, 10), α = 0.3
BER = 1E 5, α = 0.5 N (100, 10), α = 0.5
Even though applying (P1) and (P2) for admission capac- BER = 1E 6, α = 0.3 N (100, 20), α = 0.3
ity planning is useful under the condition that the required BER = 1E 6, α = 0.5 N (100, 20), α = 0.5
levels of PO (y) or PS (y), namely γO or γS , are given, these
problems are not enough for us to plan the admission capac- Figure 2: Nonconvexity of PC (y : α) with respect to y (PC (y : α) =
ity in all cases. In some cases, the required level is not given αPO (y) + (1 − α)PS (y)).
and the only information available for planning is the rela-
tive marginal utility α. In such cases, the above problem (P3)
is useful to determine the admission capacity (examples for
this case can be found in Figures 13–15). Given that the rel- 8. EXPERIMENTAL RESULTS
ative marginal utility α is 0.5, the left point y ∗ (specified by We examine the three proposed methods for various proba-
α = 0.5) is optimal. However, if the relative marginal utility bility density functions (pdf ’s) of the average channel gain Ḡ
decreases to 0.3, then the optimal point moves to the right and for various values of BER, φ, σ 2 , and p̄. In our simula-
one (specified by y ∗ at α = 0.3), causing a balance with a tion setups the transmission power is p̄ = 50 mW, the ther-
decrease in PS (denotes PS gains more weight) and an in- mal noise power is σ 2 = 10−11 W, the number of subcarriers
crease in PO (denotes PO loses more weight). The solution is C = 128 over a 3.2 MHz band, BER = 10−5 , and the mini-
methods used for solving (P1) and (P2) can be applied for mum rate requirement is φ = 100 kbps; all are used as default
(P3) after simple modifications. A simple and exact solu- values. Table 1 shows the simulation parameters values.
tion method is demonstrated in Figures 13–15 Section 8. Be- Figures 3–7 show the admission capacity y versus the
cause there is a unique inflection point for PO (y) and PS (y) threshold value of excess capacity ratio. Note that in these
and the two functions, namely PO (y) and −PS (y), are strictly figures, the actual shape of the curves are given by the step
increasing, there are at most two local minima of function functions denoting PS−1 (γS ). In Figure 3, the real shapes of
PC (y : α) = αPO (y) + (1 − α)PS (y). the curves are shown whereas the curves are smooth in the
other four figures; that is, in Figures 4–7, the curves denote
Proposition 3. The necessary condition for (local) optimality
PS−1 (γS ) instead of PS−1 (γS ).
is
In Figure 3, the admission capacities are shown with re-
dPC dPO dP spect to desired bit-error rate (BER). As we can see through
=α + (1 − α) S = 0. (27)
dy dy dy the achievable rate formula (1), the admission capacity de-
creases when BER decreases and when the targeted excess ca-
Alternatively, the necessary condition for (local) optimality can pacity ratio increases. In both cases, the admission capacity
be expressed as decreases approximately linearly with the decrease in BER. It
dPO 1−α is observed that the differences between admission capacities
=− . (28) at different values of BER decrease when the targeted outage
dPS α
ratio γO increases.
Figure 4 shows the admission capacity versus the thresh-
1
old value of excess capacity ratio with respect to transmit
This denotes the marginal utility with respect to PS (y) instead of the
marginal utility with respect to y.
power. It is observed that the admission capacity increases
2 The first difference of a function is defined as Δ f (n) = f (n + 1) − f (n) as the transmit power p̄ increases but with a decreasing rate,
and the second difference is defined as Δ2 f (n) = Δ f (n + 1) − Δ f (n). which we can conjecture from (1). In addition, it is observed
Table 1: Parameters used in experiments.
Item Value Description

p̄ 50 Avg. transmit power (mW)
σ2 1e −11 Thermal noise level (W)
C 128 No. of subcarriers
BER 1e −5 Desired bit-error rate
W 25 000 Bandwidth of subcarrier (Hz)
φ 100 Min. required rate per connection (kbps)
Ḡ ∼ N (100,5) —
1250 1250
1200 1200

1150
1150
1100
1100
1050
1050
1000
1000
950
950
900
1E 4 1E 3 0.01
900
γS 1E 4 1E 3 0.01
BER = 1E 3 BER = 1E 6 γS
BER = 1E 4 BER = 1E 7 p̄ = 30 mW p̄ = 60 mW
BER = 1E 5 p̄ = 40 mW p̄ = 70 mW
p̄ = 50 mW
Figure 3: The maximum number of connections y versus γS with Figure 4: The maximum number of connections y versus γS with
respect to BER ( p̄ = 50 mW, σ 2 = 10−11 , φ = 100 kbps, N (100, 5)). respect to p̄ (BER = 10−5 , σ 2 = 10−11 , φ = 100 kbps, N (100, 5)).
that a ±10% increase in transmit power at 50 mW can in- as admission planning with a dynamic value of admission
crease approximately ±10% of admission capacity at any threshold, is preferred compared to a static planning strat-
given threshold value of excess capacity ratio. Similarly, a egy, such as admission planning with a fixed value of ad-
±20% increase in transmit power at 50 mW results in ap- mission threshold. This is because a static planning strat-
proximately ±20% increase in admission capacity. egy does not adjust well to the high variations in the case
Figure 5 shows the admission capacity versus the targeted of a large variance. This fact demonstrates that the admis-
excess capacity ratio with respect to the minimum required sion capacity decreases as the variance of Ḡ increases, which
transmit rate per connection. It is observed that a ±1, 2% is observed in the figure. However, it is observed that an
increase in φ results in an approximately equal decrease in 8-fold increase in the variance at 5 results in a 0.5% de-
admission capacity y ∗ . This is because the total capacities, crease in admission capacity. Thus, we can safely conclude
y ∗ · φ, are approximately equal regardless of the value of φ. that under the condition that Ḡ has a large variance the ad-
Figure 6 shows the admission capacity versus the targeted ex- mission capacity decreases but the amount of decrease is
cess capacity ratio with respect to the thermal noise power. slight.
Similar patterns of admission capacity are observed. Figures 8–12 show the maximum number of connections
Figure 7 shows the admission capacity versus the tar- that can be accommodated, which is defined as the admis-
geted excess capacity ratio with respect to the pdf of the sion capacity and is denoted by y in this paper, versus the
random variable Ḡ, that is, the average channel gain, where threshold value of outage ratio. In these figures, note that
N (x, y) denote a normal distribution with mean x and vari- the actual shape of the curves are the step functions denot-
ance y. Obviously, a large variance implies a high degree ing PO−1 (γO )
. In Figure 8, the actual shapes of the curves
of variation. In this case, a dynamic planning strategy, such are shown whereas the curves are smoothed in the other four

1250 1200
1200
1100
1150
1000
1100
900
1050 1E 4 1E 3 0.01
γS
1000
N (100, 5) N (100, 20)
N (100, 10) N (100, 40)
950
(a)
900
1E 4 1E 3 0.01
γS

φ = 98 (kbps) φ = 101 (kbps) 1210
φ = 99 (kbps) φ = 102 (kbps) 1205
φ = 100 (kbps)
1200
Figure 5: The maximum number of connections y versus γS with 1195

respect to φ(BER = 10−5 , p̄ = 50 mW, σ 2 = 10−11 , N (100, 5)). 1190
1E 3
γS
N (100, 5) N (100, 20)
1250 N (100, 10) N (100, 40)
(b)
1200
Figure 7: The maximum number of connections y versus γS with

1150
respect to the pdf of Ḡ (BER = 10−5 , p̄ = 50 mW, σ 2 = 10−11 ,
φ = 100 kbps).
1100
1050
1000
950 1300
900
1E 4 1E 3 0.01
1280
γS
σ 2 = 10 10.8 σ 2 = 10 11.1
σ 2 = 10 10.9 σ 2 = 10 11.2
1260
σ 2 = 1E 11(= 10 11 )
Figure 6: The maximum number of connections y versus γS with

1240
respect to σ 2 (BER = 10−5 , p̄ = 50 mW, φ = 100 kbps, N (100, 5)).
1220
figures, that is, in Figures 9–12, the curves denote PO−1 (γO ) 1E 8 1E 6 1E 4 0.01 1
instead of PO−1 (γO )
. γO
In Figure 8, the admission capacities are shown with re- BER = 1E 3 BER = 1E 6
spect to desired bit-error rate. It is observed that the differ- BER = 1E 4 BER = 1E 7
ences between admission capacities with respect to different BER = 1E 5
values of BER are nearly equivalent regardless of the targeted
outage ratio γO . Obviously, the admission capacity increases Figure 8: The maximum number of connections y versus γO with
when BER decreases and the targeted outage ratio increases. respect to BER ( p̄ = 50 mW, σ 2 = 10−11 , φ = 100 kbps, N (100, 5)).
1300
1280
1280

1260
1260
1240
1240
1220
1220
1E 8 1E 6 1E 4 0.01 1 1E 8 1E 6 1E 4 0.01 1
γO γO
p̄ = 30 mW p̄ = 60 mW
σ 2 = 10 10.8 σ 2 = 10 11.1
p̄ = 40 mW p̄ = 70 mW σ 2 = 10 10.9 σ 2 = 10 11.2
p̄ = 50 mW
σ 2 = 1E 11(= 10 11 )
Figure 9: The maximum number of connections y versus γO with Figure 11: The maximum number of connections y versus γO with
respect to p̄ (BER = 10−5 , σ 2 = 10−11 , φ = 100 kbps, N (100, 5)). respect to σ 2 (BER = 10−5 , p̄ = 50 mW, φ = 100 kbps, N (100, 5)).
1270
1300
1280 1260
1260
1250
1240
1240
1E 3 0.01 0.1 1
1220
1E 4 1E 3 0.01 0.1 1 γO
γO N (100, 5) N (100, 15)
φ = 101 (kbps) N (100, 10) N (100, 20)
φ = 98 (kbps)
φ = 99 (kbps) φ = 102 (kbps)
φ = 100 (kbps)
Figure 10: The maximum number of connections y versus γO with Figure 12: The maximum number of connections y versus γO with
respect to φ (BER = 10−5 , p̄ = 50 mW, σ 2 = 10−11 , N (100, 5)). respect to the pdf of Ḡ (BER = 10−5 , p̄ = 50 mW, σ 2 = 10−11 ,
φ = 100 kbps).
In both situations, the quality of service, such as link error ratio γO . The rate of increase in admission capacity decreases
quality and dropping probability, is relatively bad. as the transmit power increases, following the logarithmic
Figure 9 shows the admission capacity versus the targeted scale.
outage ratio with respect to the transmit power. It is observed Figure 10 shows the admission capacity versus the tar-
that the admission capacity increases as the transmit power p̄ geted outage ratio with respect to the minimum required
increases. In addition, it is observed that the differences be- transmit rate per connection. It is observed that a ±1, 2% of
tween admission capacities with respect to different values increase in φ results in an approximately equal amount of
of p̄ are nearly equivalent regardless of the targeted outage decrease in admission capacity y ∗ . This is because the total
1E 3 1E 3
1E 4 1E 4
1E 5 y 1E 5 y
= 1268 (BER = 1E 4) = 1282 (φ = 98 kbps)
PS
PS
= 1255 (BER = 1E 5) = 1255 (φ = 100 kbps)
1E 6 = 1245 (BER = 1E 6) 1E 6 = 1230 (φ = 102 kbps)
at γO = 0.01 at γO = 0.01
1E 7 1E 7
1E 8 1E 8
1E 4 1E 3 0.01 0.1 1 1E 4 1E 3 0.01 0.1 1
PO PO
BER = 1E 4 φ = 98 kbps
BER = 1E 5 φ = 100 kbps
BER = 1E 6 φ = 102 kbps
Figure 13: PO (y) versus PS (y) with respect to BER ( p̄ = 50 mW, Figure 14: PO (y) versus PS (y) with respect to φ (BER = 10−5 ,
φ = 100 kbps, σ 2 = 10−11 , N (100, 5)). In the case that α = 0.5, y ∗ = p̄ = 50 mW, σ 2 = 10−11 , N (100, 5)). In the case that α = 0.5,
1263, 1251, 1241 for BER = 10−4 , 10−5 , 10−6 , respectively. In the case y ∗ = 1278, 1251, 1226 for φ = 98(−2%), 100, 102(+2%) (kbps), re-
that γO = 0.01, y ∗ = 1268, 1255, 1245 for BER = 10−4 , 10−5 , 10−6 , spectively.
respectively.
capacities, namely y ∗ · φ, are approximately equal regard- 102, the admission capacity y ∗ is equal to 1278, 1251, 1226 in
less of the value of φ. Figure 11 shows the admission capacity the case of α = 0.3; y ∗ is equal to 1277, 1251, 1225 in the case
versus the targeted outage ratio with respect to the thermal of α = 0.5; and y ∗ is equal to 1277, 1251, 1225 in the case of
noise power. Similar patterns of admission capacity are ob- α = 0.7.
served. Figure 15 shows the relation between excess capacity ra-
Figure 12 shows the admission capacity versus the tar- tio PS and outage ratio PO with respect to the pdf ’s of the
geted outage ratio with respect to the variance of the random average channel gain Ḡ. For the respective pdf ’s N (100, 5),
variable Ḡ, that is, the average channel gain. When γO is less N (100, 10), N (100, 20), the admission capacity y ∗ is given
than about 0.46, the larger the variance of Ḡ is, the higher the by 1251, 1239, 1206 in the case of α = 0.3; y ∗ is given by
rate of increase in the admission capacity is, and the admis- 1251, 1237, 1200 in the case of α = 0.5; y ∗ is given by
sion capacity in the case of a small variance is greater than 1250, 1236, 1193 in the case of α = 0.7. Unlike Figures 13
in the case of a large variance. However, when γO > 0.46, the and 14, the regions of Pareto efficiency between PS and PO
admission capacity in the case of a large variance is greater are quite different from each other with respect to the vari-
than that in the case of a small variance. ance of the random variable Ḡ. It is observed that the smaller
Figure 13 shows the relation between excess capacity ra- the variance is, the better both PS and PO are.
tio PS and outage ratio PO with respect to the desired bit-
error rate (BER). In Figures 8 and 3, it has been shown that
BER affects the admission capacity in both cases of (P1) and 9. CONCLUDING REMARKS
(P2). However, the effect of BER on the relation between
PS and PO is very small. This implies that the regions of Because the admission capacity, which is defined as the up-
Pareto efficiency between PS and PO are almost equivalent per bound of the number of connections that a base sta-
regardless of the desired bit-error rate. For the respective val- tion can accommodate, fluctuates in accordance with the
ues BER = 1E − 4, 1E − 5, 1E − 6, the admission capacity y ∗ signal-to-noise ratio, a portion of ongoing connections may
is equal to 1264, 1251, 1241 in the case of α = 0.3, y ∗ is equal be dropped prior to their normal completion because of out-
to 1263, 1251, 1241 in the case of α = 0.5, and y ∗ is equal to age events. In this paper, we have developed three methods
1263, 1250, 1240 in the case of α = 0.7. This implies that the for admission capacity planning of an orthogonal frequency-
larger α is, the smaller is the admission capacity. A larger α division multiple-access system. Taking into account of the
should result in a smaller outage ratio. fluctuations of the average channel gains, we have derived
Figure 14 shows the relation between excess capacity ra- outage ratio at the connection level, and the excess capac-
tio PS and outage ratio PO with respect to the minimum re- ity ratio. Based on these metrics, we have formulated three
quired transmit rate φ. For the respective values φ = 98, 100, problems to optimize admission capacity by maximizing
1E 3 [3] M. Ergen, S. Coleri, and P. Varaiya, “Qos aware adaptive re-

source allocation techniques for fair scheduling in OFDMA
based broadband wireless access systems,” IEEE Transactions
on Broadcasting, vol. 49, no. 4, pp. 362–370, 2003.
1E 4
[4] C. Mohanram and S. Bhashyam, “A sub-optimal joint sub-
carrier and power allocation algorithm for multiuser OFDM,”
y IEEE Communications Letters, vol. 9, no. 8, pp. 685–687, 2005.
= 1255 (N (100, 5))
PS
1E 5 [5] Y. J. Zhang and K. B. Letaief, “Multiuser adaptive subcarrier-

= 1250 (N (100, 10)) and-bit allocation with adaptive cell selection for OFDM sys-
= 1232 (N (100, 20)) tems,” IEEE Transactions on Wireless Communications, vol. 3,
at γO = 0.01 no. 5, pp. 1566–1575, 2004.
1E 6
[6] Z. Han, Z. Ji, and K. J. Ray Liu, “Fair multiuser channel allo-
cation for OFDMA networks using Nash bargaining solutions
and coalitions,” IEEE Transactions on Communications, vol. 53,
1E 7 no. 8, pp. 1366–1376, 2005.
1E 6 1E 5 1E 4 1E 3 0.01 0.1 1 [7] Y. Yao and G. B. Giannakis, “Rate-maximizing power allo-
PO cation in OFDM based on partial channel knowledge,” IEEE
Transactions on Wireless Communications, vol. 4, no. 3, pp.
N (100, 5) 1073–1083, 2005.
N (100, 10)
N (100, 20) [8] D. Niyato and E. Hossain, “Connection admission control al-
gorithms for OFDM wireless networks,” in Proceedings of IEEE
Global Telecommunications Conference (GLOBECOM ’05), pp.
Figure 15: PO (y) versus PS (y) with respect to the pdf of Ḡ (BER = 2455–2459, St. Louis, Mo, USA, November-December 2005.
10−5 , p̄ = 50 mW, σ 2 = 10−11 , φ = 100 kbps). In the case that α =
[9] L.-C. Wang, S.-Y. Huang, and Y.-C. Tseng, “Interference anal-
0.5, y ∗ = 1251, 1238, 1201 for N (100, 5), N (100, 10), N (100, 20),
ysis and resource allocation for TDD-CDMA systems to sup-
respectively.
port asymmetric services by using directional antennas,” IEEE
Transactions on Vehicular Technology, vol. 54, no. 3, pp. 1056–
1069, 2005.
[10] M. Casoni, G. Immovilli, and M. L. Merani, “Admission con-
the reduction of the outage ratio, the excess capacity ratio, trol in T/CDMA systems supporting voice and data applica-
tions,” IEEE Transactions on Wireless Communications, vol. 1,
and the convex combination of them. Because of the sim-
no. 3, pp. 540–548, 2002.
plicity of its formulation, each problem can be solved in
[11] S. Ross, Stochastic Processes, John Wiley & Sons, New York, NY,
real time. We believe that the proposed capacity planning
USA, 2nd edition, 1996.
method can be effectively applied in the design and dimen-
[12] K.-D. Lee, “Variable-target admission control for nonstation-
sioning of OFDMA cellular networks, especially in situations
ary handover traffic in LEO satellite networks,” IEEE Transac-
where a significant fraction of the users experience group- tions on Vehicular Technology, vol. 54, no. 1, pp. 127–135, 2005.
mobility.
ACKNOWLEDGMENTS Ki-Dong Lee received the B.S. and M.S.

degrees in operation research (OR) and
The authors are grateful to the anonymous reviewers for their the Ph.D. degree in industrial engineer-
constructive comments which greatly improved the quality ing (with applications to wireless networks)
of presentation of this paper. This work was supported in from the Korea Advanced Institute of Sci-
part by the Korea Research Foundation (KRF) under Grant ence and Technology (KAIST), Daejeon,
KRF-2005-214-D00139 and in part by the Canadian Natural Korea, in 1995, 1997, and 2001, respectively.
From 2001 to 2005, he was a Senior Mem-
Sciences and Engineering Research Council through Grant
ber of engineering staff at the Electronics
STPGP 269872-03. and Telecommunications Research Institute
(ETRI), Daejeon, where he was involved with several government-
funded research projects. Since 2005, he has been with the Depart-
REFERENCES ment of Electrical and Computer Engineering, University of British
Columbia (UBC), Canada, as a Research Associate. His research
[1] C. Y. Wong, R. S. Cheng, K. B. Letaief, and R. D. Murch, “Mul- interests are in performance evaluations, optimization techniques,
tiuser OFDM with adaptive subcarrier, bit, and power allo- and their applications to radio resource management in wireless
cation,” IEEE Journal on Selected Areas in Communications, multimedia networks. He received the IEEE ComSoc AP Outstand-
vol. 17, no. 10, pp. 1747–1758, 1999. ing Young Researcher Award in 2004 and the Asia-Pacific Oper-
[2] D. Kivanc, G. Li, and H. Liu, “Computationally efficient band- ations Research Society (APORS) Young Scholar Award in 2006,
width allocation and power control for OFDMA,” IEEE Trans- and he served as a Coguest Editor for the Special Issue on Next-
actions on Wireless Communications, vol. 2, no. 6, pp. 1150– Generation Hybrid Wireless Systems in the IEEE Wireless Com-
1158, 2003. munications.
Victor C. M. Leung received the B.A.S.

(with honors.) and Ph.D. degrees, both
in electrical engineering, from the Univer-
sity of British Columbia (UBC) in 1977
and 1981, respectively. He was the recipi-
ent of many academic awards, including the
APEBC Gold Medal as the Head of the 1977
graduate class in the Faculty of Applied Sci-
ence, UBC, and the NSERC Postgraduate
Scholarship. From 1981 to 1987, he was a
Senior Member of Technical Staff and Satellite Systems Specialist at
MPR Teltech Ltd. In 1988, he was a lecturer in electronics at the
Chinese University of Hong Kong. He returned to UBC as a Fac-
ulty Member in 1989, where he is a Professor and holder of the
TELUS Mobility Research Chair in Advanced Telecommunications
Engineering in the Department of Electrical and Computer Engi-
neering. His research interests are in mobile systems and wireless
networks. He is a Fellow of IEEE and a voting member of ACM.
He is an Editor of the IEEE Transactions on Wireless Communi-
cations, an Associate Editor of the IEEE Transactions on Vehicular
Technology, and an Editor of the International Journal of Sensor
Networks.
DOI 10.1155/WCN/2006/79148
Joint Downlink Power Control and Multicode Receivers

for Downlink Transmissions in High Speed UMTS
Bessem Sayadi, Stefan Ataman, and Inbar Fijalkow
ETIS/ENSEA, University of Clergy-Pontoise/CNRS, 6 Avenue du Ponceau, 95014 Clergy-Pontoise, France
Received 30 September 2005; Revised 28 February 2006; Accepted 19 May 2006

We propose to combine the gains of a downlink power control and a joint multicode detection, for an HSDPA link. We propose
an iterative algorithm that controls both the transmitted code powers and the joint multicode receiver filter coefficients for the
high-speed multicode user. At each iteration, the receiver filter coefficients of the multicode user are first updated (in order to
reduce the intercode interferences) and then the transmitted code powers are updated, too. In this way, each spreading code of
the multicode scheme creates the minimum possible interference to others while satisfying the quality of service requirement. The
main goals of the proposed algorithm are on one hand to decrease intercode interference and on the other hand to increase the
system capacity. Analysis for the rake receiver, joint multicode zero forcing (ZF) receiver, and joint multicode MMSE receiver is
presented. Simulation is used to show the convergence of the proposed algorithm to a fixed point power vector where the multicode
user satisfies its signal-to-interference ratio (SIR) target on each code. The results show the convergence behavior for the different
receivers as the number of codes increases. A significant gain in transmitted base station power is obtained.
Copyright © 2006 Bessem Sayadi et al. This is an open access article distributed under the Creative Commons Attribution License,
1. INTRODUCTION codes in a synchronous multipath free channel. However,

multipath propagation partially destroys the orthogonality of
As wireless access to the internet rapidly expands, the need the multicode transmission and leads to a significant self in-
for supporting multirate services (voice, data, multimedia, tercode interference which increases with the number of par-
etc.) over limited spectrum increases. CDMA technologies allel codes for a multicode scheme. Therefore, the quality of
are being considered for third-generation wireless networks, the downlink under frequency selective fading environments
UMTS. There are hence two channelization schemes for is interference limited. In this paper, we consider a single cell
achieving multirate transmissions. The first, known as the environment where one or more users employ a multicode
variable spreading factor scheme, achieves variable-data rate downlink transmission.
transmission by assigning the radio link a single variable- In order to improve the quality of the downlink which
length random spreading sequence. However, short codes, is typically defined in terms of the signal-to-interference ra-
when subjected to a large delay-spread multipath channel tio (SIR), a joint multicode reception was recently proposed
loose their orthogonality and lead to a significant intersym- in [3] with the assumption that the different codes have a
bol interference (ISI). To circumvent this limitation, we con- fixed transmitting power. Based on a description of the signal
sider the second option called multicode transmission. The received over fading code-division multiple-access channel,
high-rate data stream is split into several lower rate data sub- where many different data rates are considered, it is shown
streams [1]. Each substream is spread by a specific spreading in [3] that the problem of recovering the multicode user can
sequence and all the substreams are then transmitted syn- be expressed as a multiuser interference cancelation problem,
chronously as virtual users. A future transmission mode such where each channel code represents a virtual user.
as the high-speed downlink packet access (HSDPA [2]) will Independently in literature, power control is proposed,
make wide use of multicode to considerably increase the data classically for the link between the multiusers and the base
rate in the downlink with a peak-data rate in the range of 10– station (BS), to overcome the near-far problem, to maintain
14 Mbit/s. All the spreading sequences are orthogonal to each the mobile station power consumption, and to reduce the
other to avoid signal interference between parallel channel cochannel interference. The power control approach assumes
that a fixed receiver, usually the conventional (single user) control, the BS output power is adapted to the radio link con-
receiver, is being used. It optimizes the communication be- ditions.
tween the mobiles and the BS by controlling the transmitted The implementation of this approach, in the HSDPA
powers of the different users [4, 5]. mobile, requires interference measurements for each code.
Given the importance of power control, an extensive re- These measurements are envisaged in HSDPA standard [20].
search is focused on this subject. In [6], two optimization We show, using simulations, that the resulting algorithm
criteria are considered in a single-cell case: minimizing total converges to a fixed point power vector where the multi-
transmitted power and maximizing throughput. In [7], the code user satisfies its signal-to-interference ratio (SIR) tar-
optimum power vector is given and also statistics on the re- get on each code. The feasibility of the proposed approach
ceived power are considered. A statistical approach of the op- is based on the transmission of the requested code power
timum power solution is developed in [8]. The existence (or via a feedback link in order to update the BS output pow-
feasibility) of this optimal power allocation is also considered ers. Such a feedback is considered in the HSDPA standard
in [7, 9]. A distributed and iterative power control algorithm where the mobile transmits the channel quality indicator to
where each user’s power converges to the minimum power the base station [2]. In this study, we consider the case of the
needed to meet its quality of service (QoS) specification is joint zero forcing and the joint minimum mean square er-
proposed in [10]. A joint optimization of both receiver filters ror (MMSE) multicode linear receivers for various scenarios
and user transmit powers has been considered in [11] to find where we compare their performance to those obtained by
the jointly optimum powers and linear MMSE (minimum considering a bank of rake receivers considered, here, as the
mean square error) filter coefficients. A similar approach is conventional power control strategy.
proposed in reference [12] where the authors employ a suc- The paper is organized as follows. Section 2 introduces
cessive interference cancelation scheme. Recently, a unified the proposed linear algebraic model which describes the sig-
approach of the uplink power control that is applicable to nal received over time-dispersive fading channel including
a large family of multiuser receivers is proposed in [13, 14], a hybrid multicode/variable spreading factor transmissions.
based on the large system results published in [15]. Section 3 gives the problem statement. The proposed strat-
Based on the fact that for a fixed base station assignment egy is introduced in Sections 4 and 5, and its performance in
the feasibilities of uplink and downlink are equivalent (see a simplified HSDPA environment is assessed by means of nu-
[16] for more details), the authors in [16] present a joint merical simulations in Section 6. Finally, Section 7 presents
power control and base station assignment for the downlink. our conclusions.
Many others researchers are interested on the study of the Throughout this paper scalars, vectors, and matrices are
downlink power control such as [17–19]. In [17], the authors lower case, lower-case bold and upper-case bold characters,
studied the joint optimal power control and beamforming respectively. (·)T , (·)−1 denote transposition and inversion,
in wireless networks. In [18], the authors studied the down- respectively. Moreover, E(·) denotes the expected value op-
link power control allocation for multiclass wireless systems. erator.
However, in the case of HSDPA system, the way the base sta-
tion (BS) must allocate the power on the different codes in
the case of multicode transmission is still an open issue. It 2. SYSTEM MODEL
is indeed desirable for the BS not to use more transmission We assume a multicode CDMA frequency division duplex
power than what it needs to. This paper proposes a possible cellular system. In each cell, K mobile users, each employ-
way to solve this problem. ing a different rate, communicate with a base station. Each
In order to achieve this goal, we propose in this paper user receives a frame with a standardized number of chips
to combine the downlink power control approach and the denoted by Nchip . Based on the quality of service required by
joint multicode detection, presented in [3], for the multiuser k, the base station assigns Mk spreading codes, the pro-
code user. We propose an algorithm which controls both cessing gain is denoted by Gk , at the condition that Nchip =
the transmitted code powers at the BS and the joint mul- (k)
Gk Nbit (k)
where Nbit is the number of transmitted symbols for
ticode receiver filters implemented in the mobile. The re- user k. Under the constraint that a constant chip rate, 1/Tc ,
sulted algorithm adapts the transmitted code’s powers tak- where Tc denotes the chip period, must be maintained, the
ing into account a multicode reception strategy at the mo- symbol period, denoted here by Ts,k = Gk Tc , varies with the
bile which aims to reduce the intercode interference. Math- requested rate by user k. The index s is related to the symbol
ematically, the strategy involves two alternate optimization period and the index k is related to the kth user. In order to
problems which are resolved iteratively in the proposed algo- facilitate the description, the terminologies defined in Table 1
rithm. At each iteration first the receiver filter coefficients of are used in the rest of this paper.
the multicode user are updated to reduce the intercode in-
The path-loss attenuation between the BS and the kth
terference and then the transmitted code powers are updated
user is denoted by zk . In the no-shadowing scenario, the path
and assigned. So that, each spreading code of the multicode
loss (PL) is modeled as a simple distance-dependent loss:
scheme creates the minimum possible interference to others
while satisfying the quality of service requirement. This al-
gorithm has as main goals to decrease intercode interference
−σ
and to increase the system capacity. Using downlink power zk(PL) ≈ λdk (1)
Bessem Sayadi et al. 3
Table 1: Terminology description. channel, corresponding to user k, is described as the follow-

ing:
Notation Description
K the number of user hk = zk h. (6)
Nchip the number of chips in a one radio block
The transmit power towards the kth user on mth code will be
Gk the spreading factor assigned to the kth user (k)
denoted by pm . The transmitted signal for the kth user can
Mk the number of spreading code assigned to the kth user
be written as
(k) the number of bits or symbols transmitted in a
Nbit Nbit,k −1 Mk

one radio block (k) (k) (k)
yk (t) = pm bm (n)cm t − nTs,k , (7)
Tc the common chip period n=0 m=1
Ts,k the symbol period related to the kth user, 1 ≤ k ≤ K
where
zk the attenuation due to the path loss and the shadowing
k −1
G
L the number of paths (k) (k),(q)
cm (t) = cm ψ t − qTc (8)
τi the delay of the ith path q=0
(k)
pm the power of the mth code, 1 ≤ m ≤ Mk of the kth user
(k)
n the symbol index time with Gk the spreading factor for the kth user and bm (n) is
b(k) the transmitted symbol vector by the kth user the transmitted symbol at time n for the kth user on the
(k)
C(k) the spreading coding matrix related to the kth user mth channel-code denoted by cm (t) · ψ is a normalized chip
waveform of duration Tc . The base-band received signal at
W(k) the code’s power matrix related to the kth user
the desired user can be written as
H(k) the channel matrix related to the kth user
n the noise vector r(t)

K −1
L Nbit,k −1 Mk

(k) (k) (k)
= zk αl pm bm (n)cm t − nTs,k − τl +n(t),
or, in dB, k=1 l=0 n=0 m=1

(9)
zk(PL) [dB] ≈ 10 log10 (λ) − 10 · σ · log10 dk , (2)
where n(t) is a zero-mean additive white Gaussian noise
where the constants λ usually depend on the frequency used, (AWGN) process.
as well as the height of the base station and the wireless The received signal is time-discretized at the rate of 1/Tc ,
terminal. The dk is the distance from user k to the base sta- leading to a chip-rate discrete-time model that can be written
tion. The attenuation coefficient σ is usually between 2 and 6 as
for most indoor and outdoor environments. The model pre-
rl = r lTc
sented in (1) is a general form for the most empirical and
semiempirical path-loss attenuation model. For more details,
K −1
L Nbit,k −1 Mk

(k) (k) (k)
the reader can refer to [21]. = zk αl pm bm (n)cm l − nGk − tl,k Tc
In the shadowing case (SH), the variation due to shadow- k=1 l=0 n=0 m=1

ing is added to the path-loss value to obtain the variations. + n lTc ,
Therefore, the path-loss can be modeled as the product of a (10)
distance-dependent path-loss attenuation and a random log-
normally distributed shadowing effect [21]: where tl,k = τl /Gk is the time-discretized path delay in sam-
ple intervals (chip period).
−σ
zk(PL,SH) ≈ λdk 10ξk /10 , ξk ∼ N 0, σξ2 (3) Throughout the paper, we employ a block model. The
blocks of transmitted symbols for each user, k = 1, . . . , K, are
or, in dB, concatenated in a vector:
T
zk(PL,SH) [dB] ≈ 10 log10 (λ) − 10 · σ · log10 dk + ξk , (4) b(k) = b1(k) (0), . . . , bM
(k) (k)
(0), . . . , bM (k)
Nbit −1 (11)
k k
where N (0, σξ2 ) is the Gaussian density with mean 0 (in dB) (k)
containing Nbit bits transmitted with the different codes for
and variance σξ2 (in dB). In the rest of the paper, we denote a given user, k.
zk(PL,SH) by zk . The transmission of the data sequence over the CDMA
The effect of the downlink multipath channel is repre- channel can be expressed by the received sequence r [3]:
sented by a vector with L paths denoted, here, by T
r = r1 , . . . , rNchip +L−1
T
h = α0 , α1 , . . . , αL−1 (5)
K (12)
=
(k) W(k) b(k) + n,
C(k) H
with corresponding delays [τ0 , . . . , τL−1 ]. Therefore, the k=1

(k)= diag(hk , . . . , hk) is of size (Nbit
where H (k) (k)
Mk L, Nbit Mk ) and Let F denote the joint multicode receiver filter employed
(k)
W = diag(P
(k) (k) , P (k) , . . . , P (k) of size N M where P(k) = by the receiver of user 1, user of interest. From the output
bit k
of the joint multicode receiver, y = FT r, the SIR of virtual
diag( p1(k) , p2(k) , . . . , pM(k)
) and diag(X) represents the di-
k user of interest can be written for code m and symbol n as
agonal matrix containing only the diagonal elements of the
the following:
matrix X. The matrix C(k) represents the code matrix of size
(k)
((Nchip + L − 1), Nbit Mk L) built as follows:
(1) 2
pm E β F, hk , C(k) bm (n)
SIR(m, n) = 2 (15)
k
C(k) = v0,0,0 , . . . , vNk bit,k −1,Mk −1,L−1 , E Ω pm =m
T
T
k
vn,m,l = 0TnGk , ukm,l , 0T(Nbit,k −n−1)Gk ,
T (13) for m=1, . . . , M1 , m=1, . . . , M1 , and n=1, . . . , Nbit,1 · Ω(pm =m)
T
ukm,l = 0Ttl , ckm , 0TL−tl −1 , is the sum of the intercode interferences, the multiple access
T interference, the intersymbols interference, and the noise.
ckm = cm
k k
(1), . . . , cm Gk , β(F, hk , C(k) ) denotes the term depending on the multicode
receiver filter coefficients, the spreading code and the chan-
where n = 0, . . . , Nbit,k−1, m = 0, . . . , Mk−1, and l = 0, . . . , L−1. nel coefficients. pm denotes the power assigned to the mth
0n denotes the null vector of size n. The vector n, of length code. In the sequel, we present the expression of the terms
Nchip + L − 1, represents the channel noise vector with N0 as β(F, hk , C(k) ) and Ω(pm =m ) in the case of the rake, the zero
a power spectral density. forcing, and the MMSE multicode receivers.
The aim of the power control algorithm in CDMA sys-
The vector c(k)
m = [cm (1), . . . , cm (Gk )] denotes the spread-
k k T
tem is to assign the mobile the minimum power necessary to
ing code vector of length Gk related to the kth user. It is
achieve a certain QoS which is typically defined in terms of
obtained by the discretization at the chip rate of the func-
(k) SIR. In this context, the most employed power control algo-
tion cm (t) given by (8). The index m denotes the index of rithm was proposed by Foschini and Miljanic in [10] and it
the spreading code in the multicode scheme containing Mk is known as distributed power control (DPC). The optimum
codes. transmission power of user k, supposed monocode user, is
The model just proposed for a multirate and multicode computed iteratively in order to achieve an SIR target de-
DS-CDMA system follows the structural principles of practi- noted here by SIRtarget .
cal downlink UMTS and leads to a convenient algebraic form
which allows for a powerful receiver design for a multicode
multirate CDMA system. SIRtarget
pk (n + 1) = pk (n). (16)
For the sake of simplicity, the propagation channel is as- SIR(n)
sumed to be time invariant during the transmission of Nchip
chips. We also assume that the interferences due to symbols When the target SIR is achieved, the power’s updating
before and after Nchip data block can be completely cancelled. stops. This approach assumes a fixed receiver, usually a sin-
This is possible when those interfering symbols are known by gle receiver. To overcome this limitation, Ulukus and Yates in
the receiver via a training sequence. The model presented in [11] proposes to optimize jointly the multiuser receiver and
(12) can be generalized to incorporate scrambling codes and the user’s power in the uplink. As the main result, it is shown
multiple antenna transmissions. that the same performance as the DPC algorithm is achieved
with less transmitted power. In continuation of Yates’ idea
of a combined power control and receiver adaptation in a
3. PROBLEM STATEMENT
CDMA uplink, we develop, here, a joint power control and
multicode receiver adaptation algorithm suitable for a high-
Without loss of generality, the user 1 is chosen as the user of
speed UMTS downlink.

(k) , the received signal can
interest. By denoting A(k) = C(k) H
be expressed as So, the problem is to determine the different code pow-
ers, pm , and multicode receiver filter coefficients, such that
the allocated power to the multicode user is minimized

K
r= (1) (1) (1)
A(k) W(k) b(k) + while satisfying the quality of service requirement on each
A W b + n ,
code, SIRm ≥ SIRtarget , where SIRm = En ((SIR(m, n))), m =
k=2
desired signal + intercode interference noise
1, . . . , M1 , and SIRtarget is the minimum acceptable level of
MAI + ISI
(14) SIR for each code. En denotes the expectation over the sym-
bol index. Therefore, the problem can be stated mathemati-
where we separate the user of interest’s signal, the multiple cally as follows:
access interference (MAI), and intersymbol interference (ISI)
caused by the other users and the noise. The first term in
M1
(14) contains the useful signal and the intercode interference min pm (17)
p
caused by the multicode scheme. m=1
constrained to
2 (1) i = 0, start with initial powers p0(1) , . . . , pM
(1)
.
E Ω pm =m
1
(2) Receiver parameter calculation and receiver output SIR
pm ≥ SIRtarget (1) 2 calculation.
E β F, hk , C(k) bm (n) (18) (3) Update the code powers using
(1) (1)
pm ≤ pmax , m = 1, . . . , M1 , pm (i + 1) = (SIRtarget /En [SIR(m, n)])pm (i), for m =
1, . . . , M1 .
(1)
where pmax denoted the maximum allowed transmitted (4) [W(i + 1)] j, j = pm (i + 1), with j = m + (n − 1)M1
user’s power. where m = 1, . . . , M1 and n = 1, . . . , Nbit,1 .
The following optimization problem is difficult since the (5) i = i + 1, stop if convergence is reached; otherwise, go to
constraints denominators are also power dependent. The so- step (2).
lution is to consider a double optimization problem where
an inner optimization is inserted in the constraint set as the
following: Algorithm 1

M1
min pm (19) modified according to
p
m=1
SIR(1)
target
(1)
pm (i + 1) = min p(1) (i), pm
max
. (21)
constrained to En SIR(m, n) m

2
E Ω pm =m The new code power calculated in step (3) are transmitted
pm ≥ SIRtarget min (1) 2 , via a feedback link to the BS.
bm (n)
F E β F, h , C(k) (20) In the sequel, we present the SIR derivation in the case of
k
the zero forcing and the MMSE multicode joint receivers.
pm ≤ pmax , m = 1, . . . , M1 .
In [11], the equivalence between the optimization for- 5. JOINT MULTICODE RECEIVER STRUCTURES
mulation given by (17) and the formulation given by (19)
In this section, we derive the expression of the output SIR on
is demonstrated.
each code by considering the joint multicode receivers: ZF
The second optimization formulation is a two alternate
and MMSE.
optimization problem. The first optimization problem in-
The received signal given by (14) can be written as
volved in (19), and called the outer optimization, is defined
over the code power. Whereas the second one, called the in- r = AWb + n

(22)
ner optimization, which is involved in (20), assumes a fixed K
power vector. It is defined over the filter coefficients of the
=
by denoting n k=2 A
(k) W(k) b(k) + n.
multicode receiver. In this stage, we optimize the multicode
filter coefficients to maximally suppress the intercode inter- 5.1. Rake receiver
ference. The implementation of these two alternate optimiza-
tion problems are realized iteratively in the algorithm de- The conventional data estimator consists of a bank of rake
scribed in the next section. receivers. In this case, the output signal is
yRake = AH r = ΓWb + AH n

, (23)
4. COMBINED DOWNLINK POWER CONTROL
AND JOINT MULTICODE RECEIVERS where Γ = AH A.
We separate the desired user’s symbols, the intercode in-
In this section, we propose to combine the downlink power terference generated by the multicode transmission and the
control and the joint multicode receivers. The objective of MAI + ISI + noise generated by the noise and the other users,
the algorithm is to achieve an output SIR equal to a target
yRake = diag{ΓWb} + diag{ΓWb} + A H

n ,
SIRtarget for each assigned code to the multicode user. To do
MAI + ISI + noise
this, we exploit the linear relationship between the output desired symbols intercode interference
SIR and transmit code power as is seen in (15). The proposed (24)
algorithm is a two-stage algorithm. First, we adjust the filter
coefficients for a fixed code power vector, the inner optimiza- where diag(X) = X − diag(X) represents a matrix with zero
tion. Second, we update the transmitted code powers to meet diagonal elements containing all but the diagonal elements
the SIR constraints on each code for the chosen filter coeffi- of X.
cients using (16). The description of the proposed algorithm The useful signal for the nth transmitted symbol on the
is as follows: mth code is given by
2
The subscript 1 marks out the considered multicode user. 2 (1) 2
If we consider also a maximum transmit power limitation E [ΓW] j, j b1(1) (n) = [ΓW] j, j E b1 (n) ,
pmmax , for m = 1, . . . , M , step (3) from the above algorithm is
1 (25)
where [X] j, j denotes the element in the jth row and jth col- The SIR at the output of the MMSE receiver relating to
umn of the matrix X. the nth transmitted symbol on the mth code can be expressed
The interference and the noise are given by as follows by denoting W = W0 W as
2
I = E ΓWb− diag{ΓWb} + AH n

. (26) SIRMMSE (m, n)
2
[W ] j, j
We consider in the sequel that E{|b1(1) (n)|2 } = 1. = 2
After developing the term I and taking the jth diagonal W W H j, j − [W ] j, j + W0−1 ΓAH Rn
AΓ−1 W0H j, j
element, the SIR at the output of the rake receiver related to (34)
the nth transmitted symbol on the mth code can be expressed
T
as follows by denoting Γ = ΓW and Rn
= E[n
n

] as the for j = m+(n−1)M1 where m = 1, . . . , M1 and n = 1, . . . , Nbit,1 .
covariance matrix of the MAI, ISI and noise, The proposed approach involves complex matrix in-
2 verse computations due to the employment of instantaneous
[Γ ] j, j
SIRRake (m, n) = 2 MMSE filtering. This drawback can be recovered by replac-
(Γ )2 j, j − (Γ ) j, j + Γ Rn
Γ j, j ing instantaneous MMSE filtering with adaptive filtering. As
(27) is suggested in [22], the least mean square and the minimum
output energy algorithms present an ease implementation
for j = m+(n−1)M1 where m = 1, . . . , M1 and n = 1, . . . , Nbit,1 . and analysis. As a future work, we suggest to focus on the
complexity reduction of the proposed approach.
5.2. Joint multicode zero forcing receiver
6. SIMULATION RESULTS
In the case of the joint ZF receiver, the output signal is
Simulation results analyze the performance of the proposed
yZF = Γ−1 yRake = Wb + Γ−1 AH n

. (28) strategy considering the joint multicode MMSE and the joint
The joint ZF receiver leading to the estimate of the de- ZF receivers, and the performance obtained from the con-
sired symbols, b, is called zero forcing since it tries to force ventional power control which assumes a bank of fixed rake
the residual intercode interference to zero. receivers. We compare the different solutions by evaluating
Therefore, the SIR at the output of the joint ZF receiver the total transmit (or mean transmit) power and the SIR (or
relating to the nth transmitted symbol on the mth code can mean SIR) at the mobile receiver.
be expressed as follows: Users are placed randomly in a hexagonal cell with ra-
dius R = 1000 m around the BS. The path-loss exponent is
[W]2j, j taken σ = 4 and no shadowing is assumed. We consider a 6-
SIRZF (m, n) = −1 H (29) path downlink channel. The target SIR is fixed at SIRtarget = 4
Γ A Rn
AΓ−H j, j
(around 6 dB) for all simulations. We consider a number of
for j = m+(n−1)M1 where m = 1, . . . , M1 and n = 1, . . . , Nbit,1 . K = 20 users, among them we have K , K < K multi-
code users. The spreading factor for the single-code users is
Gk = 128 for any k = K , . . . , K. The multicode users has
5.3. Joint multicode MMSE receiver
a spreading gain Gk = 64, k = 1, . . . , K . We fix the user
The joint multicode MMSE linear receiver minimizes the 1 as user of interest. We vary its number of allocated codes
output mean squared error between M1 = 4 and M1 = 64. 1
In Figure 1, we plot the mean SIR, (1/M1 ) M m=1 SIR(m),
2
E FyRake − Wb (30) versus iteration index in the case of M1 = 4 for the con-
ventional power control algorithm (fixed rake receiver) and
with respect to F which yields the proposed strategy which optimizes the joint MMSE and
−1 ZF multicode receiver coefficients. We note the one-iteration
F = W2 ΓH ΓW2 ΓH + AH Rn
A . (31) convergence of the multicode ZF receiver, the fast conver-
gence of the multicode MMSE receiver, and the much slower
Therefore, the output signal from the MMSE receiver yields, convergence of the rake receiver.
by denoting W0 = FΓ, In the case of M1 = 16, the conventional rake receiver
cannot meet the target SIR anymore, as shown in Figure 2,
yMMSE = FyRake = W0 Wb + W0−1 ΓAH n

. (32)
where we plot the variation of the SIR(m) on each code.
Now, we can separate the desired user’s symbols, the in- However, the multicode receivers (ZF and MMSE) show
tercode interference generated by the multicode transmis- good performance. Adding more virtual users brings the
sion and the MAI + ISI + noise generated by the noise and the conventional receiver to even worse performance as is shown
other users, in Figure 3.

For M1 = 64, the different lines for each receiver type
yMMSE = diag W0 Wb + diag W0 Wb + W0 Γ−1 AH AH n

. correspond to the variation of the SIR on each code, SIR(m),
(33) versus iteration index.
4.5 4.5
Output SIR on each code, m = 1 . . . M1

4 3.5
3
3.5
Mean SIR
2.5
2
3
1.5
1
2.5
0.5
2 0
2 4 6 8 10 12 14 2 4 6 8 10 12 14
Iteration index Iteration index
SIRRake SIRRake
SIRZF SIRZF
SIRMMSE SIRMMSE
Figure 1: The SIR convergence for the rake, ZF, and MMSE re- Figure 2: The SIR convergence for the rake, ZF, and MMSE re-
ceivers in the case M1 = 4 multicode. ceivers in the case M1 = 16 multicode.
From Figures 2 and 3, we observe the difficulty of the 4.5

conventional power control to reach the target SIR because 4
Output SIR on each code, m = 1 . . . M1
of the MAI, ISI, and the intercode interferences. In the case

of low load in the cell (few users), the conventional power 3.5
control reaches the SIR target; see Figure 1. However, in this 3
case, our proposed strategy presents a faster convergence.
The variation of the base station transmit power ra- 2.5
tios pZF / pRake and pMMSE / pRake versus the iteration index is 2
shown in Figure 4 in the case of a number of codes M1 = 16
codes of the multicode user. We note a decrease of about 20% 1.5
of the transmitted BS power. 1
However, a much significant gain in transmitted BS pow-
er is noted in the case of M1 = 64, as we can deduce from the 0.5
results of Figure 5. The MMSE shows its optimality with sig- 0
nificantly improved results with respect to the ZF receiver: 2 4 6 8 10 12 14
the MMSE always gains power with respect to the rake re- Iteration index
ceiver (the ratio is smaller than 1) where the ZF increases first
SIRRake
the required power to achieve the required SIR.
SIRZF
We observe from Figures 4 and 5 that the proposed strat- SIRMMSE
egy of joint downlink power control and multicode receivers
outperforms the conventional downlink power control in
Figure 3: The SIR convergence for the rake, ZF, and MMSE re-
terms of total transmitted power of the multicode user. ceivers in the case M1 = 64 multicode.
In all simulations, we note the very fast (1 iteration) con
vergence of the ZF receiver, the fast convergence of the
MMSE receiver, and the much slower convergence of the
conventional power control. The fast convergence of the ZF fixed receiver that takes into account only the desired signal
receiver is easy to explain: since this receiver performs an or- processing the MAI, ISI, and intercode interferences as noise,
thogonal projection into the subspace formed by the inter- therefore yielding the worst performance.
fering signals, the output desired signal does not depend on The best performance in minimizing transmit powers
the interfering signals’ amplitudes. There is only one update and maximizing the cell capacity is obtained by the MMSE
of (21). In the case of the joint multicode MMSE receiver, at receiver. The ZF receiver shows slightly lower performance,
each iteration the receiver is updated since it depends on the in terms of total transmit power, at high-cell loads (case of
received powers of each code. Finally, the rake receiver is a M1 = 64, see Figure 5).
1.02 1.15
1 1.1
0.98 1.05
Transmit power’s ratio
Transmit power’s ratio

0.96 1
0.94 0.95
0.92 0.9
0.9 0.85
0.88 0.8
0.86 0.75
0.84 0.7
0.82 0.65
2 4 6 8 10 12 14 2 4 6 8 10 12 14
Iteration index Iteration index
pZF / pRake pZF / pRake
pMMSE / pRake pMMSE / pRake
Figure 4: The mean total transmit powers ratio pZF / pRake and Figure 5: The mean total transmit power ratio pZF / pRake and
pMMSE / pRake versus the iteration index for M1 = 16. pMMSE / pRake versus the iteration index for M1 = 64.
It should be noticed that at very low-cell loads (i.e., few Transmit powers on each code, MMSE receiver
interfering single-code users and few codes for the multicode 75
user (case of M1 = 4)) the three receivers show similar per- 74.5
formance, a result that is expected.
Power in dBm on each code
After the convergence of the proposed strategy using a 74

joint multicode MMSE receiver, the codes’ power alloca- 73.5
tion is shown in Figure 6. As one can notice, it is not the
same power per code. This confirms the interest of this 73
power allocation-strategy for the downlink of the multicode
72.5
user.
72
7. CONCLUSION 71.5
In this paper, we have analyzed the benefits of combining 71

the downlink power control and the joint multicode detec-
tion for a multicode user. The proposed algorithm updates 2 2.5 3 3.5
iteratively the transmitted code powers of the multicode Iteration index
users and the joint multicode receiver filter coefficients. We
have used simulations to show the convergence and perfor- Figure 6: The code power allocation in the case of M1 = 10 codes
mance of the proposed algorithm in a system of practical in- after convergence.
terest. An important gain in transmit power reduction is ob-
tained by implementing joint multicode detection. The per- ambitious challenge to be met by third-generation systems
formance of the ZF receiver allows an important reduction in order to provide high-capacity flexible services.
in computations (step 4 is avoided). The study of theoretical
convergence of the proposed algorithm is under investigation REFERENCES
based on the analysis proposed in [23].
In order to overcome the limitation of power control due [1] H. Holma and A. Toskala, Eds., WCDMA for UMTS-Radio Ac-
to temporal filtering only, a joint power control and beam- cess for Third Generation Mobile Communications, John Wiley
& Sons, New York, NY, USA, 2000.
forming for wireless network is proposed in [17] where it is [2] 3GPP TR 25.858 V5.0.0 (2002-03), “High Speed Downlink
shown that a capacity increase is possible if array observa- Packet Access: Physical layer aspects, (Release 5)”.
tions are combined in the MMSE sense. Therefore, as a di- [3] B. Sayadi and I. Fijalkow, “Joint detection for multicode trans-
rection for further research, the combination of the three ba- mission in downlink high speed UMTS,” in Proceedings of 60th
sic interference cancelation approaches (transmit power con- IEEE Vehicular Technology Conference (VTC ’04), vol. 2, pp.
trol, multiuser detection, and beamforming) represents an 837–840, Los Angeles, Calif, USA, September 2004.
[4] M. Saquib, R. D. Yates, and A. Ganti, “Power control for an [20] 3GPP TS 25.215 V6.3.0 (2005-06), “Physical Layer - Measure-
asynchronous multirate decorrelator,” IEEE Transactions on ments (FDD), (Release 6)”.
Communications, vol. 48, no. 5, pp. 804–812, 2000. [21] A. Aguiar and J. Gross, “Wireless channel models,” Tech.
[5] R. D. Yates, “A framework for uplink power control in cellular Rep. TKN-03-007, Telecommunications Networks Group,
radio systems,” IEEE Journal on Selected Areas in Communica- Technische Universität Berlin, Berlin, Germany, April 2003.
tions, vol. 13, no. 7, pp. 1341–1347, 1995. [22] C.-L. Wang, M.-H. Li, K.-M. Wu, and K.-L. Hwang, “Adap-
[6] A. Sampath, P. S. Kumar, and J. M. Holtzman, “Power control tive interference suppression with power control for CDMA
and resource management for a multimedia CDMA wireless systems,” in Proceedings of IEEE International Symposium on
system,” in Proceedings of 6th IEEE International Symposium Circuits and Systems (ISCAS ’01), vol. 4, pp. 286–289, Sydney,
on Personal, Indoor and Mobile Radio Communications, Wire- NSW, Australia, May 2001.
less: Merging onto the Information Superhighway (PIMRC ’95), [23] J. Luo, S. Ulukus, and A. Ephremides, “Probability one con-
vol. 1, pp. 21–25, Toronto, Ontario, Canada, September 1995. vergence in joint stochastic power control and blind MMSE
[7] V. V. Veeravalli and A. Sendonaris, “The coverage-capacity interference suppression,” in Proceedings of 37th Conference on
tradeoff in cellular CDMA systems,” IEEE Transactions on Ve- Information Sciences and Systems, The Johns Hopkins Univer-
hicular Technology, vol. 48, no. 5, pp. 1443–1450, 1999. sity, Baltimore, Md, USA, March 2003.
[8] L. C. Yun and D. G. Messerschmitt, “Variable quality of service
in CDMA systems by statistical power control,” in Proceedings
of IEEE International Conference on Communications, Gateway Bessem Sayadi received the B.S. Engineer-
to Globalization, vol. 2, pp. 713–719, Seattle, Wash, USA, June ing degree in signal processing from the
1995. Ecole Supérieure des Télécommunications
[9] S. V. Hanly and D.-N. Tse, “Power control and capacity de Tunis (Sup’Com Tunis), Tunisia, in 1999,
of spread spectrum wireless networks,” Automatica, vol. 35, and both the M.Phil. (2000) and the Ph.D.
no. 12, pp. 1987–2012, 1999. (2003) degrees from the Signals and Systems
[10] G. J. Foschini and Z. Miljanic, “A simple distributed au- Laboratory (LSS) at Supélec, Gif-sur-Yvette,
tonomous power control algorithm and its convergence,” IEEE the Paris XI University, Orsay, France. In
Transactions on Vehicular Technology, vol. 42, no. 4, pp. 641– 1999, he joined France Telecom where he
646, 1993. was engaged in research on echo cancelation
[11] S. Ulukus and R. D. Yates, “Adaptive power control with and adaptive filtering. He has also served as a Teaching Assistant in
MMSE multiuser detectors,” in Proceedings of IEEE Interna- several courses on digital communications, signal processing, and
tional Conference on Communications, vol. 1, pp. 361–365, electronics in the Department of Electronic and Electrical Engi-
Montreal, Quebec, Canada, June 1997. neering, SUPÉLEC, ENSEA, and University Parix IX, since Septem-
[12] J. G. Andrews, A. Agrawal, T. H. Meng, and J. M. Cioffi, “A ber 2000. From 2003 to 2005, he was an Associate Researcher in
simple iterative power control scheme for successive inter- the Image and Signal Processing Team (ETIS), at ENSEA, Cergy-
ference cancellation,” in Proceedings of 7th IEEE International Pontoise. In 2006, he joined France Telecom as a Research Engineer.
Symposium on Spread Spectrum Techniques and Applications, His current research interests include Bayesian method, multiuser
vol. 3, pp. 761–765, Prague, Czech Republic, September 2002. detection, video coding, radio resource management, IP-mobility,
[13] F. Meshkati, D. Guo, H. V. Poor, S. C. Schwartz, and N. B. Man- and cross-layer design.
dayam, “A unified approach to power control for multiuser
detectors,” in Proceedings of the 2nd International Workshop on Stefan Ataman received the B.S. and M.S.
Signal Processing for Wireless Communications, King’s College, degrees from the Polytechnic University of
London, UK, June 2004. Bucharest, Romania, in 1999 and 2000,
[14] F. Meshkati, H. V. Poor, S. C. Schwartz, and D. Guo, “A respectively, and the Ph.D. degree from
unified power control algorithm for multiuser detectors in Université Paris-Sud, France, in 2004, all
large systems: convergence and performance,” in Proceedings of in electrical engineering. Currently, he
the 43rd Allerton Conference on Communications, Control and is working as a Research Associate with
Computing, Urbana-Champaign, Ill, USA, September 2005. University Cergy-Pontoise/ETIS laboratory,
[15] D. Guo and S. Verdú, “Randomly spread CDMA: asymptotics France. His research interests are in the ar-
via statistical physics,” IEEE Transactions on Information The- eas of digital communications and signal
ory, vol. 51, no. 6, pp. 1983–2010, 2005. processing with applications to CDMA wireless communications,
[16] F. Rashid-Farrokhi, K. J. Ray Liu, and L. Tassiulas, “Downlink power control, and multiuser receivers in CDMA cellular systems.
power control and base station assignment,” IEEE Communi-
Inbar Fijalkow received her Engineering
cations Letters, vol. 1, no. 4, pp. 102–104, 1997.
and Ph.D. degrees from Ecole Nation-
[17] F. Rashid-Farrokhi, L. Tassiulas, and K. J. Ray Liu, “Joint op- ale Supérieure des Télécommunications
timal power control and beamforming in wireless networks (ENST), Paris, France, in 1990 and 1993,
using antenna arrays,” IEEE Transactions on Communications, respectively. In 1993–1994, she was a Re-
vol. 46, no. 10, pp. 1313–1324, 1998. search Associate at Cornell University, NY,
[18] J.-W. Lee, R. R. Mazumdar, and N. B. Shroff, “Downlink power USA. In 1994, she joined ETIS, UMR 8051
allocation for multi-class wireless systems,” IEEE/ACM Trans- (ENSEA - Cergy-Pontoise University -
actions on Networking, vol. 13, no. 4, pp. 854–867, 2005. CNRS) in Cergy-Pontoise, France. Since
[19] L. Song and J. M. Holtzman, “CDMA dynamic downlink 2004, she is the head of ETIS. Her cur-
power control,” in Proceedings of 48th IEEE Vehicular Technol- rent research interests are in signal processing applied to dig-
ogy Conference (VTC ’98), vol. 2, pp. 1101–1105, Ottawa, On- ital communications: iterative (turbo) processing (in particular
tario, Canada, May 1998. turbo-equalization), analysis of communication systems (including
MIMO, OFDM, CDMA, etc.) and cross-layer optimization. Until

2005, she has been Member of the board of the GDR ISIS, which is
the CNRS French national research group on signal, image, and
vision processing. She has been an Associate Editor of the IEEE
Transactions on Signal Processing 2000–2003.
DOI 10.1155/WCN/2006/17526
Adaptive Downlink Resource Allocation Strategies for

Real-Time Data Services in OFDM Cellular Systems
Navid Damji and Tho Le-Ngoc
Department of ECE, McGill University, 3480 University street, Montréal, Québec, Canada H3A 2A7
Received 1 October 2005; Revised 25 February 2006; Accepted 21 March 2006

This paper presents a detailed performance analysis of adaptive downlink resource allocation based on users’ instantaneous channel
responses using power minimization (PM) and bandwidth constrained power minimization (BCPM) strategies. This study shows
that, in cellular systems, where interference is a dominant factor, the link outage performance of a resource allocation strategy
varies significantly depending on the user channel parameters. In particular, both analytical and simulation results indicate that
the PM strategy outperforms BCPM in a mild shadowing environment. However, in severe shadowing conditions, this trend is
reversed. This assessment holds true for both flat and frequency-selective fading.
Copyright © 2006 N. Damji and T. Le-Ngoc . This is an open access article distributed under the Creative Commons Attribution
cited.
1. INTRODUCTION efficiently use the resources such as bandwidth, power and

modulation to increase the spectral efficiency of the system.
The next generation of mobile communications is envisioned Resource allocation for multiuser OFDM has been given
to offer a multitude of services that are available and acces- much attention in the literature. In [2], the authors proposed
sible anywhere and anytime. With the introduction of new an optimization criterion to minimize the transmitted power
multirate, multi-QoS services, the future networks should be while satisfying the rate requirements of the users in the sys-
designed for economic packet data transfers [1]. These ser- tem. The authors in [3] give an alternative formulation to
vices are highly asymmetrical and require high transmission maximize the rates of the users while satisfying power con-
bandwidth on the downlink. However, due to the limitations straints. A more generalized formulation in terms of max-
of the available frequency spectrum, its efficient use is cru- imizing the average utility function of the users is given in
cial to the success of the next generation wireless networks. [4] and algorithms are proposed in [5]. The above described
Novel access methods coupled with adaptive resource man- formulations are pertinent to an orthogonal frequency di-
agement techniques, specifically for downlink transmission, vision multiple access (OFDMA) type system in which fre-
require more attention in order to improve the spectral effi- quency division is used as a multiple access mechanism. Al-
ciency [1]. ternatively, spread-spectrum (SS) techniques, such as code
Orthogonal frequency division multiplexing (OFDM) is division multiple access (CDMA), can be used as an access
potentially suited for supporting high-speed downlink trans- mechanism over OFDM. In this field, several resource alloca-
mission as it can offer high spectral efficiency due to its ro- tion algorithms have been proposed that allocate subcarriers
bust performance over heavily impaired links. OFDM has and codes to mitigate the effects of frequency-selective fading
been demonstrated as an efficient way to mitigate the adverse [6] and Doppler spread for high mobility [7]. Most of the al-
effects of frequency selective multipath fading by transmit- gorithms proposed above can be applied in a single-cell sys-
ting signals over a number of flat-faded narrow-band chan- tem since interference for other cells is not considered. How-
nels. The inherent multicarrier nature of OFDM also allows ever, in a multicell environment, the inter-cell interference
the use of adaptive modulation and power allocation accord- has a significant impact on the performance of the system.
ing to the responses of the narrow-band channels, which can In this scenario, the power minimization approach as pro-
significantly enhance the system performance. In order to posed in [2] is a logical candidate for resource allocation. The
exploit fully the advantages of OFDM in cellular systems, objective of power minimization aims to reduce the trans-
dynamic allocation techniques need to be devised, which mission (and hence interference) power on all subcarriers,
which results in using the lowest possible modulation, and the link outage probability in flat fading environment with
hence a larger transmission bandwidth (i.e., number of sub- different levels of shadowing, while Section 4 gives perfor-
carriers) to meet the rate constraint. In a distributed cellular mance results of the two strategies. Section 5 briefly describes
system, where the base-stations do not know each other al- the algorithms that can be used to solve the PM and BCPM
locations in advance, this allocation strategy leads to using a optimization problems. Finally, in Section 6, the perfor-
larger proportion of bandwidth, which increases the proba- mance of the two schemes (using the algorithms in Section 5)
bility of experiencing interference in other cells. This in it- is evaluated for real-time data services in different shadowing
self may not be bad, since if the power of interference is low environments.
compared to the signal power (i.e., the channel response of
the interference is much lower than that of the signal), then 2. RESOURCE ALLOCATION IN OFDM SYSTEMS
the users will still be able to decode the data with low BER.
However, when the power of interference is high, the user The aim of resource allocation is to deliver the traffic of mul-
will experience outage with high probability since the inter- tiple users within the given system bandwidth while meeting
ference is present in a large fraction of the bandwidth. In this the QoS requirements. The users’ traffic translates to a given
scenario, it may make sense to limit the transmission band- rate requirement, and the bandwidth of the system is given
width of the users, such that if the subcarriers are chosen in a by the total number of subcarriers used in the cell, each hav-
random manner, the probability of interference will be lower ing a fixed symbol rate. The rate constraint can be met us-
than in the previous case. This implies that, in order to sat- ing adaptive modulation (also referred to as bit loading) on
isfy the user rate requirements, higher modulation and hence all subcarriers such that the total rate transmitted on all the
higher transmission power will be used on these subcarri- subcarrier equals the total rate requirement of the users. The
ers. QoS on the other hand is greatly affected by time-varying
A simple resource allocation scheme for the downlink channel conditions in a wireless cellular environment. In ad-
OFDM cellular system known as the best subcarrier allo- dition, cochannel interference is a major factor that affects
cation (BSA), presented in [8], aims to minimize the re- the link outage probability. From this perspective, one strat-
quired number of subcarriers for transmission, and then the egy of resource allocation would be to transmit the lowest
transmission power. A similar formulation was also given amount of power while satisfying the rate constraints, which
in [9] which is more suitable to point-to-point networks. A is the power minimization strategy. The other is to minimize
more detailed problem formulation and analysis was given the probability of experiencing interference, which is the ba-
in [10] and the BSA algorithm was refined to a more opti- sis of using minimum bandwidth (BCPM) required to satisfy
mal strategy known as bandwidth constrained power mini- user constraints. The mathematical formulation of the two
mization (BCPM). It was shown that in a severe shadowing problems is given as follows.
environment, BCPM gives a better outage performance than
PM since it keeps the probability of interference to a mini- 2.1. Power minimization (PM)
mum.
The objective of this paper is to present a more general The aim of power minimization is to use the least amount of
analysis of the two schemes discussed above, in order to have power to deliver all of the users’ traffic in the given band-
an insight into how the two schemes perform in different width. The optimization problem can be mathematically
environments. Specifically, under what channel conditions stated as follows:
would the one scheme perform better than the other. For this

N
K
N
K
fP cn,k
purpose we develop analytical models of link outage for both min pn,k = ρn,k (1)
severe and nonsevere shadowing conditions in flat fading n=1 k=1 n=1 k=1
γn,k
conditions. The impact of using lower power (lower modula-
tion and larger bandwidth) versus lower bandwidth (higher subject to
modulation and higher power) is studied as the load per cell
cn,k ∈ 0, c0 , c1 , . . . , cη , where ci < ci+1 ; (2)
is increased for the two cases. It is shown in Section 4 that
in less severe shadowing environments, using lower power
K
ρn,k cn,k ≥ bn ∀n ∈ {1, 2, . . . , N }; (3)
(basis of PM) performs better than using smaller bandwidth
k=1
(basis of the BCPM strategy), whereas the opposite is true
for severe shadowing. In Section 6, the performance of the
N
ρn,k ≤ 1 ∀k ∈ {1, 2, . . . , K },
PM and BCPM strategies is given for a more realistic de- n=1
ployment scenario that uses multiple antenna beams per sec- ⎧ (4)
⎨0 if cn,k = 0,
tor to isolate the effects of interference, using real-time data ρn,k = ⎩
models [11]. Both severe and nonsevere shadowing environ- 1 if cn,k > 0;
ments are simulated and the same performance trend is ob-
served. where
The remainder of the paper is organized as follows. In (i) n is user index,
Section 2, problem formulation is given for the two strate- (ii) N is total number of users,
gies. Section 3 discusses the analytical model used to evaluate (iii) k is subcarrier index,
N. Damji and T. Le-Ngoc 3
(iv) K is total number of subcarriers, Mj

(v) pn,k = ρn,k fP (cn,k )/γn,k is power on the kth subcarrier rj j
of the nth user with cn,k bits, θj j
(vi) cn,k is the bit loading level corresponding modulation Bj
and coding scheme, ri j
(vii) fP (·) is power function corresponding to the bit load- Mi
ing level cn,k , Di j
rii
(viii) γn,k is channel attenuation on the kth subcarrier of the θii
ϕi j
nth user,
Bi R
(ix) bn is the number of bits per OFDM symbol required
by the nth user.
Constraints (2)–(4) in the above-stated problem formulation
aim to satisfy the user rate requirements to achieve the objec-
tive of minimum power. In order to meet the required rates Figure 1: Interaction of two cells.
in constraint (3) with the smallest power, the selection of bit
loading levels, cn,k , in constraint (2) tends to use low values
ci and hence, a large amount of subcarriers. As a result, this 3. EFFECTS OF INTERFERENCE IN FLAT
increases the occurrence of interference in subcarriers. FADING ENVIRONMENT
For PM strategy, the goal would be to minimize the modu-

2.2. Bandwidth constrained power lation on each subcarrier and, as a consequence, the number
minimization (BCPM) of subcarriers is increased to support the rate requirement.
Whereas for BCPM strategy, the aim would be to transmit
The BCPM strategy aims to use the minimum number of in the least number of subcarriers by increasing modula-
subcarriers to satisfy the rate requirement with minimum tion level in each subcarrier. In order to understand the im-
power. Alternatively, the problem can be stated as minimiz- pacts of the two strategies on the system performance, we de-
ing power while satisfying the user rate requirements with the velop an analytical framework to evaluate the performance
smallest possible number of subcarriers. Consider the rate re- of the users in terms of expected link outages as a function
quirement of the user n represented by the required number of the system load in terms of the number of users, and
of bits per OFDM symbol, bn . The smallest possible num- the number of subcarriers used to deliver the rate require-
ber of subcarriers, Smin
n , to satisfy bn is obtained by using the ment. The framework of outage probability calculation in-
highest bit loading level, cη , that is, Smin
n = bn /cη . There-
volves accounting for flat fading and shadowing with cochan-
fore, to satisfy the user rate requirements with the small- nel interference. Given the number of users in the system, the
est possible number of subcarriers we can add another con- number of subcarriers allocated per user and the signal-to-
straint for the new bandwidth-constrained power minimiza- interference ratio (SIR) required for the modulation/coding
tion (BCPM), stated as follows: scheme in use, the analysis gives the expected outage proba-
bility. In this section, we show that the relative performance

N
K
N
K
ρn,k fP cn,k of PM and BCPM strategies highly depends on the level of
min pn,k = (5)
n=1 k=1 n=1 k=1
γn,k shadowing. Section 3.1, gives an overview of the system and
the probability density functions (pdf) that are required to
subject to derive the expected outage for a given number of users. These
results will be used in Section 3.2 where the overall expected

cn,k ∈ 0, c0 , c1 , . . . , cη , where c j < c j+1 ; (6) outage probability is derived.

K
ρn,k cn,k ≥ bn ∀n ∈ {1, 2, . . . , N }; (7) 3.1. System model
k=1

N We consider a cellular system with 3 sectors/cell, assuming
ρn,k ≤ 1 ∀k ∈ {1, 2, . . . , K }; that the path-loss model has the following general form:
n=1
⎧ (8) −β
⎨0 PLi j = ri j , (10)
if cn,k = 0,
ρn,k =
⎩1 if cn,k > 0; where i is the index of the ith base-station and j is the index

of the jth user, and ri j is jth user’s distance from the ith base

K
bn
ρn,k ≤ Smin
n = ∀n ∈ {1, 2, . . . , N }. (9) station. Let Dik be the distance and φik be the angle between
k=1
cη base-station i and base-station k, and R be the cell radius.
These relations are shown in Figure 1. Note that the subscript
It is noted that constraint (9) in the BCPM problem may in- ii represents the case where both the base-station and the user
crease the minimum power as compared to that in the PM are of the same cell. Hence, PL j j is the pathloss, r j j is the
problem. distance of user j from its cell’s base-station, and θ j j is the
angle of the user with respect to his base-station (as indicated the same subcarrier as the user 0, the SIR can be given as
in Figure 1).
pt00 · ps00
The distance from the ith base-station to the user j can SIR | n =
be expressed in terms of Di j , φi j , r j j and θ j j as follows: pcn0
(14)
1
n
= , where pcn0 = ln e pci0
.
ri2j = r 2j j + Di2j − 2r j j Di j cos φi j + π − θ j j . (11) pcn0 i=1
The probability of outage given n interfering users can be ex-

In presence of Rayleigh fading and log-normal shadowing, pressed as
the pdf of instantaneous received power psi j from the ith

base-station to the jth user can be given as [12] the following 1
Pr outage | n = Pr(SIR < z) = Pr pcn0 >
equation assuming the transmission power of base-station i z
is 0 dB: ∞ (15)

= f pcn0 pcn0 ∂pcn0 ,
∞ 1/z
1
f psi j psi j = √ exp − pli j exp − psi j exp − pli j where z is the minimum SIR protection ratio required for
2πσ −∞
2 reliable communications. Note that z is different for different
pli j − mi j
× exp − ∂pli j , modulation/coding schemes used in the system.
2σ 2 For severe shadowing conditions, where the individual
(12) interference components are approximated by log-normal
distributions, the pdf of pcn0 can be approximated by another
where pli j is the logarithmic local mean power with logarith- log-normal distribution, that is,
mic area mean power mi j and logarithmic standard devia- 2
1 ln pcn0 − mc
n0
tion σ expressed in natural units. The logarithmic area mean f pcn0 pcn0 = √ exp − 2
,
power is given as mi j = −β ln(ri j ) and the σ is related to deci- 2π σcn pcn0 2σ cn
bel standard deviation σdB as eσ = 10σdB /10 . (16)
In the PM and BCPM strategies, the power control is per-
formed on the instantaneous channel response, which has where the mean mc n0 and standard deviation σcn can be de-
both the fading and shadowing components (as described rived by using the method proposed by Schwarz and Yeh
above). The first step is to derive the pdf of the received power [14]. Hence the outage probability given users n can be stated
of one interferer, given that the signal power is normal- as follows:
ized at 1 (0 dB). Appendix A presents the derivation [10] for
Pr outage | n
shadowing standard deviation larger than 10 dB, by approxi-
∞ 2
mating the received interference as a log-normal distributed 1 pcn0 − mc
n0
component [12]. Unfortunately, the results become inaccu- = √ exp − ∂ pcn0 .
1/z 2π σcn pcn0 2
rate for lower shadowing standard deviation. For less severe 2σ cn
shadowing, an alternative method based on Padé approxima- (17)
tion [13] to derive the pdf of the received interference power
For less severe shadowing conditions, the interference is a
from one user is given in Appendix B. Our derivation shows
mixture of Pareto-distributed components, and a closed-
that the approximated distribution has a form of a mixture of
form pdf of the sum in (14) is difficult to find analytically
Pareto distributions. Finally, we give the distribution of users
by using the Padé approximation. An alternative approach is
within a cell. If the users are uniformly distributed in the sec-
to approximate the sum of independent Pareto random vari-
tor then the pdf of rii and θii for all i can be given by
ables by the pdf of the largest one, since it is the dominant
⎧
term in the sum. The cdf of the maximum of the distribu-
⎨ 2rii
⎪
0 < rii ≤ R,
tions is readily given as follows:
frii rii = ⎪ R2
⎩0
n
elsewhere, F pcn0 (pcn0 )
F pmax = F pci0 (pci0 )
⎧ (13)
⎪ 3 2π i=1
⎨ 0 < θii ≤ ,
fθii θii = ⎪ 2π 3 n
M
M −1
⎩0 qms
elsewhere. = λms λlt 1 − pci0 + 1 .
i=1 l=0 m=0
qlt
(18)
3.2. Expected outage probability
The outage probability for the n interferers can then be given
Outage occurs when the desired signal does not meet the
simply by using the following expression:
required SIR for reliable communications. The interference

arises from the closest users that use the same subcarrier. Let 1
j = 0 be the user of interest. Given n interfering users using Pr outage | n = 1 − F pcn0 . (19)
z
Table 1: Modulation/coding schemes and parameters.
Modulation and coding schemes Bandwidth efficiency Nscu for 20 kbps Nscu for 120 kbps zdB required at BER = 10−6
rate-1/2 QPSK 1 bps/Hz 2 12 2 dB
rate-1/2 16QAM 2 bps/Hz 1 6 7.01 dB
rate-3/4 16QAM 3 bps/Hz — 4 11.17 dB
The expressions of outage above are valid for a given set 20 kbps (low-rate scenario). In this case, for each user, we
of mean powers of transmit and receive power of interfer- can use 2 rate-1/2 QPSK subcarriers with a required SIR of
ing users, which in turn depend on the distance of user i 2 dB or 1 rate-1/2 16-QAM subcarrier with a required SIR
and user 0 from their corresponding base stations. Hence, of 7.01 dB. In the second scenario, each user needs 120 kbps
this expression is conditioned on the distance vector R = (high-rate scenario). This can be supported by 12 rate-1/2
{r00 , r11 , . . . , rnn } and θoo . In order to get the average outage QPSK, 6 rate-1/2 16-QAM, or 4 rate-3/4 16-QAM subcarri-
probability, we generate M samples of the vector R chosen ers with increased SIR requirements. Table 1 summarizes the
from frii (rii ) and θoo chosen from fθoo (θoo ), and average the modulation and coding techniques taken from [15] with the
result. corresponding rate and SIR protection ratio for reliable com-
The next step is to derive the user outage probability in munication (BER = 10−6 ), and also gives the number of sub-
terms of system load (represented by the number of users carriers required by each modulation for the low- and- high
N in each sector) and the number of subcarriers Nscu al- rate scenarios. The analysis is carried out for σdB = 5 dB and
located per user. Define the load per subcarrier per sector, 10 dB.
Ls = NNscu /K, where K is the number of subcarriers per sec- The analysis is also verified by simulations with a 3-
tor, and T as the total number of interfering sectors, with sector, 19-cell system shown in Figure 2(a) in the presence of
I = {It , 1 ≤ t ≤ T } as the set of indexes of the interfering sec- Rayleigh flat fading and path-loss exponent β = 3.5. Within
tors. Let the system states S j , 0 ≤ j ≤ 2T −1, represent all pos- one cell, each sector has a hard division amongst the subcar-
sible combinations of the elements of I, which are currently riers to allocate so that they do not interfere with each other.
interfering with user 0, for example, S0 = {∅}, S1 = {I1 }, For analytical results, only 2 first-tier cells and 5 second-tier
S2 = {I2 }, . . . , ST+1 = {I1 , I2 }, S2T −1 = {I1 , I2 , . . . , IT }. As- cells shown in Figure 2(b) are taken into account since they
sume that the bandwidth is divided into K/Nscu contiguous contain the dominant interferers.
sets of Nscu subcarriers each. For uniform distribution in al- Figures 3(a) and 3(b) show the results for low-rate
locating subcarrier set, the probability of being in state S j,k (20 kbps) and high-rate (120 kbps) scenarios, respectively.
is Solid lines indicate the analytical results for severe shadowing
n j T −n j with σdB = 10 dB, while dashed-dotted lines represents the
Pr S j = Ls 1 − Ls , (20)
mild shadowing with σdB = 5 dB.
where n j is the number of interferers in the state S j . It follows For both low- and high-rate data scenarios, in the case
that Pr{outage | S j } = Pr(outage | n j ) given in (17) and (19) of mild shadowing with standard deviation of 5 dB, the PM
and the total probability of outage is approach using low bandwidth-efficient modulation/coding
T −1
schemes (e.g., rate-1/2 QPSK) provides better performance
2
(i.e., lower outage probability) than the BCPM approach.
Pr{outage} = Pr outage | S j Pr S j . (21)
j =0
However, for severe shadowing, the trend is reversed. This
can be explained by the fact that, for the same margin sep-
The above derivation gives the analytical framework for the arating the mean signal and interference powers, the chance
performance evaluation of the PM and BCPM criterions. of crossing the protection ratio (i.e., margin) in mild shad-
Power minimization alone tends to increase the number of owing scenarios is smaller than that in the case with severe
subcarriers in order to reduce the required modulation level shadowing (i.e., large shadowing variance). Hence, minimiz-
and hence the transmitted power; whereas bandwidth min- ing power alone in the severe shadowing case is not enough
imization in order to reduce the probability of interference to guarantee low link outage. Instead, reducing the proba-
tends to reduce the number of subcarriers used and increase bility of interference by using minimum bandwidth yields a
the modulation levels and the corresponding power. higher protection against outage. On the other hand, in a low
shadowing scenario, as the minimum power approach al-
4. PERFORMANCE IN FLAT FADING AND ready gave sufficient protection against outage, using BCPM
SHADOWING ENVIRONMENT approach may increase the interfering power level that leads
to higher outage.
Consider an OFDM system with 180 subcarriers (60 subcar- Figure 3(b) shows that, in the higher-rate case (120
riers per cell). Each subcarrier has a bandwidth of 10 kHz kbps), for mild shadowing with σdB = 5 dB, going from
and can ideally support 10ksps. We investigate the perfor- 12 to 6 subcarriers does not have as drastic an impact on
mance of the PM and BCPM strategies for the two follow- performance as from 6 to 4 subcarriers. This implies that
ing traffic scenarios. In the first scenario, each user requires there is a limit on the SIR threshold, which would severely
(a) Simulation system with 19 cells, 2- (b) Analytical model with 2 first-tier
tier interferers. and 5 second-tier interferers.
Figure 2: Cellular system: simulation and analytical models.
100
Expected outage probability
Outage probability
10−1
10−1
10−2 10−2
6 12 18 24 30 36 42 48 54 60 3 6 9 12 15
Users per cell Users per cell
QPSK 0.5/5 dB analysis QPSK 0.5/10 dB analysis 16QAM 0.75/5 dB analysis 16QAM 0.75/10 dB analysis
QPSK 0.5/5 dB simulation QPSK 0.5/10 dB simulation 16QAM 0.75/5 dB simulation 16QAM 0.75/10 dB simulation
16QAM 0.5/5 dB analysis 16QAM 0.5/10 dB analysis 16QAM 0.5/5 dB analysis 16QAM 0.5/10 dB analysis
16QAM 0.5/5 dB simulation 16QAM 0.5/10 dB simulation 16QAM 0.5/5 dB simulation 16QAM 0.5/10 dB simulation
QPSK 0.5/5 dB analysis QPSK 0.5/10 dB analysis
QPSK 0.5/5 dB simulation QPSK 0.5/10 dB simulation
(b) Outage probability for 120 kbps.
(a) Outage probability for 20 kbps.
Figure 3: Performance comparison of various modulation/coding schemes.
degrade the performance if crossed. The results support the The advantage of the analytic model presented in this pa-
power minimization argument that keeping lower modu- per is that it can give insight into the performance of resource
lation (with lower SIR thresholds) would yield better per- allocation algorithms without performing extensive simula-
formance for mild shadowing cases. The reverse situation tions. In fact, as will be seen in Section 6 for performance
for the case of severe shadowing with σdB = 10 dB is also of PM and BCPM strategies in frequency selective fading,
shown in Figure 3(b): moving from 4 to 6 subcarriers has the same performance trend with respect to shadowing is
smaller performance penalty than from 6 to 12 subcarriers. observed. However, for frequency selective channel models
This is consistent with the argument that the probability of and user traffic models, simulation is necessary for exact per-
interference is the dominant factor indicating the system performance evaluation in terms of outage, throughput, and/or
formance in severe shadowing. delay.
5. ALGORITHMS FOR RESOURCE ALLOCATION (3) Determine the effective power reduction Δpi j =
∂pi j + ∂p ji for each user pair (i, j), where
In this section, we discuss algorithms for the two strategies
that we implement for the performance evaluation of real- (a) ∂pi j is the minimum power reduction
time data services. among all subcarriers of i when a subcar-
rier of i is reassigned to j, that is, ∂pi j =
5.1. Power minimization mink {∂pikj }, and ki j = arg mink {∂pikj } and
There have been several proposed algorithms in the litera- ∂pikj = f j /γ j,k − fi /γi,k is the potential power
ture that try to solve the power minimization problem. In reduction when subcarrier k belonging to
[2], Lagrangian relaxation technique is used to derive an it- user i is reassigned to user j.
erative technique of allocation. This approach has high com- (4) Determine Δpmin = min{Δpi j } amongst all user
plexity in the number of iterations required to solve the prob- pairs (i, j). If Δpmin < 0, perform the cor-
lem. The brute-force integer programing formulation and its responding subcarrier re-assignment; otherwise
linear programing counterpart have been discussed in [16]. stop, the power cannot be reduced further.
This approach has even higher complexity, although it per-
forms better than the previous approach. In [17], an intuitive (iii) Bit loading.
approach to power minimization was proposed that divides (1) Initialize for each user n the bit loading level on
the allocation problem into three steps: (i) bandwidth alloca- each subcarrier k assigned to user n in the previ-
tion, (ii) subcarrier assignment, and (iii) bit loading. This is ous step to be cn,k = 0. Let the initial power incre-
the approach we use in implementing the power minimiza- ment be ∂pn,k = f p (1)/γn,k on each subcarrier.
tion algorithm. The subcarrier assignment proposed in [17]
is replaced by the one in [18] since this algorithm gave bet- (2) For each user n, while Kk=1 cn,k < bn ,
ter performance in terms of minimum power. The bit load- (a) l = arg mink ∂pn,k ,
ing algorithm is the greedy approach proposed in [19], which (b) cn,k = cn,k + 1,
is optimal for single user allocation. The details of the algo- (c) ∂pn,k = f p (cn,k + 1)/(γn,k ) − f p (cn,k )/(γn,k ).
rithms can be found in respective literature. Here, we sum-
marize them as follows.
5.2. Bandwidth constrained power minimization
(i) Bandwidth allocation.
For this strategy of resource allocation, a simple algorithm
(1) Initialize the subcarrier allocation for user n to
that assigned the minimum number of best available subcar-
Sn = Smin
n = bn /cη , where Sn is the number of
riers to the users was proposed in [8]. In [10], a more optimal
subcarriers assigned to the user n.
approach was proposed that follows the three-step approach
(2) While Nn=0 Sn < N,
described above. However the first two steps are replaced by
(a) let ∂rn = ((Sn + 1)/ γ̄n ) f p (bn /(Sn + 1)) − the following.
(Sn / γ̄n ) f p (bn /Sn ) for all n, where γ̄n is the av-
(i) Bandwidth allocation.
erage channel response of user n and ∂rn is
the power reduction if one more subcarrier (1) For each user n, allocate the minimum number
was allocated to the user; of subcarriers that would satisfy the users rate re-
(b) assign the additional subcarrier to the user quirement, determined as Sn = Smin n = bn /cη .
who causes the minimum power reduction,
(ii) Subcarrier assignment.
that is,
(1) For each user n, sort the subcarriers in ascending
– m = arg minn ∂rn ,
order in terms of fn /γn,k .
– Sm = Sm + 1.
(2) Allocate in a round-robin manner the best un-
used subcarrier to user n from the above list
(ii) Subcarrier assignment.
of sorted subcarriers until its subcarrier require-
(1) Initialize each user n by sorting his subcarriers ment is satisfied.
in ascending order in terms of fn /γn,k , where (3) Determine the effective power reduction Δpi j =
fn = fP (cn ) and cn = bn /Sn . Here cn is seen as min{∂pi j + ∂p ji , ∂pi j + ∂p ji + ∂pii , ∂pi j + ∂pi j +
the average number of bits loaded per subcarrier. ∂p j j } for each user pair (i, j), where
The actual number of bits per subcarrier is only
(a) ∂pi j is the minimum power reduction
decided in the bit loading stage, and cn is used
amongst all subcarriers of i when a subcar-
here conveniently to simplify the subcarrier as-
rier of i is reassigned to j, that is, ∂pi j =
signment process.
(2) Allocate in a round-robin manner the best un- mink {∂pikj }, and ki j = arg mink {∂pikj } and
used subcarrier to user n from the above list ∂pikj = f j /γ j,k − fi /γi,k is the potential power
of sorted subcarriers until its subcarrier require- reduction when subcarrier k belonging to
ment is satisfied. user i is reassigned to user j;
Expected packet error rate
Expected packet error rate

10−1 10−1
10−2 10−2
6 12 18 24 30 36 42 48 54 3 6 9 12 15
Number of users per cell Number of users per cell
PM 5 dB PM 10 dB PM 5 dB PM 10 dB
BCPM 5 dB BCPM 10 dB BCPM 5 dB BCPM 10 dB
(a) 32 kbps. (b) 64 kbps.
Figure 4: Performance of PM and BCPM schemes in supporting real-time data services.
(b) ∂pii is the minimum power reduction per sector); each having a bandwidth of 10 kHz and can
amongst the unused subcarriers when an support a symbol rate of 10 ksps (same assumption as in
unused subcarrier k is used instead of k ji , Section 3). The modulation and coding levels are the same
where k ji was reassigned from user j to user as used in Section 3, which give a spectral efficiency of 1–
i, that is, ∂pii = mink { fi /γi,k − fi /γi,k ji for 4 bps/Hz. The resource allocation interval is 20 ms in which
all k of ununsed subcarriers }. Determine the channel is assumed to be unchanged. The modified Hata
Δpmin = min{Δpi j } among all user pairs model is used to represent the path-loss model with a path-
(i, j). If Δpmin < 0, perform the correspond- loss exponent of 3.5.
ing subcarrier reassignment; otherwise stop, Shadowing is assumed to be correlated log-normal using
the power cannot be reduced further. the method stated in [20] with standard deviation of 5 and
10 dB and correlation distance of 20 m, which is commonly
(iii) Bit loading. used for a vehicular environment. The simulated multipath
(1) Initialize for each user n the bit loading level on power delay profiles are vehicular-B [11], and Rayleigh fad-
each subcarrier k assigned to user n in the previ- ing is assumed with the Jake’s method [20]. At the beginning
ous step to be cn,k = 0. Let the initial power incre- of the simulation, the mobiles are dropped in the sectors with
ment be ∂pn,k = f p (1)/γn,k on each subcarrier. speed of 30 km/h. No handoffs are simulated, and it is as-
sumed that if the mobile leaves the sector from one edge, it
(2) For each user n, while Kk=1 cn,k < bn ,
enters from another to preserve the number of mobiles in the
(a) l = arg mink ∂pn,k , area during simulation.
(b) cn,k = cn,k + 1, We simulate two data rate scenarios. The low-rate sce-
(c) ∂pn,k = f p (cn,k + 1)/γn,k − f p (cn,k )/γn,k . nario consists of a constant rate of 32 kbps [11] with packet
Note that this is the same bit-loading algorithm used for sizes of 320 bits at a constant inter-arrival time of 10 ms. The
power minimization. high-rate scenario consists of a constant rate of 64 kbps [11]
with packet sizes of 640 bits at a constant inter-arrival time of
10 ms. The packet error rate is used as a performance mea-
6. PERFORMANCE OF REAL-TIME DATA SERVICES IN sure that captures the link outage.
FREQUENCY SELECTIVE FADING AND DIFFERENT Figure 4(a) shows the performance results for 32 kbps in
SHADOWING ENVIRONMENTS both 5 and 10 dB shadowing scenarios. It can be seen that
the same trend is followed in this figure as the analysis. In a
For performance evaluation of data services, we simulate a mild shadowing environment (σdB = 5 dB), the PM scheme
cellular environment with a 19-cell configuration that in- performs better with 1.5 times more users at PER = 10%. On
cludes the effects of up to second-tier interferers. Each hexag- the other hand, in a severe shadowing environment (σdB =
onal cell is divided into three sectors, with 3 beams per sector. 10 dB), the BCPM gives twice as many users as PM at PER =
The OFDM system has 90 traffic subcarriers (30 subcarriers 10%.
Figure 4(b) also gives similar conclusions as the analysis The interference power from user i’s base-station to user j
for 64 kbps data services: PM outperforms BCPM for mild can be given as pci j = ptii · psi j , and the corresponding pdf
shadowing (σdB = 5 dB) and the situation is reversed for se- is
vere shadowing (σdB = 10 dB). However, the difference in 2
performance between PM and BCPM is smaller. This may be 1 ln pci j − mc
ij
f pci j pci j = √ exp − , (A.4)
attributed to the fact that at high rates, a lot more subcarri- 2π σc pci j 2σc2
ers are employed for the lower bandwidth-efficient modula-

tion scheme (e.g., at 15 users per cell, rate-1/2 QPSK would where
need 105 subcarriers). However, since there are only a lim- i j = −m
mc ii + m
ij
ited number of subcarriers, the PM algorithm would have to 1 1
use higher bandwidth-efficient modulation schemes in some = β ln rii + ln(2) − β ln ri j − ln(2)
2 2
subcarriers to satisfy the rate requirements. Hence, the differ-
rii
ence in performance between PM and BCPM at higher rates = β ln
and higher loads is less pronounced. ri j
⎛ ⎞
rii
= β ln ⎝ ⎠
7. CONCLUSIONS r 2j j + Di2j − 2r j j Di j cos φi j + π − θ j j
In this paper, we have shown that in downlink OFDM mo- (A.5)
bile cellular systems, the probability of interference occur-
rence is an important factor in determining the system per- and σc2 = 2σ 2 + 2 ln(2).
formance, which should be accounted for in the resource al-
location strategy. The proposed BCPM schemes to minimize B. PDF OF RECEIVED INTERFERENCE POWER FOR
first the number of subcarriers and then power minimization MILD SHADOWING
in satisfying user rate requirements can significantly enhance
link performance. We derived a framework to analyze the Padé approximation technique [13] can be used to approx-
expected outage probability of different transmission band- imate the pdf of pci j for mild shadowing case. The power
widths and corresponding modulation schemes in flat fading series of a pdf around two points can be expressed as
and shadowing cellular environment and showed the ben- ∞

efits of constraining number of allocated subcarriers. It was h(u) = cn un , u −→ 0,
shown that in frequency-selective fading, the BCPM schemes n=0
∞

significantly outperform PM strategy alone for both voice (B.1)
and data services. h(u) = dn u−(n+1) , u −→ ∞,
n=0
μn
APPENDICES cn = (−1)n , μn = nth moment of pdf,
n!
A. PDF OF RECEIVED INTERFERENCE POWER FOR where
SEVERE SHADOWING
dn = f (n) (0), f (n) = nth derivative of pdf (B.2)
The density function of psi j can be approximated by a log-
normal distribution [12] with a reduced logarithmic area Details of the above equations are given in [13]. In the case of
mean power m i j and an increased logarithmic standard de- received power psi j given 0 dB transmit power, μn and f n (0)
viation σ given as follows: can be derived as

σ 2 = σ 2 + ln(2), 1
μn = n! exp nmi j + (nσ)2 ,
1 (A.1) 2
1 (B.3)
i j = mi j − ln(2) = −β ln ri j − ln(2).
m 1 !2
2 2 f n (0) = (−1)n exp −(n + 1)mi j + (n + 1)σ .
2
Hence the approximated pdf of psi j can be given as
2 The Padé approximation is a rational function approxima-
1 ln psi j − m
ij tion of the power series and can be used to approximate only
f psi j psi j = √ exp − . (A.2)
2π σ psi j 2σ 2 the first few terms of h(u). It has the following form:
M −1 n
Given the received power pdf based on a transmit power of [M −1/M] n=0 an u
1, for a power controlled system, the instantaneous trans- P(J,K) (u) = M
1 + n=1 bn un
mit power of user i from base-station i can be given by ptii ⎧ J −1
= 1/ psii assuming that the received power of user i is normal- ⎪
⎪
⎪ c un ,
⎪ u −→ 0,
⎪
⎪ n (B.4)
ized to 1. The corresponding pdf of ptii is ⎨n=0
2 =
⎪
⎪ −1
K
1 ln ptii + m ii ⎪
⎪
f ptii ptii = √ exp − . (A.3) ⎪
⎪ dn u−(n+1) , u −→ ∞.
⎩
2π σ ptii 2σ 2
n=0
Once the coefficients an and bn are found, then a partial frac- [6] A. Seyedi and G. J. Saulnier, “A sub-channel selective orthogo-
tion decomposition can be done: nal frequency division multiplexing spread spectrum system,”
in Proceedings of IEEE Military Communications Conference
(MILCOM ’01), vol. 2, pp. 1370–1374, McLean, Va, USA, Oc-
M −1 n
M tober 2001.
[L/M] n=0 an u λi
P(J,K) = M = . (B.5) [7] L. Loyola and T. Miki, “An OFDMA/CDMA-based bandwidth
1 + n=1 bn un m=1
u + qm
resource allocation scheme for future broadband mobile
networks,” in Proceedings of 56th IEEE Vehicular Technol-
Inverting this expression, we get a mixture of exponential dis- ogy Conference (VTC ’02), vol. 1, pp. 42–46, Vancouver, BC,
Canada, September 2002.
tributions:
[8] N. Damji and T. Le-Ngoc, “Adaptive downlink multi-carrier
resource allocation for real-time multimedia traffic in cellu-

M
lar systems,” in Proceedings of IEEE International Conference
f (x) = λi exp −qm x . (B.6) on Communications (ICC ’04), vol. 7, pp. 4258–4262, Paris,
m=1 France, June 2004.
[9] G. Kulkarni, S. Adlakha, and M. Srivastava, “Subcarrier al-
Hence, the pdf of psi j for 0 dB transmit power can be approx- location and bit loading algorithms for OFDMA-based wire-
less networks,” IEEE Transactions on Mobile Computing, vol. 4,
imated by the above expression. Following the same argu-
no. 6, pp. 652–662, 2005.
ment as in Appendix A, the transmit power of a user within
the same cell can be given by ptii = 1/ psii . The interference [10] N. Damji and T. Le-Ngoc, “Adaptive BCPM downlink resource
allocation strategies for multiuser OFDM in cellular systems,”
power from user i’s base-station to user j given as pci j = ptii .
in Proceedings of the 2nd International Conference on Broad-
psi j = psi j / psii can now be calculated by transformation of band Networks (BROADNETS ’05), Boston, Mass, USA, Octo-
random variables since both psi j and psii follow the same dis- ber 2005.
tribution as (B.6), [11] TR 101 112 V3.2.0 (1998-04), Universal Mobile Telecommuni-
cations System (UMTS); Selection procedures for the choice of
radio transmission technologies of the UMTS, ETSI, 1998.

M
M
1
f pci j (pci j ) = λms λlt 2 , (B.7) [12] R. Prasad and J. C. Arnbak, “Effects of Rayleigh fading on
l=0 m=0 qms pci j + qlt packet radio channels with shadowing,” in Proceedings of 4th
IEEE Region 10 International Conference (TENCON ’89), pp.
546–548, Bombay, India, November 1989.
where qms , λms are the coefficients of received power of inter- [13] H. Amindavar and J. A. Ritcey, “Padé approximations of prob-
ference, and qlt , λlt are the coefficients for the derived trans- ability density functions,” IEEE Transactions on Aerospace and
mit power of the interference. Electronic Systems, vol. 30, no. 2, pp. 416–424, 1994.
[14] S. C. Schwartz and Y. S. Yeh, “On the distribution function
ACKNOWLEDGMENT and moments of power sums with log-normal components,”
The Bell System Technical Journal, vol. 61, no. 7, part 1, pp.
This work is partially supported by an NSERC CRD Grant 1441–1462, 1982.
with Nortel Networks. [15] S. Vishwanath and A. Goldsmith, “Adaptive turbo-coded
modulation for flat-fading channels,” IEEE Transactions on
REFERENCES Communications, vol. 51, no. 6, pp. 964–972, 2003.
[16] I. Kim, H. L. Lee, B. Kim, and Y. H. Lee, “On the use of lin-
[1] W. Mohr and W. Konhäuser, “Access network evolution be- ear programming for dynamic subchannel and bit allocation
yond third generation mobile communications,” IEEE Com- in multiuser OFDM,” in Proceedings of IEEE Global Telecom-
munications Magazine, vol. 38, no. 12, pp. 122–133, 2000. munications Conference, vol. 6, pp. 3648–3652, San Antonio,
[2] C. Y. Wong, R. S. Cheng, K. B. Letaief, and R. D. Murch, “Mul- Tex, USA, November 2001.
tiuser OFDM with adaptive subcarrier, bit, and power allo- [17] D. Kivanc, G. Li, and H. Liu, “Computationally efficient band-
cation,” IEEE Journal on Selected Areas in Communications, width allocation and power control for OFDMA,” IEEE Trans-
vol. 17, no. 10, pp. 1747–1758, 1999. actions on Wireless Communications, vol. 2, no. 6, pp. 1150–
1158, 2003.
[3] W. Rhee and J. M. Cioffi, “Increase in capacity of multiuser
OFDM system using dynamic subchannel allocation,” in Pro- [18] C. Y. Wong, C. Y. Tsui, R. S. Cheng, and K. B. Letaief, “A real-
ceedings of 51st IEEE Vehicular Technology Conference (VTC time sub-carrier allocation scheme for multiple access down-
’00), vol. 2, pp. 1085–1089, Tokyo, Japan, May 2000. link OFDM transmission,” in Proceedings of IEEE Vehicular
Technology Conference (VTC ’99), Amsterdam, The Nether-
[4] G. Song and Y. Li, “Cross-layer optimization for OFDM wire-
lands, September 1999.
less networks - part I: theoretical framework,” IEEE Transac-
tions on Wireless Communications, vol. 4, no. 2, pp. 614–624, [19] D. Hugh-Hartogs, Ensemble modem structure for imperfect
2005. transmission media, U.S. Patents Nos. 4679227, July 1987;
4731816, March 1988; and 4833796, May 1989.
[5] G. Song and Y. Li, “Cross-layer optimization for OFDM wire-
less networks - part II: algorithm development,” IEEE Trans- [20] G. L. Stüber, Principles of Mobile Communication, Kluwer Aca-
actions on Wireless Communications, vol. 4, no. 2, pp. 625–634, demic, Boston, Mass, USA, 1996.
2005.
Navid Damji received his Bachelor’s and

Master’s degrees in electrical engineering
from McGill University, Montreal, Canada,
in 2002 and 2004, respectively. Since
September 2004, he has been with SRTele-
com as a Software DSP Engineer involved
in the development of the physical layer for
the IEEE 802.16 standards. He is pursing
his Ph.D. in electrical engineering at McGill
University. His research interests are in the
area of broadband wireless communications with emphasis on re-
source allocation and interference mitigation in OFDM cellular
systems.
Tho Le-Ngoc obtained his B.Eng. degree

(with distinction) in electrical engineering
in 1976, his M.Eng. degree in micropro-
cessor applications in 1978 from McGill
University, Montreal, and his Ph.D. degree
in digital communications in 1983 from
the University of Ottawa, Canada. During
1977–1982, he was with Spar Aerospace Ltd.
as a Design Engineer and then a Senior De-
sign Engineer, involved in the development
and design of the microprocessor-based controller of Canadarm
(of the Space Shuttle), and SCPC/FM, SCPC/PSK, TDMA satellite
communications systems. During 1982–1985, he was an Engineer-
ing Manager of the Radio Group in the Department of Develop-
ment Engineering of SRTelecom Inc., developed the new point-to-
multipoint DA-TDMA/TDM subscriber radio system SR500. He
was the System Architect of this first digital point-to-multipoint
wireless TDMA system. During 1985–2000, he was a Professor
at the Department of Electrical and Computer Engineering of
Concordia University. Since 2000, he has been with the Depart-
ment of Electrical and Computer Engineering of McGill Univer-
sity. His research interest is in the area of broadband digital com-
munications with a special emphasis on modulation, coding, and
multiple-access techniques. He is a Senior Member of the Ordre
des Ingénieur du Quebec, a Fellow of the Institute of Electrical
and Electronics Engineers (IEEE), a Fellow of the Engineering In-
stitute of Canada (EIC), and a Fellow of the Canadian Academy
of Engineering (CAE). He is the recipient of the 2004 Canadian
Award in Telecommunications Research, and the recipient of the
IEEE Canada Fessenden Award 2005.
DOI 10.1155/WCN/2006/80493
Opportunistic Nonorthogonal Packet Scheduling in Fixed

Broadband Wireless Access Networks
Mahmudur Rahman,1 Halim Yanikomeroglu,1 Mohamed H. Ahmed,2 and Samy Mahmoud1

1 Broadband Communications and Wireless Systems (BCWS) Centre, Department of Systems and Computer Engineering,
Carleton University, Ottawa, Ontario, Canada K1S 5B6
2 Faculty of Engineering and Applied Science, Memorial University of Newfoundland, St. John’s, NL, Canada A1B 3X5
Received 14 October 2005; Revised 11 March 2006; Accepted 13 March 2006

In order to mitigate high cochannel interference resulting from dense channel reuse, the interference management issues are
often considered as essential part of scheduling schemes in fixed broadband wireless access (FBWA) networks. To that end, a series
of literature has been published recently, in which a group of base stations forms an interferer group (downlink transmissions
from each base station become dominant interference for the users in other in-group base stations), and the scheduling scheme
deployed in the group allows only one base station to transmit at a time. As a result of time orthogonality in transmissions, the
dominant cochannel interferers are prevented, and hence the packet error rate can be improved. However, prohibiting concurrent
transmissions in these orthogonal schemes introduces throughput penalty as well as higher end-to-end packet delay which might
not be desirable for real-time services. In this paper, we utilize opportunistic nonorthogonality among the in-group transmissions
whenever possible and propose a novel transmission scheduling scheme for FBWA networks. The proposed scheme, in contrast
to the proactive interference avoidance techniques, strives for the improvements in delay and throughput efficiency. To facilitate
opportunistic nonorthogonal transmissions in the interferer group, estimation of signal-to-interference-plus-noise ratio (SINR) is
required at the scheduler. We have observed from simulations that the proposed scheme outperforms the reference orthogonal
scheme in terms of spectral efficiency, mean packet delay, and packet dropping rate.
Copyright © 2006 Mahmudur Rahman et al. This is an open access article distributed under the Creative Commons Attribution
cited.
1. INTRODUCTION FBWA networks [11–16]. In our previous works [12, 13], we

have shown that a very effective means of managing inter-
Fixed broadband wireless access (FBWA) [1, 2] is recognized ference is to employ coordinated orthogonal transmissions
to be a promising alternative technology to existing copper among dominant interferers achieved by inter-base station
line asymmetric digital subscriber loop (ADSL) [3, 4] and hy- (BS) signaling. The main idea of this scheme is to group a
brid fiber-coaxial (HFC) [5] cable broadband services for its number of BSs (termed as interferer group) that are domi-
fast, simple, and less expensive deployment. However, effi- nant interferers to each other and to schedule transmission
cient system planning and resource allocation policies are orthogonally so that only one BS in the group transmits at
warranted for such systems, because in addition to the chal- a particular time. This scheme is composed of two indepen-
lenges posed by the dynamic nature of wireless links, interfer- dent scheduling disciplines and hence named as intrasector
ence resulting from aggressive channel reuse is a major design and intersector scheduling (ISISS) [13].
concern. Therefore, resource allocation strategies play a ma- High end-to-end packet delay is the main drawback of
jor role for the successful evolution of FBWA. In this paper, the ISISS scheme. Packet delay is an important quality-of-
we focus on one of the most important aspects of resource service (QoS) parameter for a variety of delay-sensitive ap-
allocation, packet scheduling. plications, which is directly related to the throughput for a
Wireless scheduling techniques [6–10] have emerged as given data rate. Therefore, improving throughput and de-
tailored versions of wireline scheduling to cope with the dy- lay in an orthogonal scheduling scheme is essential. In this
namic nature of wireless links. To account for cochannel in- paper, we propose a novel scheduling scheme that improves
terference, it is common to consider the issues of interference both packet delay and resource utilization in terms of area
management as an integral part of scheduling techniques in spectral efficiency. The performance of the proposed scheme
I ¼¼ J ¼¼ K ¼¼ I¼ J¼ K¼
I¼ J¼ K¼ J K I
I J K K ¼¼ I ¼¼ J ¼¼
(a) (b)
Figure 1: (a) Nine-cell network, (b) wraparound interferer positions for SSs in BS I.
is compared to that of a reference scheme adapted from ba- controlled in-group interference, which functions adaptively
sic ISISS [13]. This reference scheme is named as intrasector in an optimistic manner yielding the capability of improv-
and orthogonal intersector scheduling with fixed modulation ing throughput and the delay. The details of the proposed
(ISOISS-FM). scheme are illustrated in Section 3.
Proposed scheme in this paper considers interference Similar notion of concurrent cochannel transmissions
management issues, integrates adaptive modulation and cod- based on terminal classifications has been previously con-
ing (AMC), and makes channel-state-based scheduling deci- sidered in the enhanced staggered resource allocation (ESRA)
sions to enhance network performance. We investigate the scheme [14]. However, the time slot allocation in that scheme
performance of the proposed scheme in two steps. First, we is static, which might result in low resource utilization es-
introduce AMC instead of fixed modulation and evaluate the pecially for bursty traffic such as in FBWA. The proposed
performance of the scheme. The resulting scheme is still or- scheme in this paper, on the contrary, is dynamic in nature,
thogonal, while it makes channel-state-based scheduling de- adaptive according to the channel state, and optimistic.
cisions. This intermediate scheme is named as intrasector and The intermediate and proposed schemes are more prone
orthogonal intersector scheduling with adaptive modulation to packet errors compared to the reference ISOISS-FM, pri-
and coding (ISOISS-AMC). Investigation of this intermediate marily because the predicted SINRs in these schemes do not
scheme quantifies the performance gain achieved from the account for the out-of-group interference. We define param-
use of AMC in an orthogonal scheme. We then employ op- eter interference compensation guard to offset overestimation
portunistic nonorthogonality in transmissions, where mul- in the predicted SINR. This guard acts as a method of pro-
tiple cochannel BSs are allowed to transmit simultaneously. tecting the in-group transmissions to a certain degree from
This final scheme is named as intrasector and opportunistic out-of-group interference. The effect of interference compen-
nonorthogonal intersector scheduling with adaptive modula- sation guard on the performance of proposed ISONOISS-
tion and coding (ISONOISS-AMC). Basically, if a number of AMC scheme has also been investigated.
cochannel BSs transmit simultaneously, each becomes inter- The rest of this paper is organized as follows. Section 2
ferer for the users in other BSs. The idea is that if the interfer- describes the reference ISOISS-FM scheme. The intermedi-
ence levels (hence the SINRs) are predicted and are transpar- ate ISOISS-AMC and proposed ISONOISS-AMC schemes
ent to each BS in the group, then every BS in the interferer are illustrated in Section 3. Section 4 describes system model.
group would potentially be able to transmit simultaneously Simulation results are presented in Section 5 followed by
with its feasible AMC mode in the presence of others being conclusions in Section 6.
interferers.
Opportunistic scheduling, in general, implies a scheduling 2. REFERENCE SCHEME: ISOISS-FM
mechanism that exploits channel variations and schedules a
A downlink time-division multiple-access (TDMA) system
user having the best channel condition at the time of inter-
in a hexagonal six-sectored nine-cell network as shown in
est [17]. However, according to the context of our study in
Figure 1(a) is considered in the reference as well as in the in-
this paper, opportunistic nonorthogonal scheduling means ex-
termediate and proposed schemes. It is assumed that a fre-
ploitation of channel variations among a group of mutually
quency reuse plan with a reuse factor of 1/6 is employed in
interfering BSs and scheduling concurrent in-group trans-
the network. The shaded sectors1 (e.g., sector 1 in Figure 1)
missions opportunistically based on the mutual interference
in all cells use the same frequency band. It should be noted
situation.
here that an alternative assignment technique for sectors,
The proposed scheme in contrast to the widely stud-
ied proactive interference avoidance techniques predicts the
interference and achievable SINR on the fly. It then de- 1 Only the shaded cochannel sectors (one sector per cell site) are simulated
cides whether or not concurrent transmissions in the in- in this study. Therefore, BS I, for instance, implies shaded sector of BS
terferer group should be allowed at a particular instant. I throughout this paper. Note that the reuse factor is 1/6, and therefore
This reactive interference-aware scheduling scheme allows there is no intersector interference among the sectors of a particular cell.
Mahmudur Rahman et al. 3
Tx change information with each other as illustrated in the fig-

Intersector
scheduler ure. The intrasector scheduling discipline decides the service
order of each SS inside the sector, while the intersector disci-
FCFS pline determines the service order among different BSs in the
Intrasector
scheduler group to ensure orthogonal or opportunistic nonorthogonal
transmissions in the interferer group. As the contributions of
I J K
the schemes lie in the intersector scheduler, for simplicity the
first-come-first-serve (FCFS) principle is considered as the in-
trasector discipline in the reference system as well as in the
intermediate and proposed schemes.
Transmissions use fixed 16-quadrature amplitude modu-
lation (16-QAM) bit-interleaved coded modulation (BICM)
Information
exchange
with a coding rate of 1/2 in the reference ISOISS-FM scheme.
Base stations in the interferer group exchange traffic-related
information, such as the arrival times of the packets (with
Figure 2: Block diagram of the scheduling scheme in group the packet lengths) arrived in previous data frame duration.
{I, J, K }.
Therefore, each BS in the group is aware of the arrival times
of the packets of its own queue as well as the packets of the
queues of the other BSs in the group. The intersector sched-
uler checks the arrival times of the head-of-line (HOL) pack-
such as the rotational or staggering approach used in [11] ets in all three queues in the group and selects the candidate
or [14], is also possible in order to reduce intersector inter- packet to be transmitted that has the earliest arrival time; for
ference, especially for lower network loading. The rationale example, in group {I, J, K }at a particular instant,
behind the assignment used in this study, where cochannel
sectors are positioned in a line, is to investigate the worst- j

case intersector interference scenario. However, the proposed w = arg min tai , ta , tak , (1)
I,J,K
scheduling scheme can be employed with any other fre-
quency planning to enhance the performance in addition where w is the BS that wins the service opportunity at that in-
to what can be obtained by the static frequency assignment j
stant, and tai , ta , and tak are the arrival times of the HOL pack-
alone. We assume that base stations and subscriber station ets at BSs I, J, and K destined to SSs i, j, and k, respectively.
(SS) terminals are equipped with directional antennas with
60◦ and 30◦ beamwidths, respectively. The SS antennas are
pointing towards the serving BSs. The effective gains of BS 3. DESCRIPTIONS OF THE INTERMEDIATE
transmit and SS receive antennas are considered to be 20 dB AND PROPOSED SCHEMES
(10 dB main and −10 dB side lobe) and 10 dB (5 dB main and Schematically, the reference, intermediate, and proposed
−5 dB side lobe), respectively.
schemes are alike in the sense that they all are composed
We have considered wraparound interference model such of two independent schedulers (intrasector and intersector).
that an interferer BS position is taken to be at a place from The main difference is in the function of the intersector
where it contributes the maximum interference for the SSs in schedulers and modulation (fixed or adaptive). The inter-
the BS of interest (see [18] for details). Figure 1(b) shows the mediate and proposed schemes make channel-state-based
positions of the interferer BSs for the SSs in BS I. Base sta- scheduling decisions and employ AMC based on the pre-
tion sets {J, K } and {I , J , K , I , J , K } are potential in- dicted SINR for transmissions towards particular SSs. In this
group and out-of-group interferers for the SSs in BS I, re- section, we provide an overview of the SINR estimation first,
spectively. A similar approach can be followed to find out and then we describe how the intersector schedulers work in
the positions of interferers for SSs in other BSs. It can easily ISOISS-AMC and ISONONISS-AMC schemes.
be conceived that as a result of combined effects of the an-
tenna directivities, gains, and relative positions of the cells,
the downlink transmissions from BSs I and Jwill be the two 3.1. SINR estimation and BS information exchange
most dominant interferers for the SSs in BS K. Similarly, BS In order for the intermediate and proposed schemes to be
I and wraparound BS K (considered to be at the left of BS I) able to execute link-state-based scheduling decisions and em-
would be the most dominant interferers for the SSs in BS J. ploy AMC, SINR would have to be estimated at each BS. For
Moreover, wraparound BSs J and K are the most dominant the nine-cell network shown in Figure 1(a), every transmis-
interferers for SSs in BS I. Following these arguments, BSs I, sion will have eight potential interferers. Let us consider the
J, and K form an interferer group. Similarly, BSs {I , J , K } scenario shown in Figure 1(b). The SINR of a received packet
and {I , J , K } form two other interferer groups in the at SS i served by BS I can be expressed as
network.
The scheduling scheme (reference, intermediate, or pro-
Pt GiI
posed) is employed in each interferer group as shown in γi =
i

i i , (2)
Figure 2 for interferer group {I, J, K }. The in-group BSs ex- Pt =I Ax Gx + Pt
x∈IG, x y ∈OG A y G y + PN
where Pt is the fixed transmit power. The first term in the de- Table 1: Lookup table for AMC modes. Data for BICM modulation
nominator is the summation of interference from in-group curves are provided by Dr. Sirikiat Lek Ariyavisitakul.
BSs (IG) and the second term expresses the total interfer- Efficiency,
ence from out-of-group BSs (OG). For the given scenario, SINR range (dB) AMC mode
(bits/s/Hz)
IG ≈ {I, J, K }and OG ≈ {I , J , K , I , J , K }. Parameter
3.39 ≤ γ < 5.12 QPSK rate 1/2 1.0
GiI is the link gain between the serving BS I and SS i. Param-
eters Gix and Giy are the link gains to the desired SS i from 5.12 ≤ γ < 6.02 QPSK rate 2/3 1.33
the interfering in-group and out-of-group BSs, respectively. 6.02 ≤ γ < 7.78 QPSK rate 3/4 1.5
These link gain parameters include the effect of antenna gains 7.78 ≤ γ < 9.23 QPSK rate 7/8 1.75
at the BS and the SS terminals, as well as the propagation loss 9.23 ≤ γ < 11.36 16-QAM rate 1/2 2.0
(including shadowing and fading) of the link. In (2), PNi is 11.36 ≤ γ < 12.50 16-QAM rate 2/3 2.67
the average thermal noise computed at the receiver of SS i. 12.5 ≤ γ < 14.21 16-QAM rate 3/4 3.0
We note that all BSs do not necessarily transmit simul- 14.21 ≤ γ < 16.78 16-QAM rate 7/8 3.5
taneously because of either algorithm dictation or empty 16.78 ≤ γ < 18.16 64-QAM rate 2/3 4.0
queues. The parameters Ax and A y in (2) denote activity fac- 18.16 ≤ γ < 20.13 64-QAM rate 3/4 4.5
tors which take value of 1 if the interferer BS is transmitting
20.13 ≤ γ < 24.30 64-QAM rate 7/8 5.25
and 0 if it is idle. An expression similar to (2) is applicable for
γ ≥ 24.30 64-QAM rate 1 6.0
the SINR at any SS in other BSs.
The link gain parameters are monitored at the SS termi-
nal and reported back to the serving BS from where they are
exchanged among in-group BSs by inter-BS signaling. For ex- any time, three HOL packets in the in-group BSs are com-
ample, SS i in the interferer group of {I, J, K } keeps track of pared by the intersector scheduler to select the candidate BS
GiI , GiJ , and GiK , and reports this information to the serving that has the best link to the SS. If SSs i, j, and k are the can-
BS I as often as necessary. BS I shares this information with didates for HOL packets in BSs I, J, and K in the interferer
j
in-group BSs J and K. It is important to note that the channel group, and GiI , GJ , and GkK are the link gains from BSs to SSs,
changes slowly because of the fixed SS locations; this yields respectively, then
low Doppler shifts in FBWA networks. Therefore, link state
reporting does not have to be very frequent, which makes it j
w = arg max GiI , GJ , GkK , (5)
completely feasible in such systems. I,J,K
Since the inter-BS signaling is performed only among in-
group interferers, BSs do not have knowledge about the out- where w is the BS that wins the scheduling opportunity.
of-group interference, and hence the estimated SINRs do not The selected BS predicts the SINR according to (3) or a
include the second denominator term in (2). The estimated similar expression. Using this estimated SINR, the feasible
SINRs for orthogonal ISOISS-AMC scheme, γOi , and for op- AMC mode is chosen from Table 1 and the packet is sched-
portunistic nonorthogonal ISONOISS-AMC scheme, γONO i
, uled for the instant. It should be noted that the modula-
for SS i are given as follows: tion schemes listed in Table 1 are the mandatory schemes for
downlink transmissions recommended by the 802.16 a stan-
Pt GiI dard [1].
γOi = , (3)
PNi
3.3. Intersector scheduler in the proposed
i Pt GiI
γONO = i . (4) ISONOISS-AMC scheme
Pt i
=I Ax Gx + PN
x∈IG,x
Using estimated SINRs from (4), the intersector scheduler
From (3), we see that only the link gains from the serv- finds a combination of concurrent transmissions that gives
j
ing BSs to desired SSs, for example, {GiI , GJ , GkK } for BS the highest aggregate throughput efficiency. If queues of all
group {I, J, K }, are required in order to estimate SINRs in-group BSs are nonempty, there are seven possible combi-
in ISOISS-AMC, while additional link gain information nations of transmissions at a particular instant. For exam-
j j
{GI , GkI , GiJ , GkJ , GiK , GK } are to be exchanged in ISONOISS- ple, all three BSs transmit (1 choice) or two BSs transmit (3
AMC as in (4). The number of in-group interference con- choices), or only one BS transmits (3 choices). We note that
tributing terms in the denominator of (4) equals the number the last 3 choices are only available transmission options in
of in-group BSs transmitting simultaneously, minus one. ISOISS-AMC. For each combination, first, the SINRs are es-
timated from exchanged information as discussed. Then, the
3.2. Intersector scheduler in the intermediate spectral efficiency for each transmission is calculated. Finally,
ISOISS-AMC scheme the aggregate spectral efficiency for the combination of si-
multaneous transmissions is predicted.
Similar to ISOISS-FM scheme, this scheme is orthogonal as Let us illustrate the steps for the first combination when
well; however, it employs AMC instead of fixed modulation all three BSs I, J, and K have potential to transmit con-
and makes channel-state-based scheduling decisions as op- currently to respective SSs i, j, and k. Each reception will
posed to the arrival-time-based decisions in ISOISS-FM. At have two in-group interferers. Therefore, according to (4) the
estimated SINR at SS i’s packet, given I, J, and K are trans- Table 2: Out-of-group interference compensation values for
mitting simultaneously, is ISONOISS-AMC.
Network loading Compensation guard
i|(I,J,K) Pt GiI (SSs/sector) (dB)
γONO = . (6)
Pt GiJ + Pt GiK + PNi 4 0.17
j |(I,J,K)
k|(I,J,K)
8 0.84
Similarly, for BSs J and K, γONO and γONO can be 12 1.29
found in a straightforward manner. 14 2.20
From these estimated SINRs, the achievable AMC modes,
16 2.08
and corresponding spectral efficiencies ηI , ηJ , and ηK can be
18 2.33
obtained from Table 1. Then, the aggregate spectral efficiency
ΓI,J,K for the combination is calculated from the following 20 2.40
relation: 24 2.44
j
ti t tk
ΓI,J,K = ηI × d + ηJ × d + ηK × d , (7) interference guard while making SINR estimations. Let us
tr tr tr
denote that 50th percentile value of the error between the ac-
j tual and estimated SINR is φ(l) (dB), which is a function of
where tdi , td , and tdk are the transmission durations for BSs the network loading l users/sector. There could be numerous
I, J, and K’s packet determined by the packet length and ways to find this error in a real network. For example, the
AMC modes as discussed later. The longest transmission network can be equipped with a mechanism to track out-
time among all three transmission durations is denoted as of-group interference from history. However, in this study,
j
tr , that is, tr = max(tdi , td , tdk ). we find this error from simulations as follows. First, a set
Similarly, aggregate spectral efficiencies for other combi- of SINRs for different loading values is noted in the pres-
nations, namely ΓI,J , ΓJ,K , ΓK,I , ΓI , ΓJ , and ΓK , can be calcu- ence of out-of-group interferers. Then, a second set is gener-
lated. Service opportunity is granted to the combination of ated where the out-of-group interferers are neglected. Now,
BSs that gives highest aggregate spectral efficiency according the difference of the 50th percentile SINR (dB) of these two
to the following: sets gives φ(l). Table 2 shows different φ(l) values for differ-
ent network loading levels obtained in the ISONOISS-AMC
w = arg max(ΓI,J,K , ΓI,J , ΓJ,K , ΓK,I , ΓI , ΓJ , ΓK ), (8) scheme. We investigate the effect of this guard only for the
proposed scheme.
where w is the set of BSs transmiting concurrently. The amount of error φ(l) (dB) is subtracted from (4)
We note here that packets in different BSs take different (dB) to obtain the expected SINR in ISONOISS-AMC. The
lengths of frame time due to the variability of packet size, estimated SINR with guard at SS i’s packet, given I, J, and K
modulation level, and coding rate. In order to avoid excessive are transmitting simultaneously, is
interference, a new scheduling event cannot be made until
the largest transmission time tr of the previous event elapses. i|(i,J,K) Pt GiI
γONO,guard = 10 log10 − φ(l). (9)
Pt GJ + Pt GiK + PNi
i
3.4. Out-of-group interference guard

However, while employing this guard is expected to
An effort has been made in order to avoid out-of-group in- improve the packet error rate performance of the pro-
terference as much as possible in all simulated scheduling posed scheme, it will lower the throughput, as the scheduler
schemes by using groupwise time partitioning in the frame. chooses the AMC modes more conservatively. Therefore, this
The frame is partitioned into three subframes (SFs), indexed interference guard can be regarded as a system design param-
as SF1, SF2, and SF3 from start to the end of the frame. eter to be adjusted according to desired tradeoff between the
BSs in the interferer group {I, J, K } schedule their traffic packet error rate and throughput efficiency.
with the subframe sequence of {SF1, SF2, SF3}, while, group
{I , J , K } and {I , J , K } use the sub-frames in the se- 3.5. A note on implementations
quences of {SF2, SF3, SF1} and {SF3, SF1, SF2}, respec-
tively. Clearly, this technique is effective as long as the ar- It should be mentioned that in a practical deployment sce-
riving traffic in each group is such that it can be accommo- nario, a single BS would qualify as a member of three in-
dated into 1/3 of the frame. However, the system must be de- dependent interferer groups for the above-described setting.
signed for loaded network where out-of-group interference Therefore, there is an issue of resolving the conflicts that
is inevitable. might arise from the commands of three different groups.
SINR estimations discussed in Section 3.1 do not take the Our focus in this paper is to present the basic concept of
out-of-group interference into account. As a result, the esti- opportunistic nonorthogonal scheduling; nevertheless, we
mations are optimistic, which might result in higher packet state a number of solutions to this issue. First, the interferer
error rate. To investigate the effects of out-of-group interfer- groups can be determined in such a way that each BS can only
ence on network performance, we consider an out-of-group be a member of only one interferer group. This deployment
Table 3: System parameters. Table 4: Traffic model parameters of the video stream [22].
Parameters Values Packet Pareto parameter Pareto parameter
Hexagonal six-sectored cell radius (km) 2.0 IRP arrival rate for ON for OFF
Propagation exponent, n 3.75 (packets/s) distribution distribution
Fixed transmit power (Watts) 6.5 IRP#1 112.38 1.14 1.22
BS antenna (600 beam width) gain (dB) 20 (front 10, back −10) IRP#2 154.75 1.54 1.28
SS antenna (300 beam width) gain (dB) 10 (front 5, back −5)
Transmission direction Downlink
Uplink-downlink duplexing FDD considered independent lognormal random variables with a
Multiple access TDMA standard deviation of 8 dB for shadowing.
Frequency reuse factor 1/6 Time-correlated flat Rayleigh fading with Doppler fre-
Carrier frequency, f (GHz) 2.5 quency of 2.0 Hz has been considered in this study, where
Channel bandwidth, B (MHz) 3.0 the Doppler spectrum S( f ) is given by the following equa-
Time-correlated Rayleigh fading: tion [20, 21]:
max. Doppler freq., fm (Hz) 2.0 ⎧
⎨1 − 7.2 f 2 + 0.785 f 4 ,
Independent lognormal shadowing: f0 ≤ 1,
0 0
standard deviation (dB) 8.0 S( f ) = ⎩ (11)
0, f0 > 1.
Noise power, PN (dBW) −134.06
Frame length (ms) 5.0
In the above, f0 = f / fm , where fm is the maximum Doppler
Average data rate per user (kbps) 404.16
frequency.
Simulation tool used OPNET Modeler 9.1 [19]
With a channel bandwidth of 3.0 MHz and noise figure
(NF) of 5 dB, the average noise power is −134.06 dBW.
To evaluate the proposed scheme, real-time video traffic
solution would result in some degradation in performance in is used in this study. Two interrupted renewal process (2IRP)
terms of overall network interference; however, this solution sources are superimposed to model the user’s video traffic in
would still control in-group interference for a subset BSs in the downlink transmission as indicated in [22]. The average
the group. Secondly, even when a BS is a member of different packet rate of one 2IRP generator is 126.3 packets per second
interferer groups and receives different commands, a second determined from parameters given in Table 4. The length of
tier of the control scheme (e.g., the majority rule algorithm) packets is assumed to be variable and is uniformly distributed
can be employed to resolve the conflicts. For instance, when between 250 to 550 bytes. Therefore, the average downlink
a BS is a member of three groups, it can only transmit when data rate for each SS is 404.16 kbps.
the decisions from two or more groups go in favor of trans- End-to-end packet delay is the summation of queuing de-
missions. lay and packet transmission delay. Packet transmission delay
depends on the packet size L p , symbol rate of the transmis-
4. SYSTEM MODEL sion channel rs , modulation level M, and coding rate rc , and
is expressed as
Table 3 summarizes system parameters used in this study.
The path-loss model has been taken from [20, 21]. For a Lp
td = . (12)
transmitter-receiver (T-R) separation of d meters the large- rs rc log2 M
scale path-loss (in linear scale) PL including shadowing is
given by the following relation: We assume asynchronous transmission such that inter-
ferers may arrive or leave anytime during the transmission
⎧ 0.6 −2
⎪
⎪ 4πd0 2 d n f hr time of a packet of interest. Therefore, SINR varies, and the
⎪
⎪ 10Xσ /10 , d ≥ d0 ,
⎪
⎨ λ d0 2000 2 packet experiences different bit error rates at different seg-
PL = ⎪ ments of the packet. The number of erroneous bits in a seg-
⎪
⎪
⎪
⎪ 4πd 2 Xσ /10 ment s is given by the product of the probability of the bit er-
⎩ 10 , d < d0 , ror in the segment Prb(s) and the number of bits correspond-
λ
(10) ing to the segment length Nb(s) . The total number of bits in
error in the packet Ne can be written by the following rela-
where n is the propagation exponent (we have taken n = 3.75 tion:
for 50-meter antenna height in terrain type C; see [20, 21]
for details). Parameter d0 is the close-in reference distance
S
Ne = prb(s) Nb(s) , (13)
considered to be 50 m, f is the operating frequency in MHz, s=1
λ is the operating wavelength related to speed of light c and
operating frequency f , and hr is the receiver antenna height where S is the total number of segments in that packet expe-
in meters which is considered to be 3 meters. Parameter Xσ riencing different SINR.
is a Gaussian distributed random variable with a mean of 0 The total number of erroneous bits is used to decide
and a standard deviation of σ used for shadowing. We have whether the packet is received correctly. In simulations, we
90 0.14
Percentage of orthogonal and non-orthogonal
80
0.12
transmissions in ISONOISS-AMC
70
0.1
60
Packet error rate

50 0.08
40 0.06
30
0.04
20
0.02
10
0 0
4 8 12 14 16 18 20 24 4 6 8 12 14 16 18 20 24
Network loading (SSs/sector) Network loading (SSs/sector)
1 BS transmits ISOISS-FM ISONOISS-AMC

2 BSs transmit ISOISS-AMC ISONOISS-AMC with guard
3 BSs transmit
Figure 4: Packet error rate in different schemes.
Figure 3: Percentage of single and multiple transmissions in
ISONOISS-AMC.
assume that a packet is considered to be in error if more than The network simulation is executed in real time, using
1% of the total bits present in the packet are erroneous. Re- OPNET [19] Modeler and wireless module, and the statistics
transmissions of erroneous packets by automatic repeat re- are taken over a long enough time for the observed param-
quest (ARQ) are not considered in this study. eters to converge. It should be noted that shadowing for a
The frame length is considered to be 5 milliseconds. particular SS does not change over simulation time as the SS
Packets are scheduled in a frame-by-frame basis at the start location is fixed. At any loading, a set of shadowing values is
of every frame. Any packet arriving at current frame time will assigned for all SSs (randomly placed) in the network. Dur-
have to wait at least until the start of the next frame. ing the course of simulation time, neither the locations of SSs
nor the shadowing values are changed. For any particular SS,
5. SIMULATION RESULTS fading is correlated and it changes over time. Therefore, per-
formed simulation is Monte Carlo in the time axis, but not
The performance of the proposed scheduling scheme for SS locations and shadowing. However, statistics are col-
ISONOISS-AMC is evaluated by comparing it with that of lected in sectors of all nine cells in the network, and hence
the reference scheme ISOISS-FM in terms of the essential there is a certain degree of averaging with respect to the SS
network performance parameters such as packet error rate, locations.
area spectral efficiency, packet dropping rate, and the mean Figure 3 shows the percentage of the scheduling deci-
end-to-end packet delay. Also, the performance of ISOISS- sions that yields into 1, 2, and 3 (all) in-group BSs transmis-
AMC is shown in order to quantify the benefits of employ- sions in ISONOISS-AMC scheme. We observe that around
ing AMC alone. These performance metrics are functions of 35% of the time, the scheme is capable of using opportunis-
network loading and are observed against the number of SSs tic nonorthogonality in transmissions (all three BSs trans-
per sector (varied from 4 to 24). mit 5% of the time and any 2 BSs transmit 30% of the
The packet error rate is the ratio of the number of erro- time) giving higher aggregate spectral efficiency than single
neous packets to the total packets received during the sim- transmission.
ulation period. The area spectral efficiency is expressed as Figure 4 compares the packet error rate performance
the correctly received information bits per second per Hz per of the proposed, reference, and intermediate schemes. The
sector. Packet is dropped from the BS queue when the queu- modulation and coding level used in the reference ISOISS-
ing delay exceeds 195 milliseconds. The delay constraint is as- FM scheme is more robust than the channel-state-based cho-
sumed to be 200 milliseconds (For interactive video, such as sen AMC modes in the proposed ISONOISS-AMC scheme.
videoconferencing) with a 5- milliseconds safety margin pro- Also, increased number of packets in the air results in in-
vided to ensure that every packet received by the SS meets the creased number of out-of-group interferers in ISONOISS-
delay requirement. We express packet dropping rate in pack- AMC scheme. Consequently, the packet error rate in pro-
ets per frame per sector. The mean end-to-end delay mea- posed scheme is higher. The packet error rate of ISOISS-
sure does not include the delays of the dropped packets in AMC fall in between the reference and proposed schemes as
the queue at transmitter side. ISOISS-AMC suffers less from interference in comparison to
2.2 0.2
0.18
Area spectral efficiency (bps/Hz/sector)
1.8 0.16
0.14
Mean end-to-end delay (s)

1.6
0.12
1.4
0.1
1.2
0.08
1
0.06
0.8 0.04
0.6 0.02
0.4 0
4 6 8 12 16 18 20 24 4 6 8 12 14 16 18 20 24
Network loading (SSs/sector) Network loading (SSs/sector)
ISOISS-FM ISONOISS-AMC ISOISS-FM ISONOISS-AMC

ISOISS-AMC ISONOISS-AMC with guard ISOISS-AMC ISONOISS-AMC with guard
Figure 5: Area spectral efficiency in different schemes. Figure 7: Mean end-to-end packet delay in different schemes.
11 ber of packets per frame can be transmitted in these schemes.

10 While the area spectral efficiency in ISOISS-FM is limited
Net throughput (packets/frame/sector)
by around 0.6 bps/Hz/sector, the proposed ISONOISS-AMC

9 shows an area spectral efficiency of around 2.2 bps/Hz/sector
8 at the network loading of 24 SSs per sector. ISOISS-AMC
delivers spectral efficiency of around 1.65 bps/Hz/sector at
7 the same loading. At this loading value, around 3 times
6 higher area spectral efficiency and throughput are achieved
in the ISONOISS-AMC compared to those obtained in the
5 ISOISS-FM. Improvements in ISONOISS-AMC compared
4 to ISOISS-AMC are solely due to the benefits of in-group
opportunistic multiple transmissions. As employing out-of-
3 group interference guard in the proposed scheme led the
2 schedulers to choose AMC modes conservatively, the area
4 6 8 12 14 16 18 20 24 spectral efficiency and net throughput are reduced slightly.
Network loading (SSs/sector) However, while packet error rates are similar in ISONOISS-
ISOISS-FM ISONOISS-AMC
AMC with guard and in ISOISS-AMC, the former achieves
ISOISS-AMC ISONOISS-AMC with guard much higher area spectral efficiency and net throughput.
Figure 7 illustrates the mean end-to-end delay. We ob-
serve that the delay reaches the threshold 200 milliseconds
Figure 6: Net throughput in different schemes.
for a loading level as low as 6 SSs per sector in the ISOISS-
FM scheme. Because of less efficient AMC mode usage, fewer
packets get transmitted per frame in the ISOISS-FM scheme.
ISONOISS-AMC. However, when out-of-group interference As a result, the queue length grows even at very low load-
guard is considered in ISONOISS-AMC, packet error rate is ing levels such as 5 or 6 SSs per sector, causing high mean
reduced drastically and the resulting error rate is comparable end-to-end delay. In ISONOISS-AMC, on the other hand,
to that of ISOISS-AMC. the queues grow at much higher loading levels, as the pro-
We present area spectral efficiency and net throughput in posed scheme is able to use efficient AMC modes, and it al-
Figures 5 and 6, respectively. Although packet error rate is lows concurrent transmissions among in-group BSs. There-
high, ISOISS-AMC and ISONOISS-AMC show tremendous fore, we notice a much better delay performance in the pro-
improvements in terms of area spectral efficiency and net posed scheme compared to the reference scheme. For in-
throughput. This is because the intermediate and proposed stance, for a mean delay of 50 milliseconds, ISOISS-FM sup-
schemes are capable of using much higher AMC modes ports only 4 SSs, while ISONOISS-AMC is able to support
whenever possible in comparison to 16-QAM with a coding 16 SSs in a sector. Once again, the mean end-to-end de-
rate of 1/2 mode used in ISOISS-FM; therefore, a larger num- lay in ISOISS-AMC falls between those in ISOISS-FM and
14 AMC, is still orthogonal, while it makes link-state-based

Packet dropping rate (packets/frame/sector)
scheduling decisions and uses AMC. Finally, the proposed

12 interference-aware scheme, ISONOISS-AMC, makes link-
state-based scheduling decisions, employs AMC, and al-
10
lows controlled in-group interference in order to improve
throughput and packet delay.
8
It has been observed that the area spectral efficiency in
ISONOISS-AMC is around three times higher than that in
6
ISOISS-FM. Moreover, higher throughput results in signif-
4
icant improvements in end-to-end packet delay and packet
dropping rate in ISONOISS-AMC. To quantify the ben-
2 efits of AMC alone, we also have studied ISOISS-AMC,
which outperforms the reference scheme in terms of area
0 spectral efficiency, net throughput, mean end-to-end delay,
4 6 8 12 14 16 18 20 24 and packet dropping rate. The proposed ISONOISS-AMC
Network loading (SSs/sector) achieves up to 33% better area spectral efficiency than the in-
termediate ISOISS-AMC scheme. This improvement is solely
ISOISS-FM ISONOISS-AMC
ISOISS-AMC ISONOISS-AMC with guard
due to the opportunistic nonorthogonal transmissions in the
proposed scheme.
While the proposed scheme shows performance im-
Figure 8: Packet dropping rate in different schemes.
provements in terms of area spectral efficiency, delay, and
packet dropping rate, it experiences higher packet error rate
due to increased number of uncontrolled out-of-group in-
terferers. However, when out-of-group interference guard
in ISONOISS-AMC as expected. Observed improved delay is used in ISONOISS-AMC, the packet error rate becomes
performance in ISONOISS-AMC compared to ISOISS-AMC comparable to that observed in ISOISS-AMC. Nevertheless,
is due to the simultaneous in-group transmissions in the if even 10% packet error rate is allowed by the upper layer,
ISONOISS-AMC scheme. When out-of-group interference the proposed ISONOISS-AMC can support as many as 16 SSs
guard is used in ISONOISS-AMC, the mean end-to-end de- per sector with mean packet delay of around 50 milliseconds
lay increases slightly, however, it is always less than that in and the reasonable packet dropping rate, while ISOISS-FM
ISOISS-AMC. supports only 4 SSs. For the similar packet error rate and
The comparison of packet dropping rate is shown in mean end-to-end delay, the ISOISS-AMC scheme can ac-
Figure 8. ISONOISS-AMC shows much better performance commodate 13 SSs per sector.
than ISOISS-FM in terms of packet dropping rate for the
same reasons as for the delay. The packet dropping rate in
ACKNOWLEDGMENTS
the intermediate scheme ISOISS-AMC is lower than that
obtained in ISOISS-FM and higher than that observed in The authors would like to thank OPNET Technologies, Inc.
ISONOISS-AMC. for providing software license to carry out the simulations of
It is observed that the performances of ISOISS-AMC and this research. The authors are grateful to Dr. Keivan Navaie
ISONOISS-AMC are comparable until the loading level of for his review and comments. This research has been funded
12 users/sector. This is due to the fact that at this point of in part by National Capital Institute of Telecommunications
loading, ISOISS-AMC becomes fully loaded and packets start (NCIT), Ottawa, Canada. Part of this paper has been pre-
to drop, while ISONOISS-AMC still has some capacity left sented at the Proceedings of IEEE International Conference
in the frame. The difference in performance increases as the on Communications (ICC), 16–20 May 2005, Seoul, Korea.
loading values grow further beyond this point. Simulations
are prevented from going beyond 24 users/sector due to the REFERENCES
long simulation time needed. However, the trends of the per-
formance curves show that the benefits in ISONOISS-AMC [1] IEEE Std. 802.16a for Local and Metropolitan Area Networks,
are even higher at higher loading than presented here. Air interface for fixed broadband wireless access systems,
IEEE, April 2003.
[2] H. Bölcskei, A. J. Paulraj, K. V. S. Hari, R. U. Nabar, and W.
6. CONCLUSIONS W. Lu, “Fixed broadband wireless access: state of the art, chal-
lenges, and future directions,” IEEE Communications Maga-
The benefits of combining link-state-based scheduling de-
zine, vol. 39, no. 1, pp. 100–108, 2001.
cisions, AMC, and opportunistic nonorthogonal transmis- [3] D. L. Waring, “The asymmetrical digital subscriber line
sions in fixed broadband wireless access networks have been (ADSL): a new transport technology for delivering wideband
investigated in this paper. A reference orthogonal schedul- capabilities to the residence,” in Proceedings of IEEE Global
ing scheme that makes arrival-time-based scheduling deci- Telecommunications Conference and Exhibition (GLOBECOM
sions and uses fixed modulation, namely ISOISS-FM, has ’91), vol. 3, pp. 1979–1986, Phoenix, Ariz, USA, December
been adapted from [13]. The intermediate scheme, ISOISS- 1991.
[4] K. Maxwell, “Asymmetric digital subscriber line (ADSL): in- [18] M. Rahman, “Adaptive modulation & coding-based packet
terim technology for the next forty years,” IEEE Communica- scheduling with inter-base station coordination in cellular
tions Magazine, vol. 34, no. 10, pp. 100–106, 1996. fixed broadband wireless networks,” M.S. thesis, Carleton
[5] C. Bisdikian, K. Maruyama, D. I. Seidman, and D. N. Ser- University, Ottawa, Ontario, Canada, 2004, http://www.sce
panos, “Cable access beyond the hype: on residential broad- .carleton.ca/faculty/yanikomeroglu/cv/publications.shtml.
band data services over HFC networks,” IEEE Communications [19] OPNET University Program: http://www.opnet.com/services/
Magazine, vol. 34, no. 11, pp. 128–135, 1996. university.
[6] P. Bhagwat, P. Bhattacharya, A. Krishna, and S. K. Tripathi, [20] V. Erceg, L. J. Greenstein, S. Y. Tjandra, et al., “An empirically
“Enhancing throughput over wireless LANs using channel based path loss model for wireless channels in suburban envi-
state dependent packet scheduling,” in Proceedings IEEE of ronments,” IEEE Journal on Selected Areas in Communications,
15th Annual Joint Conference of the IEEE Computer Societies vol. 17, no. 7, pp. 1205–1211, 1999.
(INFOCOM ’96), vol. 3, pp. 1133–1140, San Francisco, Calif, [21] V. Erceg, et al., Channel models for fixed wireless applications,
USA, March 1996. IEEE 802.16 work in progress document # 802.16.3c-01/29r4,
[7] T. S. E. Ng, I. Stoica, and H. Zhang, “Packet fair queueing al- July 2001.
gorithms for wireless networks with location-dependent er- [22] C. R. Baugh and J. Huang, Traffic model for 802.16 TG3
rors,” in Proceedings of IEEE 17th Annual Joint Conference of MAC/PHY simulations, IEEE 802.16 work in progress docu-
the IEEE Computer and Communications Societies (INFOCOM ment # 802.16.3c-01/30r1, March 2001.
’98), vol. 3, pp. 1103–1111, San Francisco, Calif, USA, March-
April 1998.
[8] S. Desilva and S. M. Das, “Experimental evaluation of channel Mahmudur Rahman received the B.S. de-
state dependent scheduling in an in-building wireless LAN,” gree in electrical and electronic engineer-
in Proceedings of 7th International Conference on Computer ing from Bangladesh University of Engi-
Communications and Networks (ICCCN ’98), pp. 414–421, neering and Technology (BUET), Dhaka,
Lafayette, La, USA, October 1998. Bangladesh, in 1991. He obtained an
[9] C. Fragouli, V. Sivaraman, and M. B. Srivastava, “Controlled M.Eng. degree in telecommunications
multimedia wireless link sharing via enhanced class-based from Asian Institute of Technology (AIT),
queuing with channel-state-dependent packet scheduling,” in Bangkok, Thailand, and an M.A.S. degree
Proceedings of IEEE 17th Annual Joint Conference of the IEEE in electrical engineering from Carleton
Computer and Communications Societies (INFOCOM ’98), University, Ottawa, Canada, in 1994 and 2004, respectively. He
vol. 2, pp. 572–580, San Francisco, Calif, USA, March-April received Finnish International Development Agency (FINNIDA)
1998. Scholarship for his studies at AIT. He worked as an Electron-
[10] S. Lu, V. Bharghavan, and R. Srikant, “Fair scheduling in wire- ics Engineer in Bangladesh Atomic Energy Commission, Dhaka,
less packet networks,” IEEE/ACM Transactions on Networking, Bangladesh, from 1991 to 1993. From 1995 to 1996, he was a
vol. 7, no. 4, pp. 473–489, 1999. Process Engineer in Johnson Electric Industrial Manufactory, Ltd.,
[11] T. K. Fong, P. S. Henry, K. K. Leung, X. Qiu, and N. (Thailand). Initially appointed to the position of Senior R&D En-
K. Shankaranarayanan, “Radio resource allocation in fixed gineer in 1996, he served ACE Electronics Industries Co., Ltd.,
broadband wireless networks,” IEEE Transactions on Commu- Bangkok, Thailand, as an R&D Division Manager from 1997 to
nications, vol. 46, no. 6, pp. 806–818, 1998. 1999. He is currently working towards a Ph.D. degree in electri-
[12] M. H. Ahmed, H. Yanikomeroglu, and S. Mahmoud, “Interfer- cal engineering at Carleton University. He is involved in the Wire-
ence management using packet sheduling in broadband wireless World Initiative New Radio (WINNER) Project. His current
less access networks,” Willey Journal on Wireless Communica- research interests include radio resource management, multihop
tions and Mobile Computing, vol. 6, no. 1, pp. 95–103, 2006. wireless networks, and intercell coordination.
[13] M. H. Ahmed, H. Yanikomeroglu, S. Mahmoud, and D. D.
Falconer, “Scheduling of multimedia traffic in interference- Halim Yanikomeroglu received a B.S. de-
limited broadband wireless access networks,” in Proceedings of gree in electrical and electronics engi-
5th International Symposium on Wireless Personal Multimedia neering from the Middle East Technical
Communications (WPMC ’02), Honolulu, Hawaii, USA, Octo- University, Ankara, Turkey, in 1990, and
ber 2002. an M.A.S. degree in electrical engineer-
[14] K. K. Leung and A. Srivastava, “Dynamic allocation of down- ing (now ECE), and a Ph.D. degree in
link and uplink resource for broadband services in fixed wire- electrical and computer engineering from
less networks,” IEEE Journal on Selected Areas in Communica- the University of Toronto, Canada, in
tions, vol. 17, no. 5, pp. 990–1006, 1999. 1992 and 1998, respectively. He was with
[15] F. Borgonovo, M. Zorzi, L. Fratta, V. Trecordi, and G. Bianchi, the Research and Development Group of
“Capture-division packet access for wireless personal commu- Marconi Kominikasyon A.S., Ankara, Turkey, from January 1993
nications,” IEEE Journal on Selected Areas in Communications, to July 1994. Since 1998, he has been with the Department of
vol. 14, no. 4, pp. 609–622, 1996. Systems and Computer Engineering at Carleton University, Ot-
[16] K. Chawla and X. Qiu, “Quasi-static resource allocation with tawa, where he is now an Associate Professor and Associate Chair
interference avoidance for fixed wireless systems,” IEEE Jour- for Graduate Studies. His research interests include almost all
nal on Selected Areas in Communications, vol. 17, no. 3, pp. aspects of wireless communications with a special emphasis on
493–504, 1999. infrastructure-based multihop/mesh/relay networks. He has been
[17] X. Liu, E. K. P. Chong, and N. B. Shroff, “Opportunistic trans- involved in the steering committees and technical program com-
mission scheduling with resource-sharing constraints in wire- mittees of numerous international conferences in communications;
less networks,” IEEE Journal on Selected Areas in Communica- he has also given several tutorials in such conferences. He was the
tions, vol. 19, no. 10, pp. 2053–2064, October 2001. Technical Program Cochair of the IEEE Wireless Communications
and Networking Conference 2004 (WCNC’04). He was an Editor

for IEEE Transactions on Wireless Communications during 2002–
2005, and a Guest Editor for Wiley Journal on Wireless Commu-
nications & Mobile Computing; he was an Editor for IEEE Com-
munications Surveys & Tutorials for 2002–2003. Currently he is
serving as the Chair of the IEEE Communications Society’s Tech-
nical Committee on Personal Communications (TCPC), he is also
a Member of IEEE ComSoc’s Technical Activities Counsel (TAC).
Mohamed H. Ahmed received his B.S. and

M.S. degrees in electronics and communi-
cations engineering from Ain Shams Uni-
versity, Cairo, Egypt, in 1990 and 1994, re-
spectively. He obtained his Ph.D. degree in
electrical engineering in 2001 from Car-
leton University, Ottawa. From 2001 to
2003, he worked as a Senior Research As-
sociate at the Department of Systems and
Computer Engineering, Carleton University. In April 2003, he
joined the Faculty of Engineering and Applied Science, Memorial
University of Newfoundland as an Assistant Professor of electrical
and computer engineering. He served as a Technical Program Com-
mittee Member of various conferences and a Guest Editor for Wi-
ley Journal on Wireless Communications & Mobile Computing. He
won the Ontario Graduate Scholarship for Science and Technology
in 1997, the Ontario Graduate Scholarship in 1998, 1999, and 2000,
and Communication and Information Technology Ontario (CITO)
Graduate Award in 2000. His research interests include wireless ac-
cess techniques, resource management in wireless networks, multi-
hop wireless networks, MIMO antenna systems, cooperative diver-
sity, and broadband wireless networks.
Samy Mahmoud is the Dean of the Fac-

ulty of Engineering and Design, Carleton
University. He was the Chair of the De-
partment of Systems and Computer Engi-
neering, Carleton University from 1987 to
1997. His main research work is in the areas
of wireless communication systems and the
transmission of voice and video signals over
high-speed networks. He has published over
200 papers in these fields in recent years and supervised 32 doc-
toral and 85 M.Eng. theses to completion. He graduated with the
M.Eng. and Ph.D. degrees in electrical engineering from Carleton
University in 1971 and 1975, respectively. He coauthored a major
book on “Communication Systems Analysis and Design,” published
in 2004 by Pearson Prentice Hall. His recent research work in the
field of radio-over-fiber has resulted in several archival publications
and 8 patent applications. His research work received several in-
ternational and national awards, including two Best Paper Awards
for publications in the IEEE Transactions on Vehicular Technol-
ogy and two TRIO Feedback Awards for best technology transfer
to industry. He recently led a major initiative to establish the Na-
tional Capital Institute of Telecommunications (NCIT), a joint re-
search organization involving several large international companies
in the telecommunications and computer industries, leading uni-
versity researchers, scientists, and engineers from two major Cana-
dian Government Research Laboratories (CRC and NRC).
DOI 10.1155/WCN/2006/36424
Rate-Optimal Multiuser Scheduling with Reduced Feedback

Load and Analysis of Delay Effects
Vegard Hassel,1 Mohamed-Slim Alouini,2 Geir E. Øien,1 and David Gesbert3

1 Departmentof Electronics and Telecommunications, Norwegian University of Science and Technology, 7491 Trondheim, Norway
2 Departmentof Electrical Engineering, Texas A&M University at Qatar (TAMUQ), Education City, P.O. Box 5825, Doha, Qatar
3 Mobile Communications Department, Eurécom Institute, 06904 Sophia Antipolis, France
Received 30 September 2005; Revised 13 March 2006; Accepted 26 May 2006

We propose a feedback algorithm for wireless networks that always collects feedback from the user with the best channel conditions
and has a significant reduction in feedback load compared to full feedback. The algorithm is based on a carrier-to-noise threshold,
and closed-form expressions for the feedback load as well as the threshold value that minimizes the feedback load have been found.
We analyze two delay scenarios. The first scenario is where the scheduling decision is based on outdated channel estimates, and the
second scenario is where both the scheduling decision and the adaptive modulation are based on outdated channel estimates.
Copyright © 2006 Vegard Hassel et al. This is an open access article distributed under the Creative Commons Attribution License,
1. INTRODUCTION uler does not receive a feedback, a random user is chosen.

Because the best user is not chosen for every time slot, the
In a wireless network, the signals transmitted between the SMUD algorithm however introduces a reduction in system
base station and the mobile users most often have different spectral efficiency. In addition it can be hard to set the thresh-
channel fluctuation characteristics. This diversity that exists old value for this algorithm. Applying a high threshold value
between users is called multiuser diversity (MUD) and can be will lead to low feedback load, but will additionally reduce
exploited to enhance the capacity of wireless networks [1]. the MUD gain and hence the system spectral efficiency. Using
One way of exploiting MUD is by opportunistic scheduling a low threshold value will have the opposite effect: the feed-
of users, giving priority to users having good channel condi- back load reduction is reduced, but the spectral efficiency will
tions [2, 3]. Ignoring the feedback loss, the scheduling algo- be higher.
rithm, that maximizes the average system spectral efficiency The feedback algorithm proposed here is inspired by the
among all time division multiplexing- (TDM-) based algo- SMUD algorithm, in the sense that this new algorithm also
rithms, is the one where the user with the highest carrier-to- employs a feedback threshold. However, if none of the users
noise ratio (CNR) is served in every time slot [2]. Here, we succeeds to exceed the CNR threshold, the scheduler requests
refer to this algorithm as max CNR scheduling (MCS). full feedback, and selects the user with the highest CNR.
To be able to take advantage of the MUD, a base station Consequently, the MUD gain [1] is maximized, and still the
needs feedback from the mobile users. Ideally, the base sta- feedback load is significantly reduced compared to the MCS
tion only wants feedback from the user with the best channel algorithm. Another advantage with this novel algorithm is
conditions, but unfortunately each user does not know the that for a specific set of system parameters it is possible to
CNR of the other users. Therefore, in current systems like find a threshold value that minimizes the feedback load.
Qualcomm’s high data rate (HDR) system, the base station For the new feedback algorithm we choose to investigate
collects feedback from all the users [4]. two important issues, namely, (i) how the algorithm can be
One way to reduce the number of users giving feedback optimized, and (ii) the consequences of delay in the system.
is by using a CNR threshold. For the selective multiuser diver- The first issue is important because it gives theoretical lim-
sity (SMUD) algorithm, it is shown that the feedback load its for how well the algorithm will perform. The second issue
is reduced significantly by using such a threshold [5]. For is important because the duration of the feedback collection
this algorithm only the users that have a CNR above a CNR process will often be significant and this will lead to a re-
threshold should send feedback to the scheduler. If the sched- duced performance of the opportunistic scheduling since the
feedback information will be outdated. The consequences of given point in time. The system uses the first half of the time
delay are analyzed by looking separately at two different ef- slot for downlink and the last half for uplink transmission.
fects: (a) the system spectral efficiency degradation arising The users measure their channel for each downlink transmis-
because the scheduler does not have access to instantaneous sion and this measurement is fed back to the base station so
information about CNRs of the users, and (b) the bit error that it can decide which user is going to be assigned the next
rate (BER) degradation arising when both the scheduler and time slot. The second system model is a system where differ-
the mobile users do not have access to instantaneous channel ent carriers are used for uplink and downlink. For the base
measurements. station to be able to schedule the user with the best down-
link channel quality, the users must measure their channel
Contributions for each downlink transmission and feed back their CNR
measurement. For both system models the users are notified
We develop closed-form expressions for the feedback load of about the scheduling decision in a short broadcast message
the new feedback algorithm. The expression for the thresh- from the base station between each time slot.
old value which minimizes the feedback load is also derived.
In addition we obtain new closed-form expressions for the 3. ANALYSIS OF THE FEEDBACK LOAD
system spectral efficiency degradation due to the scheduling
delay. Finally, closed-form expressions for the effects of out- The first step of the new feedback algorithm is to ask for feed-
dated channel estimates are obtained. Parts of the results have back from the users that are above a CNR threshold value γth .
previously been presented in [6]. The number of users n being above the threshold value γth is
random and follow a binomial distribution given by

Organization N n
Pr(n) = 1 − Pγ γth PγN −n γth , n = 1, 2, . . . , N,
The rest of this paper is organized as follows. In Section 2, we n
present the system model. The feedback load is analyzed in (1)
Section 3, while Sections 4 and 5 analyze the system spectral where Pγ (γ) is the cumulative distribution function (CDF)
efficiency and BER, respectively. In Section 6 the effects of of the CNR for a single user. The second step of the feedback
delay are discussed. Finally, Section 7 lists our conclusions. algorithm is to collect full feedback. Full feedback is only
needed if all users’ CNRs fail to exceed the threshold value.
2. SYSTEM MODEL The probability of this event is given by inserting γ = γth into
Pγ∗ (γ) = PγN (γ), (2)
We consider a single cell in a wireless network where the base
station exchanges information with a constant number N of where γ∗ denotes the CNR of the user with the best channel
mobile users which have identically and independently dis- quality.
tributed (i.i.d.) CNRs with an average of γ. The system con- We now define the normalized feedback load (NFL) to be
sidered is TDM-based, that is, the information transmitted the ratio between the average number of users transmitting
in time slots with a fixed length. We assume flat-fading chan- feedback, and the total number of users. The NFL can be ex-
nels with a coherence time of one time slot, which means that pressed as a the average of the ratio n/N, where n is the num-
the channel quality remains roughly the same over the whole ber of users giving feedback:
time slot duration and that this channel quality is uncorre-
N N n N
N
n
lated from one time slot to the next. The system uses adap- F= Pγ γth + 1 − Pγ γth PγN −n γth
tive coding and modulation, that is, the coding scheme, the N n=1
N n
modulation constellation, and the transmission power used N
N −1
depend on the CNR of the selected user [7]. This has two ad- = PγN γth + 1 − Pγ γth
vantages. On one hand, the spectral efficiency for each user n=1
n−1
n−1 N −n
is increased. On the other hand, because the rate of the users × 1 − Pγ γth Pγ γth = PγN γth
is varied according to their channel conditions, it makes it −1

possible to exploit MUD. N N −1 k
+ 1 − Pγ γth 1 − Pγ γth PγN −1−k γth
We will assume that the users always have data to send k=0
k
and that these user data are robust with respect to delay, that N

= 1 − Pγ γth + Pγ γth , N = 2, 3, 4, . . . ,
is, no real-time traffic is transmitted. Consequently, the base
station only has to take the channel quality of the users into (3)
account when it is performing scheduling. where the last equality is obtained by using binomial expan-
The proposed feedback algorithm is applicable in at least sion [8, equation (1.111)]. For N = 1 full feedback is needed,
two different types of cellular systems. The first system model and F = 1. In that case the feedback is not useful for mul-
is a time-division duplex (TDD) scenario, where the same tiuser scheduling, but for being able to adapt the base sta-
carrier frequency is used for both uplink and downlink. We tion’s modulation according to the channel quality in the re-
can therefore assume a reciprocal channel for each user, that ciprocal TDD system model described in the previous sec-
is, the CNR is the same for the uplink and the downlink for a tion.
Vegard Hassel et al. 3
Normalized feedback load for average CNR of 15 dB This pdf can be obtained by differentiating (2) with respect
100
to γ. Inserting the CDF and pdf for Rayleigh fading chan-
90 nels (pγ (γ) = (1/γ)e−γ/γ ), and using binomial expansion [8,
Feedback load relative to full feedback (%)
2 users
equation (1.111)], we obtain
80
N −1
70 N N −1
pγ∗ (γ) = (−1)n e−(1+n)γ/γ . (6)
60 γ n=0 n
5 users
50 Inserting (6) into the expression for the spectral efficiency
40 10 users for optimal rate adaptation found in [9], the following ex-
pression for the MASSE can be obtained [10, equation (44)]:
30
50 users ∞
C ora
20 = log2 (1 + γ)pγ∗ (γ)dγ
W 0
10 N −1 (7)
N N −1 ne
(1+n)/γ 1+n
0 = (−1) E1 ,
0 5 10 15 20 25 30 ln 2 n=0 n 1+n γ
CNR threshold (dB)
where ora denotes optimal rate adaptation and E1 (·) is the
first-order exponential integral function [8].
Figure 1: Normalized feedback load as a function of γth with γ =
15 dB.
4.2. Optimal power and rate adaptation
It has been shown that the MASSE for optimal power and
A plot of the feedback load as a function of γth is shown
rate adaptation can be obtained as [10, equation (27)]
in Figure 1 for γ= 15 dB. It can be observed that the new
algorithm reduces the feedback significantly compared to a ∞
C opra γ
system with full feedback. It can also be observed that one = log2 pγ∗ (γ)dγ
W 0 γ0
threshold value will minimize the feedback load in the sys-
N −1 (8)
tem for a given number of users. N N − 1 (−1)n (1 + n)γ0
= E1 ,
The expression for the threshold value that minimizes ln 2 n=0 n 1+n γ
the average feedback load can be found by differentiating (3)
with respect to γth and setting the result equal to zero: where opra denotes optimal power and rate adaptation and γ0
1/(N −1) is the optimal cutoff CNR level below which data transmis-
∗ 1 sion is suspended. This cutoff value must satisfy [9]
γth = Pγ−1 , N = 2, 3, 4, . . . , (4)
N ∞
1 1
−1 − pγ∗ (γ)dγ = 1. (9)
where Pγ (·) is the inverse CDF of the CNR. In particular, γ0 γ0 γ
for a Rayleigh fading channel, with CDF Pγ (γ) = 1 − e−γ/γ ,
the optimum threshold can be found in a simple closed form Inserting (6) into (9), it can subsequently be shown that the
as following cutoff value can be obtained for Rayleigh fading
1/(N −1) channels [10, equation (24)]:
∗ 1
γth = −γ ln 1 − , N = 2, 3, 4, . . . . (5) −1
N
N N −1 e−(1+n)γ0 /γ (1 + n)γ0 γ
(−1)n − E1 = .
n=0
n (1 + n)γ0 /γ γ N
4. SYSTEM SPECTRAL EFFICIENCIES FOR DIFFERENT (10)
POWER AND RATE ADAPTATION TECHNIQUES
To be able to analyze the system spectral efficiency we choose 5. M-QAM BIT ERROR RATES
to investigate the maximum average system spectral efficiency The BER of coherent M-ary quadrature amplitude modula-
(MASSE) theoretically attainable. The MASSE (bit/s/Hz) is tion (M-QAM) with two-dimensional Gray coding over an
defined as the maximum average sum of spectral efficiency additive white Gaussian noise (AWGN) channel can be ap-
for a carrier with bandwidth W (Hz). proximated by [11]

4.1. Constant power and optimal rate adaptation 3γ
BER(M, γ) ≈ 0.2 exp − . (11)
2(M − 1)
Since the best user is always selected, the MASSE of the new
algorithm is the same as for the MCS algorithm. To find the The constant-power adaptive continuous rate (ACR) M-
MASSE for such a scenario, the probability density function QAM scheme can always adapt the rate to the instanta-
(pdf) of the highest CNR among all the users has to be found. neous CNR. From [12] we know that the constellation size
for continuous-rate M-QAM can be approximated by M ≈ take delay into consideration. We analyze, in what follows,
(1+3γ/2K0 ), where K0 = − ln(5 BER0 ) and BER0 is the target two delay scenarios. In the first scenario, a scheduling delay
BER. Consequently, it can be easily shown that the theoreti- arises because the scheduler receives channel estimates, takes
cal constant-power ACR M-QAM scheme always operates at a scheduling decision, and notifies the selected user. This user
the target BER. then transmits, but at a possibly different rate. The second
For physical systems only integer constellation sizes are scenario deals with outdated channel estimates, which leads
practical, so now we restrict the constellation size Mk to 2k , to both a scheduling delay as well as suboptimal modulation
where k is a positive integer. This adaptation policy is called constellations with increased BERs.
adaptive discrete rate (ADR) M-QAM, and the CNR range is Outdated channel estimates have been treated to some
divided into K + 1 fading regions with constellation size Mk extent in previous publications [12, 13]. However, the con-
assigned to the kth fading region. Because of the discrete as- cept of scheduling delay has in most cases been analyzed for
signment of constellation sizes in ADR M-QAM, this scheme wire-line networks only [14, 15]. Although some previous
has to operate at a BER lower than the target. The average work has been done on scheduling delay in wireless networks
BER for ADR M-QAM using constant power can be calcu- [16], scheduling delay has to the best of our knowledge not
lated as [12] been looked into for cellular networks.

K
k=1 kBERk 6.1. Impact of scheduling delay
BERadr =
K , (12)
k=1 k pk
In this section, we will assume that the scheduling decision is
where based on a perfect estimate of the channel at time t, whereas
γk+1 the data are sent over the channel at time t + τ. We will as-

BERk = BER Mk , γ pγ∗ (γ)dγ, (13) sume that the link adaptation done at time t+τ is based on yet
γk another channel estimate taken at t + τ. To investigate the in-
N N
pk = 1 − e−γk+1 /γ − 1 − e−γk /γ (14) fluence of this type of scheduling delay, we need to develop a
pdf for the CNR at time t +τ, conditioned on channel knowl-
is the probability that the scheduled user is in the fading re- edge at time t. Let α and ατ be the channel gains at times t
gion k for CNRs between γk and γk+1 . and t + τ, respectively. Assuming that the average power gain
Inserting (11) and (6) into (13) we obtain the following remains constant over the time delay τ for a slowly-varying
expression for the average BER within a fading region: Rayleigh channel (i.e., Ω = E[α2 ] = E[α2τ ]) and using the

same approach as in [12] it can be shown that the conditional
N −1
0.2N N − 1 e−γk ak,n − e−γk+1 ak,n pdf pατ |α (ατ | α) is given by
BERk = (−1)n , (15) √
γ n=0 n ak,n 2ατ 2 ραατ −(α2τ +ρα2 )/(1−ρ)Ω
pατ |α ατ | α = I0 e ,
(1 − ρ)Ω (1 − ρ)Ω
where ak,n is given by
(18)
1+n 3
ak,n = + . (16) where ρ is the correlation factor between α and ατ and I0 (·)
γ 2 Mk − 1 is the zeroth-order modified Bessel function of the first kind
When power adaptation is applied, the BER approxima- [8]. Assuming Jakes Doppler spectrum, the correlation co-
tion in (11) can be written as [11] efficient can be expressed as ρ = J02 (2π fD τ), where J0 (·) is
the zeroth-order Bessel function of the first kind and fD [Hz]
is the maximum Doppler frequency shift [12]. Recognizing
3γ Sk (γ)
BERpa (M, γ) ≈ 0.2 exp − , (17) that (18) is similar to [17, equation (A-4)] gives the follow-
2(M − 1) Sav
ing pdf at time t+τ for the new feedback algorithm, expressed
where Sk (γ) is the power used in fading region k and Sav in terms of γτ and γ [17, equation (5)]:
is the average transmit power. Inserting the continuous
−1
power adaptation policy given by [11, equation (29)] into N N exp − γτ /γ 1 − ρ n/(n + 1)
pγτ∗ γτ = (−1) n .
(17) shows that the ADR M-QAM scheme using optimal n=0
n+1 γ 1 − ρ n/(n + 1)
power adaptation always operates at the target BER. Cor- (19)
respondingly, it can be shown that the continuous-power,
continuous-rate M-QAM scheme always operates at the tar- Note that for τ = 0 (ρ = 1) this expression reduces to (6), as
get BER. expected. When τ approaches infinity (ρ = 0) (19) reduces
to the Rayleigh pdf for one user. This is logical since for large
6. CONSEQUENCES OF DELAY τs, the scheduler will have completely outdated and as such
useless feedback information, and will end up selecting users
In the previous sections, it has been assumed that there is independent of their CNRs.
no delay from the instant where the channel estimates are Inserting (19) into the capacity expression for opti-
obtained and fed back to the scheduler, to the time when the mal rate adaptation in [9, equation (2)], then using bino-
optimal user is transmitting. For real-life systems, we have to mial expansion, integration by parts, L’Hôpital’s rule, and
[8, equation (3.352.2)], it can be shown that we get the fol- Average MASSE degradation due to scheduling delay
8
lowing expression for the MASSE:
∞ 7
C ora 10 users
= log2 1 + γτ pγτ∗ γτ dγτ 5 users
W 0 6
MASSE C /W (bits/s/Hz)

−1
N 2 users
1 N
= (−1)n e1/γ(1−ρ(n/(n+1))) (20) 5
ln 2 n=0
n+1 1 user

1 4
× E1 .
γ 1 − ρ n/(n + 1)
3
Using a similar derivation as for the expression above it
2
can furthermore be shown that we get the following expres-
sion for the MASSE using both optimal power and rate adap- 1
tation:
∞ 0
C opra γτ 10 3 10 2 10 1 100 101
= log2 pγτ∗ γτ dγτ
W 0 γ0 Normalized time delay
−1
N
1 N γ Optimal power and rate adaptation
= (−1)n E1 0 , Constant power and optimal rate adaptation
ln 2 n=0
n+1 γ 1 − ρ n/(n + 1)
(21)
Figure 2: Average degradation in MASSE due to scheduling de-
with the following power constraint: lay for (i) optimal power and rate adaptation and (ii) optimal rate
adaptation.
−1
N
N e−1/γ(1−ρ(n/(n+1)))
(−1)n
n=0
n+1 γ0
The average BER can be found by using the following equa-
E1 1/γ 1 − ρ n/(n + 1) tion:
− = 1.
γ 1 − ρ n/(n + 1) ∞
(22) BERacr = BER(γ)pγ∗ (γ)dγ. (24)
0
Again, for zero time delay (ρ = 1), (20) reduces to (7), (21) For discrete rate adaptation with constant power, the BER
reduces to (8), and (22) reduces to (10), as expected.
can be expressed by (12), replacing BERk with BERk , where
Figure 2 shows how scheduling delay affects the MASSE γk+1 ∞
for 1, 2, 5, and 10 users. We see that both optimal power and
BERk = BER Mk , γτ pγτ |γ γτ | γ dγτ pγ∗ (γ)dγ.
rate adaptation and optimal rate adaptation are equally ro- γk 0
bust with regard to the scheduling delay. Independent of the (25)
number of users, we see that the system will be able to oper-
ate satisfactorily if the normalized delay is below the critical Inserting (6), (11), and (18) expressed in terms of γτ and γ
value of 2 · 10−2 . For normalized time delays above this value, into (25), we obtain the following expression for the average
we see that the MASSE converges towards the MASSE for one BER within a fading region:
user, as one may expect.
N −1
0.2N N − 1
e−γk ck,n − e−γk+1 ck,n
BERk = (−1)n , (26)
6.2. Impact of outdated channel estimates γ n=0 n dk,n
We will now assume that the transmitter does not have a per- where ck,n is given by
fect outdated channel estimate available at time t+τ, but only
at time t. Consequently, both the selection of a user and the 1+n 3ρ
ck,n = + , (27)
decision of the constellation size have to be done at time t. γ 3γ(1 − ρ) + 2 Mk − 1
This means that the channel estimates are outdated by the
same amount of time as the scheduling delay. The constella- and dk,n by
tion size is thus not dependent on γτ , and the time delay in 1 + n 3(1 + n − ρn)
this case does not affect the MASSE. However, now the BER dk,n = + . (28)
γ 2 Mk − 1
will suffer from degradation because of the delay. It is shown
in [12] that the average BER, conditioned on γ, is Note that for zero delay (ρ = 1), ck,n = dk,n = ak,n , and (26)
reduces to (15), as expected.
0.2γ
BER(γ) = · e−ρK0 γ/(γ+γ(1−ρ)K0 ) . (23) Because we are interested in the average BER only for the
γ + γ(1 − ρ)K0 CNRs for which we have transmission, the average BER for
Average BER degradation due to time delay for BER0 = 10 3

ACKNOWLEDGMENTS
10 2
The work of Vegard Hassel and Geir E. Øien was supported in

part by the EU Network of Excellence NEWCOM and by the
Average bit error rate (BER)
1 user NTNU Project CUBAN (http://www.iet.ntnu.no/projects/

cuban). The work of Mohamed-Slim Alouini was in part
supported by the Center for Transportation Studies (CTS),
3 Minneapolis, USA.
10
10 users
REFERENCES
[1] P. Viswanath, D. N. C. Tse, and R. Laroia, “Opportunistic
beamforming using dumb antennas,” IEEE Transactions on In-
formation Theory, vol. 48, no. 6, pp. 1277–1294, 2002.
10 4 [2] R. Knopp and P. A. Humblet, “Information capacity and
10 3 10 2 10 1 power control in single-cell multiuser communications,” in
Normalized time delay Proceedings of IEEE International Conference on Communica-
tions (ICC ’95), vol. 1, pp. 331–335, Seattle, Wash, USA, June
Adaptive continuous-rate, constant-power M-QAM 1995.
Adaptive discrete-rate, constant-power M-QAM [3] M. Andrews, K. Kumaran, K. Ramanan, A. Stolyar, P. Whit-
Adaptive continuous-rate, continuous-power M-QAM ing, and R. Vijayakumar, “Providing quality of service over a
Adaptive discrete-rate, continuous-power M-QAM shared wireless link,” IEEE Communications Magazine, vol. 39,
no. 2, pp. 150–153, 2001.
Figure 3: Average BER degradation due to time delay for M-QAM [4] P. Bender, P. Black, M. Grob, R. Padovani, N. Sindhushayana,
rate adaptation with γ=15 dB, 5 fading regions, and BER0 = 10−3 . and S. Viterbi, “CDMA/HDR: a bandwidth-efficient high-
speed wireless data service for nomadic users,” IEEE Commu-
nications Magazine, vol. 38, no. 7, pp. 70–77, 2000.
[5] D. Gesbert and M.-S. Alouini, “How much feedback is multi-
continuous-power, continuous-rate M-QAM is user diversity really worth?” in Proceedings of IEEE Interna-
tional Conference on Communications (ICC ’04), vol. 1, pp.
∞ 234–238, Paris, France, June 2004.
γK BER(γ)pγ∗ (γ)dγ
BERacr,pa = ∞ . (29) [6] V. Hassel, M.-S. Alouini, G. E. Øien, and D. Gesbert, “Rate-
γK pγ∗ (γ)dγ optimal multiuser scheduling with reduced feedback load and
analysis of delay effects,” in Proceedings of the 13th European
Correspondingly, the average BER for the continuous-power, Signal Processing Conference (EUSIPCO ’05), Antalya, Turkey,
September 2005.
discrete-rate M-QAM case is given by
[7] K. J. Hole and G. E. Øien, “Spectral efficiency of adaptive
∞ coded modulation in urban microcellular networks,” IEEE
γ0∗ M1 BER(γ)pγ∗ (γ)dγ Transactions on Vehicular Technology, vol. 50, no. 1, pp. 205–
BERadr,pa = ∞ . (30) 222, 2001.
γ0∗ M1 pγ∗ (γ)dγ
[8] I. S. Gradshteyn and I. M. Ryzhik, Table of Integrals, Series, and
Products, Academic Press, San Diego, Calif, USA, 6th edition,
Figure 3 shows how outdated channel estimates affect the 2000.
average BER for 1 and 10 users. We see that the average sys- [9] A. J. Goldsmith and P. P. Varaiya, “Capacity of fading channels
tem BER is satisfactory as long as the normalized time de- with channel side information,” IEEE Transactions on Informa-
lay again is below the critical value 10−2 for the adaptation tion Theory, vol. 43, no. 6, pp. 1986–1992, 1997.
schemes using continuous power and/or continuous rate. [10] M.-S. Alouini and A. J. Goldsmith, “Capacity of Rayleigh
The constant-power, discrete-rate adaptation policy is more fading channels under different adaptive transmission and
robust with regard to time delay. diversity-combining techniques,” IEEE Transactions on Vehic-
ular Technology, vol. 48, no. 4, pp. 1165–1181, 1999.
[11] A. J. Goldsmith and S.-G. Chua, “Variable-rate variable-power
7. CONCLUSION MQAM for fading channels,” IEEE Transactions on Communi-
cations, vol. 45, no. 10, pp. 1218–1230, 1997.
We have analyzed a scheduling algorithm that has optimal [12] M.-S. Alouini and A. J. Goldsmith, “Adaptive modulation over
spectral efficiency and reduced feedback compared with full Nakagami fading channels,” Kluwer Journal on Wireless Com-
munications, vol. 13, pp. 119–143, 2000.
feedback load. We obtain a closed-form expression for the
[13] D. L. Goeckel, “Adaptive coding for time-varying channels us-
CNR threshold that minimizes the feedback load for this al- ing outdated fading estimates algorithms,” IEEE Transactions
gorithm. Both the impact of scheduling delay and outdated on Communications, pp. 844–855, 1999.
channel estimates are analytically and numerically described. [14] S. Bolis, E. G. Economou, and P. G. Philokyprou, “Scheduling
For both delay scenarios plots show that the system will be delay protocols integrating voice and data on a bus lan,” IEE
able to operate satisfactorily with regard to BER when the Proceedings I: Communications, Speech and Vision, vol. 139, pp.
normalized time delays are below certain critical values. 402–412, 1992.
[15] H.-H. Chen and W.-T. Tea, “Hierarchy schedule sensing pro- David Gesbert is a Professor at Eurécom In-
tocol for CDMA wireless networksperformance study under stitute, France. He obtained the Ph.D. de-
multipath, multiuser interference, and collision-capture ef- gree from Ecole Nationale Supérieure des
fect,” IEEE Transactions on Mobile Computing, vol. 4, pp. 178– Télécommunications, in 1997. From 1993
188, 2005. to 1997, he was with France Telecom Re-
[16] K.-W. Hung and T.-S. Yum, “Fair and efficient transmission search, Paris. From April 1997 to October
scheduling in multihop packet radio networks,” in Proceedings 1998, he has been a Research Fellow at the
of IEEE Global Telecommunications Conference (GLOBECOM Information Systems Laboratory, Stanford
’92), vol. 1, pp. 6–10, Orlando, Fla, USA, December 1992. University. He took part in the founding
[17] J. H. Barnard and C. K. Pauw, “Probability of error for selec- team of Iospan Wireless Inc., San Jose, Calif,
tion diversity as a function of dwell time,” IEEE Transactions a startup company pioneering MIMO-OFDM. Starting in 2001, he
on Communications, vol. 37, pp. 800–803, 1989. has been with the University of Oslo as an Adjunct Professor. He
has published about 100 papers and several patents all in the area
of signal processing and communications. He coedited several spe-
Vegard Hassel is currently pursuing the cial issues for IEEE JSAC (2003), EURASIP JASP (2004), and IEEE
Ph.D. degree at the Department of Electron- Communications Magazine (2006). He is an elected Member of the
ics and Telecommunications at the Norwe- IEEE Signal Processing for Communications Technical Committee.
gian University of Science and Technology He authored or coauthored papers winning the 2004 IEEE Best Tu-
(NTNU). He received the M.S.E.E. degree torial Paper Award (Communications Society) for a 2003 JSAC pa-
from NTNU in 1998 and the M.T.M. degree per on MIMO systems, 2005 Best Paper (Young Author) Award for
from the University of New South Wales Signal Processing Society journals, and the Best Paper Award for the
(UNSW), Sydney, Australia, in 2002. Dur- 2004 ACM MSWiM Workshop. He is coorganizer, with Professor
ing the years 1999–2001 and 2002-2003, he Dirk Slock, of the IEEE Workshop on Signal Processing Advances
was with the Norwegian Defence, working in Wireless Communications, 2006 (Cannes, France).
with video conferencing and mobile emergency networks. His re-
search interests include wireless networks, radio resource manage-
ment, and information theory.
Mohamed-Slim Alouini was born in Tu-

nis, Tunisia. He received the Ph.D. de-
gree in electrical engineering from the Cal-
ifornia Institute of Technology (Caltech),
Pasadena, Calif, USA, in 1998. He was an
Associate Professor with the Department
of Electrical and Computer Engineering of
the University of Minnesota, Minneapolis,
Minn, USA. Since September 2005, he has
been an Associate Professor of electrical en-
gineering with the Texas A&M University at Qatar, Education City,
Doha, Qatar, where his current research interests include the design
and performance analysis of wireless communication systems.
Geir E. Øien was born in Trondheim, Nor-

way, in 1965. He received the M.S.E.E. and
the Ph.D. degrees, both from the Norwe-
gian Institute of Technology (NTH), Trond-
heim, Norway, in 1989 and 1993, respec-
tively. From 1994 to 1996, he was an As-
sociate Professor with Stavanger Univer-
sity College, Stavanger, Norway. In 1996, he
joined the Norwegian University of Science
and Technology (NTNU) where in 2001 he
was promoted to Full Professor. During the academic year 2005-
2006, he has been a Visiting Professor with Eurécom Institute,
Sophia Antipolis, France. His current research interests are within
wireless communications, communication theory, and informa-
tion theory, in particular analysis and optimization of link adap-
tation schemes, radio resource allocation, and cross-layer design.
He has coauthored more than 70 scientific papers in international
fora, and is actively used as a reviewer for several international jour-
nals and conferences. He is a Member of the IEEE Communications
Society and of the Norwegian Signal Processing Society (NORSIG).
DOI 10.1155/WCN/2006/43759
Adaptive Rate-Scheduling with Reactive Delay Control for

Next Generation CDMA Wireless Mobile Systems
Oliver Yu, Emir Saric, and Anfei Li
Department of ECE, University of Illinois at Chicago, 851 S. Morgan Street, 1020 SEO, Chicago, IL 60607, USA
Received 1 October 2005; Revised 11 March 2006; Accepted 26 May 2006

To minimize QoS degradations during nonstationary packet loadings, predictive rate schedulers adapt the operation according to
anticipated packet arrival rates deduced via specified estimation algorithm. Existing predictive rate schedulers are developed under
the assumption of perfect estimation, which may not be possible in future CDMA-based cellular networks characterized with
highly nonstationary and bursty traffic. Additional shortcoming of existing rate schedulers is the coupling of delay and bandwidth,
that is, close interdependence of delay and bandwidth (rate), whereby controlling one is accomplished solely by changing the other.
In order to mitigate for the arrival rate estimation errors and delay-bandwidth coupling, this paper presents the feedback-enhanced
target-tracking weighted fair queuing (FT-WFQ) rate scheduler. It is an adaptive rate scheduler over multiclass CDMA systems with
predictive adaptation control to adapt to nonstationary loadings; and feedback-enhanced reactive adaptation control to counteract
arrival rate estimation errors. When the predictive adaptation control is not able to maintain long-term delay targets, feedback
information will trigger reactive adaptation control. The objective of FT-WFQ scheduler is to minimize deviations from delay
targets subject to maximum throughput utilization. Analytical and simulation results indicate that FT-WFQ is able to substantially
reduce degradations caused by arrival rate estimation errors and to minimize delay degradations during nonstationary loading
conditions.
Copyright © 2006 Oliver Yu et al. This is an open access article distributed under the Creative Commons Attribution License,
1. INTRODUCTION Given limited wireless resources, CAC enables connec-

tion-level QoS guarantees by implementing class-prioritized
Next generation CDMA-based cellular wireless networks admission control. It also enables minimum application-
are slated to provide wide range of integrated multime- level performance guarantees by limiting the total num-
dia services with a guaranteed quality of service (QoS) ber of admitted connections. However, due to the bursty
(e.g., voice, video, high-speed data). This, in turn, will cre- nature of packet traffic (especially from the connec-
ate heterogeneous traffic environment characterized with tions of nonreal-time classes) CAC alone is not adequate
highly nonstationary and bursty transmissions. The uni- to provide optimal resource utilizations and application-
versal mobile telecommunication system (UMTS) is a 3rd level performance. MAC algorithm that includes effi-
generation (3G) mobile communication system developed cient packet scheduler needs to accompany an admission
by 3rd generation partnership project (3GPP). It defines controller. It is responsible for provisioning differenti-
“per-class” QoS provisioning, and classifies all traffic into ated application-level QoS requirements to admitted con-
four QoS classes, namely conversational, streaming, in- nections by providing optimal resource allocations. This
teractive, and background [1]. Each class has its own paper focuses on packet scheduler part of MAC algo-
connection-level (or call-level) QoS requirements in terms rithm that accompanies admission controller proposed
of connection blocking/dropping probabilities, as well as in [2].
application-level QoS requirements in terms of delay, jit- Efficient packet scheduler is crucial for QoS provision-
ter, throughput, BER, and burstiness. QoS provisioning ing in an integrated multiclass packetized network. Some
with performance differentiation in a heterogeneous non- of the desirable properties of a packet scheduler providing
stationary environment requires efficient call admission “per-class” QoS support in a wireless network include ef-
control (CAC) and medium access control (MAC) proto- ficient link utilization with optimal resource distributions,
cols. delay bound guarantees for each class, bit-error-rate (BER)
guarantees, throughput guarantees, delay-bandwidth decou- size); estimation errors would degrade their performance.
pling, and low complexity. Due to nonstationary traffic expected in the future wire-
Many packet scheduling algorithms have been proposed less networks, arrival rate estimation errors are immi-
for CDMA-based wireless networks. The capacity of CDMA nent. Consequently, estimation errors could lead to inef-
systems (especially in uplink) is interference-limited, sub- ficient and erroneous resource distributions (i.e., rate as-
ject to the variation of signal-to-interference ratios (SIRs), signments) whereby over-provisioning of some traffic classes
and bandwidth demands of users with limited power con- might occur even when other classes are not meeting
straints. CDMA system loading factor can be derived to de- QoS targets. Moreover, TVFQ adapts weights (or rates)
note interference-based CDMA resources occupied by trans- based on the future queue size (and predefined priority in-
mitting users. The schemes in [3–6] utilize interference- dices) without any regard to absolute delay targets. In a
based loading and the variants of generalized processor shar- highly nonstationary environment characterized with fre-
ing (GPS) fair scheduling discipline to dynamically allo- quent packet bursts, however, it is possible to have a con-
cate transmission rates and schedule packets in a CDMA- nection with large instantaneous queue size (due to sud-
based system. Specifically, the authors in [3, 4] propose den arrival burst) but whose mean delay is significantly
code-division GPS (CDGPS) scheduling scheme that max- below its delay target. Hence, to utilize resources effi-
imizes throughput by providing “weighted fairness” (i.e., ciently in a nonstationary traffic environment, adaptive
relative provisioning) in terms of the rate and signal-to- rate scheduler needs additional delay target-tracking con-
interference ratio (SIR) guarantees. Similarly the scheme in straints so as to minimize delay deviations from absolute tar-
[5] proposes a rate scheduler with explicit BER guarantees gets.
in a wideband CDMA system. The scheme in [6] is a rate In order to dispense with delay-bandwidth coupling as
scheduler based on the adjusted GPS concept that explic- well as to counteract arrival rate estimation errors and to
itly takes into account current channel conditions. It maxi- achieve efficient resource distributions with absolute de-
mizes total throughput by providing “weighted-fair” rate al- lay target-tracking, this paper proposes feedback-enhanced
locations with BER guarantees. The scheduler in [7] con- target-tracking weighted fair queuing (FT-WFQ) scheduler.
trols transmission power and dynamically allocates transmis- It dynamically adapts transmission rates on a “per-class” ba-
sion rates so as to maximize the number of users whose BER sis such as to minimize overall delay deviations from abso-
is satisfied. To solve such an optimization problem, the au- lute delay targets subject to maximum throughput utiliza-
thors suggest search procedure based on the genetic algo- tion. FT-WFQ utilizes predictive adaptation control based
rithm. One of the drawbacks of the aforementioned schemes on estimated arrival rates, but it also implements concur-
[3–7] is the delay-bandwidth coupling whereby interdepen- rent feedback-enhanced reactive control that detects imper-
dence of delay and bandwidth (e.g., reducing delay im- fections, such as estimation errors, and counteracts them.
plies a larger bandwidth allocation) could lead to resource Feedback control unit monitors average delays of each class
underutilizations. The importance of delay-bandwidth de- and if it detects that a class is degraded (possibly be-
coupling is even more signified in the future multime- cause of estimation errors) it corrects the problem in or-
dia wireless networks supporting traffic with similar delay der to achieve efficient resource distributions and mini-
but considerably different bandwidth requirements or vice mize overall delay deviations from corresponding delay tar-
versa. gets.
Schedulers in [8, 9] dispense with delay-bandwidth cou- This paper is organized as follows. In Section 2, system
pling in a time-division-duplex (TDD) CDMA system by uti- model as well as problem statements are described. Then, in
lizing packet-prioritization that arrange transmissions in or- Section 3, CDMA “bandwidth” in terms of the interference-
der to explicitly reduce packet delays. Similarly, authors in based loading is derived. Also, maximum loading-capacity is
[10, 11] propose token bank fair queuing (TBFQ) schedul- computed. The proposed FT-WFQ scheduler is thoroughly
ing algorithm that provides soft QoS guarantees. TBFQ keeps presented in Section 4. In Section 5 analysis and simulation
track of previous transmissions and introduces a priority in- models for performance evaluation are presented. Section 6
dex that determines which connections can utilize excess re- displays numerical results and comparison. Finally, Section 7
sources. concludes the paper.
Time varying fair queuing (TVFQ) scheme in [12] is
motivated by the delay-bandwidth decoupling problem. It
extends dynamic (weighted) fair queuing concept into mul- 2. SYSTEM MODEL AND PROBLEM
ticode (MC) CDMA systems. TVFQ decouples delay and STATEMENTS
bandwidth by solving a nonlinear integer programming
problem that explicitly minimizes queuing delays and pro- 2.1. System model
duces optimal weight (rate) assignments on a time-varying
basis. The authors present computationally efficient solu- Uplink scheduling in a single cell of a wireless cellular sys-
tion method based on dynamic programming. However, tem that uses CDMA is considered. The cell contains mo-
the problem with TVFQ algorithm as well as the adaptive bile users requesting packet transmission (i.e., seeking ac-
rate schedulers in [3, 4] is that they rely upon the per- cess to CDMA resources) and the base station (BS) which
fect estimations of the future traffic arrival rates (or queue centrally implements scheduling algorithm and optimally
Oliver Yu et al. 3
allocates resources on a dynamic basis (i.e., every schedul- Conversational class

ing time interval). Transmission is packetized with fixed-
λ1 R1
length packets and time is divided into frames of equal
Streaming class
length T f (e.g., T f = 10 ms in UMTS). A class-based sys- λ2 R2
tem is assumed where packets of each user belong to one Requests FT-WFQ
of N traffic-classes (e.g., N = 4 in UMTS). Packets of Classifier λ3 Interactive class R3 scheduler
(assign Ri )
each class have a distinct time-out value (i.e., delay require-
λ4 R4
ment denoted by a delay target) measured in frames and Background class
if not transmitted by this time, they are useless. It is as-
sumed that each mobile user has a large enough buffer, so
that packets are lost only if not scheduled and transmitted
on time (time-out expiration), and not due to buffer over- Figure 1: Packet scheduler at base station (with N = 4 traffic
flow. classes).
Let the maximum uplink capacity of CDMA system
(i.e., resource capacity measured in terms of interference-
based loading) in the nth time interval be denoted by
ηT [n]. The maximum capacity in terms of CDMA load- the next interval. The notification is through a downlink
ing, subject to BER constraints, is analytically derived in the broadcast control channel. After listening to broadcast con-
next section (Section 3). In each scheduling time interval trol channel, mobile users, which are granted permission to
n, the job of a packet scheduler is to optimally allocate the transmit, forward their packets to BS on uplink dedicated
available capacity among active (i.e., transmitting) mobile channels in the corresponding frame of time interval n + 1.
users. The whole process is repeated every scheduling time inter-
It is implied that the admission control has been con- val.
ducted previously, and only users that are admitted into the
system can send packet transmission requests. Admission
control is such that each traffic class i (i = 1, 2, . . . , N) is 2.2. Problem statements
guaranteed minimum allocation rate Ri,min (packets/frame).
Packet scheduling is performed to further exploit bursty na- 2.2.1. Packet arrival rate estimation errors and
ture of user traffic. efficient resource distributions
Consequently, whenever active admitted users have
packets ready for transmission in the next time interval, Adaptation of the existing dynamic scheduling schemes
they send packet transmission requests (i.e., small signal- such as time varying fair queuing (TVFQ) [12] is highly
ing packets) to the base station (BS) in the current time sensitive on the real-time estimation of future packet ar-
interval on special uplink random access request chan- rival rates (or some other measure of future traffic). Traf-
nels. It is assumed in this paper that some efficient ran- fic in the future wireless networks is, however, expected
dom access technique is employed and that random ac- to be highly nonstationary. Due to small cell size and in-
cess delay is negligible. Users seeking medium access increased handoff rates even traffic of real-time classes (con-
dicate the number of packets ready for transmission in versational and streaming) observed at BS is expected to
the next time interval, as well as traffic class of each fluctuate and be nonstationary. The performance of the
packet. adaptive scheduler degrades in the presence of arrival
BS collects all transmission requests (i.e., small sig- rate estimation errors inherent in nonstationary environ-
naling packets) for the following time interval. It first ment.
classifies all requests according to their traffic class and Estimation errors could lead to inefficient resource (i.e.,
then places classified packet requests (one for each packet rate) distributions and unequal delay deviations from the
requested) into N traffic queues on a first-come-first- targets. For instance, classes whose arrival rate is over-
served basis (see Figure 1). Note that besides new packet estimated will (erroneously) allocate more resources than
requests each queue may also contain unexpired back- needed to keep their delays at the corresponding targets.
logged packet requests that were invoked in previous This, in turn, will capture resources from other classes
intervals but were not accommodated for transmission whose delay as a result might rise above targets. Con-
yet. sequently, this could lead to a situation where for some
At the end of the scheduling time interval BS performs classes, large negative deviations from delay targets could
the proposed adaptive scheduling algorithm as explained in be present even when positive delay deviations are ob-
next sections. The algorithm returns the optimal rate allo- served for other classes. Ideally, however, there should not
cations (in packets/frame) for each traffic-queue that would be any negative deviations when positive ones are ob-
minimize delay cost function, namely R∗i , i = 1, 2, . . . , N. served.
Based on these, BS notifies the owners (i.e., users) of the Authors in [13] suggest three prediction techniques for
R∗i head-of-line packet requests in the traffic-queue i (i = estimating packet arrival rate λi [n] of class i at the current
1, 2, . . . , N) that they are granted permission to transmit in time interval n. First one is to use the arrival rate observed
at the previous time interval. The second one is to use av- of a user of class i, i = 1, 2, . . . , N (observed at BS) is given as
erage arrival rate based on observed history (i.e., λi [n] =
n−1 Eb Si
j =1 λi [ j]/(n − 1)). The third method they suggest is based = G p,i · , (1)
on moving average of the first two methods. From these es- I0 i Itotal − Si
timation techniques it is evident that they are prone to er-
rors in a highly nonstationary environment subject to sud- where Si is the received signal power of a user of class i, and
den bursts of packet arrivals. Itotal is the total received wideband power including thermal
noise power PN in the BS. Assume a perfect power control
Furthermore, TVFQ scheduler does not consider abso-
such that the received power levels Si of all users belonging
lute delay targets when dynamically adapting weights (or
to the same class i are equal. Let γi be the minimum value
rates). In a highly nonstationary environment even under
of (Eb /I0 )i required for acceptable BER (for a user of class
perfect traffic estimations it is possible to have connections
i). Therefore, for satisfactory BER, the following constraints
whose traffic queue size is large due to sudden traffic bursts
need to be satisfied (∀i):
but whose mean delay is significantly below corresponding
delay target. Thus, for efficient rate adaptations, target track-
Eb Si
ing constraints that minimize delay deviations need to be in- = G p,i · ≥ γi . (2)
I0 i Itotal − Si
corporated.
It can be shown that the received power levels are mini-
2.2.2. Delay-bandwidth coupling mized when the above equation is satisfied with equality. Let
S∗i be the received power level of a user of class i such that the
One of the major shortcomings of dynamic rate scheduler, above equation is satisfied with equality. Thus,
such as the ones based on GPS, is the coupling of delay
1
and bandwidth. It refers to close interdependence of delay S∗i = · Itotal . (3)
and rate (i.e., bandwidth) parameters, whereby provision- 1 + G p,i /γi
ing one parameter (e.g., delay) can only be accomplished
by changing the other (e.g., rate). For instance, in GPS, the Note, however, that the received power level Si is
delay of a class-queue is controlled by changing its allo- bounded by the maximum value Si,max which is dependent on
cated rate (i.e., bandwidth). Since delay and bandwidth can- (mobile) transmit power, and achieving feasible S∗i ≤ Si,max
not be modified independently, the BS scheduler would al- is a requirement that limits maximum interference Itotal that
locate high rate to a class-queue with low delay requirement a system is able to tolerate, as elaborated in the next sub-
even if this class has low bandwidth requirement. This would section. Let the load factor increment Δηi of a user of class
lead to high bandwidth underutilizations. Delay-bandwidth i be defined as Δηi ≡ S∗i /Itotal .Therefore,
coupling problem is even more signified in a future multi-
1
class environment where classes with similar delay require- Δηi = . (4)
ments might have significantly different bandwidth require- 1 + G p,i /γi
ments (e.g., voice and video). In order to utilize resources
efficiently, a dynamic scheduler needs to decouple delay and Assuming Ni users of class i are in the system, Itotal is given as
bandwidth such that both parameters can be guaranteed in-
dependently.
N
Itotal = Ni · S∗i + PN . (5)
i=1
3. LOADING AND MAXIMUM LOADING

Using terminology of the last section, note that the bit
CAPACITY IN CDMA SYSTEM rate of the “class” i is given as Ri = Ni · ri . Let noise rise NR be
defined as the ratio of total received wideband noise power in
This section presents the concept of loading as an integrated BS to the thermal noise power (NR = Itotal /PN ). Substituting
measure of resource-usage in a multiclass CDMA system. into the above formulas,
The maximum possible loading capacity subject to BER con-
straints is also derived. These results are used by the dynamic Itotal 1 1
resource monitor of the proposed scheduler as explained in NR = = N = , (6)
PN 1 − i=1 Ni · Δηi 1 − η
detail in Section 4.
where η (η ≥ 0) is defined as loading:
3.1. CDMA interference-based loading

N
PN
η= Ni · Δηi = 1 − ≤ ηT . (7)
Let G p,i be the processing gain (or the spreading factor) of i=1
Itotal
a user that belongs to traffic-class i (i = 1, . . . , N), defined
as G p,i = W/ri , where W is the system bandwidth in Hz (or The loading represents the amount of resources used in a
chip rate), and ri is the bit rate of a user of traffic class i. The CDMA system (when corresponding bit rates are allocated),
signal energy per bit to noise-plus-interference ratio (Eb /I0 )i and it defines the so-called “CDMA bandwidth.”
Oliver Yu et al. 5
3.2. Maximum loading capacity 4.1. FT-WFQ architecture
Theoretically, the maximum loading, denoted as ηT , is 1. The unifying architecture that applies to both versions
In reality, however, ηT is limited by utmost interference (or (heuristic and optimal) of feedback-enhanced target-track-
loading) a system is able to tolerate (given BER and limited ing weighted fair queuing (FT-WFQ) scheduler is shown
power Si,max constraints). From the above (5) and (7), the to- in Figure 2. The scheduler consists of feedback-enhanced
tal interference Itotal can be expressed in terms of loading η as scheduling unit (F-SU) fed and controlled by arrival rate
Itotal = PN /(1 − η). Then, the BER constraints of (2) become estimator block (AE), feedback control unit (FCU) and dy-
Si,max namic resource monitor (DRM). F-SU defines an optimiza-
G p,i · ≥ γi , ∀i. (8) tion problem that optimally allocates transmission rates
PN /(1 − η) − Si,max
every scheduling time interval. The optimization problem
Equivalently, within F-SU is shaped by the information provided by AE,
PN G p,i · Si,max FCU, and DRM, and its objective is to minimize delay cost
≤ + Si,max , ∀i (9) function as defined in the next subsections. AE block pro-
1−η γi
vides estimated arrival rates for the following time inter-
or, in terms of loading η, val, while FCU monitors average delay incurred by each
class, and adjusts optimization problem within F-SU if de-
PN
η ≤ 1−
< 1, ∀i. (10) lays exceed pre-defined targets (i.e., it provides a corrective
G p,i · Si,max /γi + Si,max feedback). The feedback adjustment (as well as optimiza-
Therefore, loading bound, or the maximum loading ηT tion problem within F-SU) is heuristic or optimal depend-
tolerated by a system is given as ing on the version of scheduler and as elaborated in the fol-
lowing subsections. DRM on the other hand dynamically
PN recalculates total resources (i.e., CDMA capacity) available
ηT = min 1 −
. (11)
∀i G p,i · Si,max /γi + Si,max and checks if scheduling assignment is feasible by adding
(cross-layer) resource constraint in the optimization prob-
lem.
4. FEEDBACK-ENHANCED TARGET-TRACKING
WEIGHTED FAIR QUEUING (FT-WFQ)
4.2. Heuristic-based scheme
Two versions of FT-WFQ rate scheduling scheme are pro-
posed, namely, heuristic and optimal. The proposed scheme Let λi [n] be the estimated arrival rate of class i (i =
is characterized with the following features. 1, 2, . . . , N) for the nth scheduling time interval measured in
packets per frame (note that the actual estimation method
(i) It supports a multiclass prioritized adaptive rate is not considered in this paper). It is provided by the ar-
scheduling with “per-class” QoS support including rival rate estimator block (AE) (Figure 2). Also, let Qi [n]
guaranteed rate, delay, and BER. To maintain QoS be the queue size (in packets) of class i at the beginning
guarantees the proposed scheme adapts to changing of the nth (scheduling) time interval. Note that Qi [n] is
traffic conditions by employing predictive adaptation known to the BS scheduler as it represents the current packet
based on estimation of future packet arrival rates as backlog. Considering the nth time interval in isolation, the
well as feedback-enhanced reactive adaptation control. scheduling delay (in frames) of class i packet-queue is given
(ii) It exploits feedback-enhanced reactive control in or- by
der to maintain delay targets (target tracking) and to
counteract arrival rate estimation errors. When pre- Qi [n] + λi [n] · T
dictive adaptation fails to maintain delay targets (due Di [n] = , (12)
Ri [n]
to arrival rate estimation errors or high congestions)
feedback information is utilized to correct rate alloca- where Ri [n] is the allocated rate (in packets/frame) to class i
tions. Feedback control ensures that deviations from packet-queue in the nth time interval and T is the schedul-
delay targets are minimized by efficient allocation of ing time interval duration measured in frames (T = 1 if
resources during failure condition. scheduling is done on a frame-by-frame basis). The objec-
(iii) It decouples delay and bandwidth (i.e., rate) param- tive of the (heuristic) F-SU in the nth scheduling time in-
eters. Maintaining delay targets and rate allocation terval is to allocate rates Ri [n] (i = 1, 2, . . . , N) such as
are accomplished through a separate control. Total to minimize overall delay cost function Ni=1 Di [n], while
scheduling delay is explicitly minimized while rate keeping mean delay of all classes as close as possible to
guarantees are still met. their respective delay targets. Note, however, that the de-
(iv) It utilizes cross-layered design, whereby dynamic relay cost function defined above is highly dependent on the
source monitor ensures that allocated rates are feasible estimated arrival rates λi [n]. Even slight estimation errors
in the sense that BER is satisfied for all transmitting by AE block could degrade performance, and lead to er-
users. Interference-based loading is used to denote re- roneous rate assignments with inefficient resource distribu-
source usage in a CDMA system. tions.
Assign rates Ri [n]
Class 1
R1 [n]
Feedback
control unit
(FCU)
Class 2
R2 [n] Delay
Feedback-enhanced
scheduling unit monitoring
F-SU Di
R3 [n]
Class 3
R4 [n] Feedback
adjustment
Class 4
(heuristic or
optimal)
Arrival Dynamic
rate resource
estimator monitor Reactive control
(AE) (DRM)
Predictive control
Figure 2: Architecture of FT-WFQ scheduler (with four traffic classes).
In order to mitigate for the estimation error, as well as interval n is formulated (for clarity of presentation index n is
to meet mean delay objectives as efficiently as possible, the dropped) as follows.
following heuristic-based feedback control unit (FCU) that Find the optimal rate allocations R∗i , i = 1, 2, . . . , N, so as
initiated adjustment of the optimization problem in F-SU to
is proposed. Let Td,i denote mean delay target for pack-

ets of class i (measured in frames). It is an operator spe-
N
pi · Qi + λi · T
minimize (13)
cific value based on the level of QoS guarantee provided. i=1
Ri
The FCU monitors mean packet scheduling delays of each
class. Let the running average of monitored packet delay of subject to
class i at the time interval n be denoted as Di [n]. Start-

ing from the highest priority class (class 1) with descend- Ri ≥ min Ri,min , Qi + λi · T /Td,i / P,
∀i ∈ (14a)
ing priority, FCU finds class i (if any) whose delay Di [n] is

above targeted threshold Td,i (i.e., Di [n] > Td,i ). This sig- Ri = min Ri,min , Qi + λi · T /Td,i ∀i ∈ P , (14b)
nals that the estimation error occurred (with high proba-
N
1
bility) and that class i was degraded due to wrong assign-
≤ ηT , (14c)
i=1
1 + W/R i /γi
ments. FCU then “preempts” all classes j = i whose mean
delay D j [n] is below corresponding targeted threshold (i.e., Ri ≥ 0 i = 1, 2, . . . , N. (14d)
all classes j for which D j [n] < Td, j ). A “preempted” class
is constrained to minimum guaranteed rate and it is pre- The term (Qi + λi ) · T/Td,i , appearing in constraints of
vented from sharing excess resources (in that time inter- (14a), (14b), represents the rate needed to keep class i de-
val). “Preemption” is conducted by sending feedback infor- lay below its delay target Td,i . However, in order to make
mation that changes corresponding constraints in optimiza- the solution feasible in the case of unpredicted bursts, each
tion problem within (heuristic) F-SU in the nth interval. class is only guaranteed service rate Ri,min , which is the
This ensures that class j receives only minimum guaranteed minimum rate for class i guaranteed by admission control
service rate until delay of class i has stabilized. The pseu- (see min(·) term in (14a) and (14b)). Note that if class i
docode of FCU-initiated heuristic adjustment is shown in is preempted by heuristic FCU the inequality constraint in
Figure 3. (14a) is changed to the corresponding equality constraint
Let the set of preempted classes (in the nth time interval) in (14b). The constraint in (14c) is due to DRM. It en-
be denoted by P . Let ηT [n] denote the total capacity avail- sures that the rate allocation is feasible in the sense that
able as evaluated by dynamic resource monitor (DRM), and BER is satisfied for all transmitting users. DRM constraint
let constant pi indicate different priorities in the system, such in (14c) follows from Section 3 with class i rate given as
that if class i has higher priority than class j, then pi > p j . Ri = Ni · ri and with the maximum loading ηT given by
Then, the optimization problem of (heuristic) F-SU in the (11).
Oliver Yu et al. 7
Heuristic FCU in interval n:

subject to

(1) i = 1 Ri ≥ min Ri,min , Qi + λi · T/Td,i ∀i, (17a)
(2) if (Di [n] > Tdi ) {
(3) Preempt Classes j for which D j [n] < Tdj
N
1
≤ ηT , (17b)
(4) → Set R j [n] = R j,min 1 + (W/Ri )/γi
i=1
(5) DONE
(6) } Ri ≥ 0 i = 1, 2, . . . , N. (17c)
(7) else {
Note that the proposed optimization problem will mini-
(8) i=i+1 mize total deviations from delay targets if FCU detects that
(9) GO TO 2 mean delay of any class exceeds corresponding delay tar-
(10) } get (i.e., if the indicator function ID [n] is set to 1), oth-
erwise it will minimize the total delay (i.e., if the indica-
Figure 3: Pseudocode of FCU-initiated heuristic. tor function ID [n] is set to 0). The reasoning behind this
is that if the mean delay of all classes is below their re-
spective delay targets, then the objective is to minimize
4.3. Optimal scheme the overall delay, whereas if delay of any class is above its
corresponding delay target, the resources should be redis-
The objective of the optimal scheduling scheme (i.e., F-SU) tributed so as to keep delay of all classes as close to their
is to minimize the overall delay and in the case of arrival rate delay targets as possible. In other words if there is any
estimation errors or high loading congestions to minimize class whose mean delay is above its corresponding delay tar-
mean delay deviations from the corresponding targeted ob- get, there should be no classes whose mean delay is below
jectives. It is “optimal” in the sense that it explicitly mini- theirs.
mizes delay deviations from targeted objectives and as such The constraints in (17a) and (17b) are analogous to the
allocates resources as efficiently as possible. It is, however, not corresponding constraints in a heuristic-based problem of
overall optimal as it only considers single time interval in iso- the last subsection with constraint in (17b) due to DRM.
lation, whereas the overall optimal scheme would consider a
larger time horizon.
The optimization problem is defined as follows. Let the 5. PERFORMANCE ANALYSIS AND
indicator function ID [n] in the nth time interval be defined as SIMULATION MODELS
⎧
⎪
⎨0, if Di [n] ≤ Td,i ∀i = 1, 2, . . . , N, In this section, the analysis and simulation models are devel-
ID [n] = ⎪ (15) oped for the proposed FT-WFQ scheduler in nonstationary
⎩1, otherwise, traffic environment. Four traffic classes (i.e., N = 4) defined
in UMTS network are considered (see Table 1). Performance
where as in the last subsection Di [n] denotes the mean (FCU) measures are mean delay and service rate assigned to each
monitored scheduling delay of class i at the time interval class.
n, and Td,i is the mean delay target for packets of class i as
measured in frames. Therefore, the binary indicator function
ID [n] is set to 1 if mean delay of any class exceeds its delay tar- 5.1. Delay analysis with nonstationary packet
get Td,i . This signals that resources were assigned erroneously arrival rate and estimation error
either due to arrival rate estimation errors or due to very
high congestion. The indicator function is set by the (opti- Assume that packets of class i (i = 1, 2, 3, 4) arrive ac-
mal) feedback control unit (FCU) (recall that FCU explicitly cording to a nonstationary Poisson arrival process with
monitors mean packet scheduling delays Di [n] of each class mean arrival rate of λi (t) (packets/frame). Nonstationary
i). Using the same terminology as in the last subsection, the Poisson arrival process is characterized by time-varying
optimization problem of the optimal F-SU in the interval n mean arrival rate λi (t) modeled as follows. Time is di-
is formulated (for clarity of presentation index n is dropped) vided into equal length (scheduling) time intervals of du-
as follows. ration T frames. In the nth interval (n = 1, 2, 3, . . .)
Find the optimal rate allocations R∗i , i = 1, 2, . . . , N, so mean arrival rate λi (n · t) (denoted as λi [n]) takes a ran-
as to dom value according to a uniform distribution. It re-
tains this value for the duration of interval n. Without

N
pi · Qi + λi · T any loss of generality, the four aforementioned nonstation-
minimize 1 − ID · ary Poisson arrival processes are assumed to be indepen-
i=1
Ri
dent.

2 (16)
Let λi [n] as defined above be the actual mean arrival rate

N
Qi + λi · T
+ ID · pi · − Td,i of class i arrival process for the nth time interval. Let λi [n] be
i=1
Ri the estimated arrival rate that is observed at BS and used by
Table 1: Numerical values of QoS parameters for each class.
Delay tolerance (frames) or QoS requirements

Traffic class i Traffic type (UMTS QoS class)
packet timeout value Tdi Minimum rate Ri,min BER requirement
1 Conversational <1 12 kbps 10−3
2 Streaming 1–2 128 kps 10−4
3 Interactive 2–4 32 kbps 10−5
4 Background >8 0 10−7
the scheduling algorithm (in the nth interval). As discussed In the accordance with the proposed scheduler, FCU
previously the estimator is not perfect, and consequently it monitors PSA running average delay of each class i (20) and
is assumed that an additive white Gaussian error εn is intro- adjusts optimization problem in the nth time interval ac-
duced in each interval n, that is, cordingly, as explained in Section 4. Hence, the optimization
problem is solved in each time interval n as given in (13) for
λi [n] = λi [n] + εn . (18) the heuristic-based or (16) for the optimal scheduler. MAT-
As noted above, εn is a white Gaussian random process with LAB (optimization toolbox) was utilized to solve the actual
mean ε, and variance 0.1 · ε, for all n. Also, E[εn · εk ] = 0 for optimization problem in the nth time interval.
all n = k (E[·] is the expectation operator).
In the nth time interval, class i (i = 1, 2, 3, 4) packet- 5.2. Simulation model with nonstationary packet
queue receives service rate R∗i [n] (packets/frame) in accor- arrival rate and estimation error
dance with the solution of the optimization problem de-
fined in (13) for the heuristic-based scheduler or (16) for Proposed scheduling scheme was simulated in a nonstation-
the optimal scheduler. Hence, each class i queue can be con- ary environment using an event-driven simulation tool OP-
sidered in isolation with time-varying arrival rate λi [n] and NET [17]. The model consists of four traffic generators (one
time-varying service rate R∗i [n]. Such a queue can be rep- for each class), and the base station (BS) where the schedul-
resented by an M[n]/D[n]/1 system, where M[n] represents ing algorithm is implemented. As in the analysis section, time
the nonstationary packet arrival process as defined above is divided into equal-duration intervals n of length T frames.
and D[n] stands for deterministic server operating at opti- Traffic generators generate traffic according to four indepen-
mal rates of R∗i [n]. Because of its time-varying nature, it is dent nonstationary Poisson processes as in the last subsec-
very difficult to analyze M[n]/D[n]/1 system directly (i.e., tion. As in the analysis, it is assumed that the additive white
to solve Kolmogorov forward equations). However, various Gaussian estimation error εn is present when estimating the
approximations have been proposed in the literature. One actual arrival rate. The mean ε of the estimation error was
very simple approximation is called point-wise stationary used as a simulation parameter.
approximation (PSA) also known as quasistationary approx- In order to solve the optimization problem in (13) or
imation [14, 15]. According to PSA, in each time interval (16) using optimization toolbox provided by MATLAB, a co-
n, M[n]/D[n]/1 system can be approximated by a station- simulation interface model of OPNET and MATLAB was de-
ary M/D/1 model where the current value of λi [n] is used veloped. The “mx” interface provided by MATLAB was used,
as “stationary” arrival rate and the current value of R∗i [n] as explained in detail in [18]. (This is very useful if one needs
is used as deterministic service rate in that particular in- to use MATLAB algorithms when simulating complex com-
terval. For PSA approximation to be valid, duration of the munications systems with discrete event simulator.) The run-
time interval (T) should be 4–5 times greater than the packet ning average delay statistic was collected for each class dur-
service time, so that the system can asymptotically reach ing simulation run-time. In accordance with the proposed
a steady-state. Consequently, in analytical approximation T scheduling scheme, this information was used by feedback
frames (for some large enough T) constitute one time inter- control unit (FCU) to adjust optimization problem in each
val n. time interval.
Assuming M/D/1 model in each time-interval n (n =
1, 2, . . .), instantaneous PSA delay for class i (denoted as 6. ANALYSIS AND SIMULATION
Di [n]) is given by Pollaczek-Khinchin delay formula [16]: NUMERICAL RESULTS

2
λi [n]/ R∗i [n] 1 The numerical parameters used in the analysis as well as
Di [n] =
+ ∗ . (19)
∗
2 1 − λi [n]/Ri [n] Ri [n] in simulations are summarized in Table 2. The proposed
heuristic-based and optimal FT-WFQ scheduling schemes
Then, PSA running average delay of class i used by feedback
are evaluated in nonstationary packet arrival environment
control unit (FCU) is defined as
with and without the presence of arrival rate estima-
Di [n − 1] · (n − 1) + Di [n] tion error. The proposed scheduling schemes are compared
Di [n] = . (20) to the TVFQ scheme without reactive control as originally
n
Oliver Yu et al. 9
1.4 2
1.8
1.2
1.6
1 1.4
Delay (frames)
Delay (frames)
1.2
0.8
1
0.6
0.8
0.4 0.6
0.4
0.2
0.2
0 0
0 80 160 240 320 400 0 80 160 240 320 400
Time Time
FT-WFQ (optimal) TVFQ FT-WFQ (optimal) TVFQ

FT-WFQ (heuristic) Target Td1 FT-WFQ (heuristic) Target Td2
(a) (b)
5 10
4.5 9
4 8
3.5 7
Delay (frames)
Delay (frames)
3 6
2.5 5
2 4
1.5 3
1 2
0.5 1
0 0
0 80 160 240 320 400 0 80 160 240 320 400
Time Time

(c) (d)
Figure 4: Analysis: delay for classes 1–4 (no estimation error).
proposed in [12]. Consistent with the last section, four traf- with estimation error εn = 0). Priority weights pi are selected
fic classes are considered. such that under average stationary (arrival) conditions de-
lay targets are met; the numerical values are listed in Table 2.
6.1. No arrival rate estimation error The running average of delay (measured in frames) versus
simulation time (i.e., time instant) for each class are obtained
Due to nonstationary traffic conditions, even under perfor the compared schemes following analysis and simulations
fect traffic estimations, the selection of priority weights pi models presented in the last section. Note that, as mentioned
needed to maintain delay targets becomes a difficult task for before, TVFQ scheme is the one without reactive adaptation
the TVFQ scheduler. In this subsection, the performance of control. Delay results for the compared schemes, obtained
TVFQ scheme [12] is compared to the proposed scheduling from an analytical model, are shown in Figure 4 for classes
schemes under such conditions (i.e., nonstationary arrivals 1–4, respectively. The results are further compared by the
Table 2: Summary of analysis and simulation parameters.
Parameter Value
Class 1 arrival process Poisson, uniform range: 0 ≤ λ1 ≤ 0.4 · ηT
Delay targets Tdi (in frames) Td1 = 0.9 Td2 = 1.7 Td3 = 3.5 Td4 = 8
Interval size T (in frames) 20
Maximum power Si,max 100 (mW)
Mean estimation error ε −0.5 (also varied)
Priority weight pi p1 = 6 p2 = 4 p3 = 2 p4 = 1
Normalized percent Total deviation even when large negative delay deviations for classes 3 and 4
change from target Dtotal are present (see left part of Figure 5). Thus it wastes resources
20
by over-feeding classes 3 and 4 during their light packet ar-
rivals, when these excess resources could have been allocated
15 to classes 1 and 2, respectively. From Figure 4 and bar chart in
Figure 5, it can be seen that the proposed FT-WFQ schemes
Percentage (%)
10 (heuristic and optimal) achieve far better resource distribu-

tions and that the total delay deviations from the targets are
minimized. By utilizing feedback-enhanced reactive control
5
designed to explicitly minimize delay deviations from the
corresponding delay targets, the heuristic-based and optimal
0 FT-WFQ schemes slightly increase the mean delay of classes
3 and 4, respectively by reducing resources (i.e., rates) allo-
cated to them, but nevertheless keeps them close to their re-
5
1 2 3 4 5 spective targets. As evident from Figure 4 and the bar chart
Class on the left of Figure 5, this in turn provides more resources
to accommodate heavy traffic arrival from classes 1 and 2, re-
Optimal FT-WFQ spectively, thereby reducing their mean delay deviations from
Heuristic FT-WFQ the targets. It can also be seen that the optimal FT-WFQ
TVFQ
scheme achieves better resource allocations than heuristic-
based scheme as its objective is to explicitly minimize delay
Figure 5: Analysis: delay percent change from target (no estimation deviations.
error). Total performance gain/loss is quantified as follows.
From the bar charts on the left in Figure 5, the total de-
bar charts shown on the left part of Figure 5 that show nor- viation from
targets Dtotal is defined and calculated as
malized mean delay deviation of each class from the respec- Dtotal = |DVi | where DVi is the normalized deviation of
tive delay target (normalization is with respect to priority class i (i = 1, 2, 3, and 4). Hence, evaluating from the left part
weights, that is, for class i shown is pi / p1 · actual deviation). of Figure 5 for TVFQ: Dtotal = 0.11(11%) + 0.0275(2.75%) +
As evident from Figure 4 and the bar chart on the left part 0.025(2.5%) + 0.0025(0.25%) = 0.165(16.5%). Similarly, the
of Figure 5, TVFQ scheme performs the worst resource allo- total deviations of heuristic and optimal FT-WFQ schemes
cations among the compared schemes as it does not imple- can be obtained as 12.6% and 8.5%, respectively. The to-
ment any target-tracking constraints. It forces high positive tal deviation Dtotal bar chart is shown on the right part of
delay deviations from targets for classes 1 and 2, respectively, Figure 5.
Oliver Yu et al. 11
1.4
2
1.2
1 1.5
Delay (frames)
Delay (frames)
0.8
1
0.6
0.4
0.5
0.2
0 0
0 100 200 300 400 0 100 200 300 400
Time Time

(a) (b)
5 10
4 8
Delay (frames)
Delay (frames)
3 6
2 4
1 2
0 0
0 100 200 300 400 0 100 200 300 400
Time Time

(c) (d)
Figure 6: Delay: for classes 1–4 (no estimation error).
The corresponding OPNET simulations results, namely, 16.5%, 11.7%, and 8.9%, respectively. It is also worth men-
running averages of delays and delay deviations are shown tioning that analysis and simulation results in Figures 4
in Figures 6 and 7, respectively. The results are in close and 6, respectively, show that the system stabilizes relatively
agreement with the corresponding analytical results (from quickly.
Figures 4 and 5) with differences in instantaneous values
mainly due to highly random nature of nonstationary ar-
rival process as defined in the previous section. As the an- 6.2. Results under arrival rate estimation error
alytical results, simulations from Figures 6 and 7 demon-
strate that the proposed schemes are able to minimize de- PSA analysis and OPNET simulations are employed to com-
viations from corresponding delay targets and allocate re- pare scheduling schemes (TVFQ, FT-WFQ optimal and heu-
sources efficiently. The total deviations Dtotal observed form ristic) in nonstationary environment with imperfect arrival
Figure 7 for TVFQ, heuristic, and optimal FT-WFQ are rate estimation. Consequently, class 1 arrival rate estimation
Normalized percent Total deviation

change from target Dtotal
20
15
Percentage (%)
10
5
1 2 3 4 5
Class
Optimal FT-WFQ
Heuristic FT-WFQ
TVFQ
Figure 7: Simulation: delay percent change from target (no estimation error).
error is assumed in this subsection with mean error ε of −0.5. the heuristic FT-WFQ achieves a reduction in mean devia-
Subsequently, the study was repeated using different error- tion as compared to regular TVFQ, while the optimal FT-
mean values. WFQ as expected distributes resources the most efficiently
The analytical delay running averages (measured in and achieves the largest reduction in the total deviation. It
frames) versus time for classes 1–4, respectively, are shown is also worth mentioning that when other mean-error val-
in Figure 8 (for ε = −0.5). The corresponding bar charts ues ε are used instead, similar results are observed. These
showing (normalized) mean delay deviations from delay tar- results are verified by OPNET simulation model, but are
gets for each class as well as the total deviations Dtotal are not shown due to space limitations. It is worth mentioning
shown on the left and right parts, respectively, of Figure 9. that simulation and analytical results are in strong agree-
As expected and evident from Figures 8 and 9, the regu- ment.
lar TVFQ is the most affected by an estimation error as To compare results under different error conditions, the
it is not able to correctly adapt resource (i.e., rate) alloca- above analysis (and simulation) is repeated for other mean
tions. From Figure 9 it can be seen that class 1 is affected the values of error ε. Figure 10 shows the total deviation (Dtotal )
most as TVFQ degrades its mean delay to 15% above cor- versus mean-estimation error (ε) observed for the compared
responding delay target. Note that resources not allocated schemes using PSA analysis. Shown are results for both un-
to class 1 are utilized by lower priority classes as evident derestimation (ε < 0) and overestimation (ε > 0) cases. Ob-
by unintended decrease of their mean delay (see Figure 9). serve that the optimal FT-WFQ achieves the lowest total
Thus, TVFQ wastes resources by allowing positive (nor- deviations for all mean-estimation errors (ε), and as the
malized) deviations from the targets (classes 1 and 2) even magnitude of error increases so does the gain of the op-
when negative (normalized) deviations are present (classes timal FT-WFQ compared to the other schemes. Almost
3 and 4). By employing feedback-enhanced reactive adapta- identical results are obtained by OPNET simulation (not
tion control, proposed FT-WFQ schemes achieve much bet- shown).
ter resource distributions thereby keeping mean (normal-
ized) delays closer to corresponding targets (see Figures 8
and 9, respectively) and are thus able to substantially mitigate 7. CONCLUSIONS
for the arrival rate error effect. From Figure 9, mean (nor-
malized) delay deviations of class 1 for the heuristic-based In this work, two versions of feedback-enhanced target-
and optimal FT-WFQs are 14% and 7.5%, respectively. Cal- tracking weighted fair queuing (FT-WFQ) schedulers, name-
culating similarly as in the last subsection from Figure 9, ly, heuristic and optimal, were presented. FT-WFQ is adap-
the total deviation from targets Dtotal for TVFQ, heuris- tive rate scheduler able to provide QoS guarantees on a “per-
tic, and optimal FT-WFQs are 29.5%, 24.5%, and 19.5%, class” basis. It dynamically adapts allocation rates based on
respectively (shown on the right part of Figure 9). Thus, the predictive and feedback-enhanced reactive adaptation
Oliver Yu et al. 13
1.4 2.5
1.2
2
1
Delay (frames)
Delay (frames)
1.5
0.8
0.6
1
0.4
0.5
0.2
0 0
0 50 100 150 200 250 300 350 400 0 50 100 150 200 250 300 350 400
Time Time

(a) (b)
4 9
3.5 8
3 7
6
2.5
Delay (frames)
Delay (frames)
5
2
4
1.5
3
1
2
0.5 1
0 0
0 50 100 150 200 250 300 350 400 0 50 100 150 200 250 300 350 400
Time Time

(c) (d)
Figure 8: Analysis: delay for classes 1–4 (estimation error).
control. Predictive control adjusts rate allocations based on Performance of the proposed scheduler was evaluated
the anticipated network conditions reflected by estimated analytically and by simulation in a nonstationary environ-
arrival rates. When the predictive adaptation control fails ment with and without arrival rate estimation error. PSA ap-
to maintain delay targets (due to estimation errors or ex- proximation was used for analytical delay modeling. It was
tremely high loading), the corrective feedback-enhanced re- shown that the proposed schemes are able to substantially
active control ensures that deviations from delay targets are reduce the error effect and minimize deviations from delay
minimized. targets.
Normalized percent Total deviation REFERENCES

change from target Dtotal
40 [1] H. Holma and A. Toskala, WCDMA for UMTS, John Wiley &
Sons, West Sussex, UK, 2000.
30 [2] O. Yu, E. Saric, and A. Li, “Fairly adjusted multimode dynamic
guard bandwidth admission control over CDMA systems,”
IEEE Journal on Selected Areas in Communications, vol. 24,
Percentage (%)
20 no. 3, pp. 579–592, 2006.

[3] L. Xu, X. Shen, and J. W. Mark, “Dynamic bandwidth alloca-
tion with fair scheduling for WCDMA systems,” IEEE Wireless
10
Communications, vol. 9, no. 2, pp. 26–32, 2002.
[4] L. Xu, X. Shen, and J. W. Mark, “Dynamic fair scheduling with
0 QoS constraints in multimedia wideband CDMA cellular net-
works,” IEEE Transactions on Wireless Communications, vol. 3,
no. 1, pp. 60–73, 2004.
10
1 2 3 4 5 [5] X. Wang, “An FDD wideband CDMA MAC protocol with
Class minimum-power allocation and GPS-scheduling for wireless
wide area multimedia networks,” IEEE Transactions on Mobile
Optimal FT-WFQ Computing, vol. 4, no. 1, pp. 16–28, 2005.
Heuristic FT-WFQ [6] C. Li and S. Papavassiliou, “Fair channel-adaptive rate
TVFQ scheduling in wireless networks with multirate multimedia
services,” IEEE Journal on Selected Areas in Communications,
vol. 21, no. 10, pp. 1604–1614, 2003.
Figure 9: Analysis: delay percent change from target (estimation
error). [7] M. Moustafa, I. Habib, and M. N. Naghshineh, “Efficient radio
resource control in wireless networks,” IEEE Transactions on
Wireless Communications, vol. 3, no. 6, pp. 2385–2395, 2004.
[8] I. F. Akyildiz, D. A. Levine, and I. Joe, “A slotted CDMA pro-
tocol with BER scheduling for wireless multimedia networks,”
IEEE/ACM Transactions on Networking, vol. 7, no. 2, pp. 146–
30 158, 1999.
[9] X. Wang, “A new scheduling scheme for the wideband TD-
CDMA MAC protocol,” in Proceedings of IEEE International
25 Conference on Communications (ICC ’01), vol. 3, pp. 974–978,
Total deviation Dtotal
Helsinki, Finland, June 2001.

20 [10] W. K. Wong, H. Zhu, and V. C.M. Leung, “Soft QoS provision-
ing using the token bank fair queuing scheduling algorithm,”
IEEE Wireless Communications, vol. 10, no. 3, pp. 8–16, 2003.
15
[11] W. K. Wong, H. Y. Tang, and V. C. M. Leung, “Token bank
fair queuing: a new scheduling algorithm for wireless multi-
10 media services,” International Journal of Communication Sys-
tems, vol. 17, no. 6, pp. 591–614, 2004.
[12] A. Stamoulis, N. D. Sidiropoulos, and G. B. Giannakis, “Time-
5
1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1 varying fair queueing scheduling for multicode CDMA based
Estimation error on dynamic programming,” IEEE Transactions on Wireless
TVFQ [13] C.-C. Li, S.-L. Tsao, M. C. Chen, Y. Sun, and Y.-M. Huang,
FT-WFQ (heuristic) “Proportional delay differentiation service based on weighted
FT-WFQ (optimal) fair queuing,” in Proceedings of the 9th International Conference
on Computer Communications and Networks (ICCCN ’00), pp.
Figure 10: Analysis: total deviation versus estimation error. 418–423, Las Vegas, Nev, USA, October 2000.
[14] O. B. Jennings and W. A. Massey, “A modified offered load
approximation for nonstationary circuit switched networks,”
Telecommunication Systems, vol. 7, no. 1–3, pp. 229–251, 1997.
[15] L. Green and P. Kolesar, “The pointwise stationary approxi-
mation for queues with nonstationary arrivals,” Management
ACKNOWLEDGMENTS Science, vol. 37, no. 1, pp. 84–97, 1991.
[16] T. Robertazzi, Computer Networks and Systems: Queueing The-
This work was supported in part by the National Science ory and Performance Evaluation, Springer, New York, NY, USA,
Foundation (NSF) under Contract SCI-0129527. The au- 2000.
thors are thankful to the editors of this special issue and to [17] I. Katzela, Modeling and Simulating Communication Networks:
the anonymous reviewers for their valuable comments and A Hands-on Approach Using OPNET, Prentice Hall, Upper
suggestions. Saddle River, NJ, USA, 1998.
Oliver Yu et al. 15
[18] V. Dham, L. A. DaSilva, A. Annamalai, and V. Srivastava, “A

phased array antenna model for space division multiple access
in wireless networks,” in Proceedings of OPNETWORK Confer-
ence, Washington, DC, USA, August 2002.
Oliver Yu received the B.A.S. (1981), M.A.S.

(1991), and Ph.D. (1997) degrees in elec-
trical and computer engineering from the
University of British Columbia in Canada.
He had extensive industry research and de-
velopment experiences. He had held techni-
cal leadership and engineering management
positions in Microtel, Nortel, Hughes, and
Motorola. He is currently a Research Asso-
ciate Professor at the Department of Elec-
trical and Computer Engineering in the University of Illinois at
Chicago. His research interests are in the areas of fourth-generation
wireless systems, all-optical photonic networks, next generation In-
ternet, and computing grid systems. He has been working on re-
search projects that are funded by the National Science Foundation
(NSF) and the Department of Energy (DOE). He received the Ca-
reer Award from DOE in 2003.
Emir Saric received the B.S. and M.S.
degrees in electrical and computer engi-
neering from the University of Illinois at
Chicago (UIC) in 2001, and 2004, respec-
tively. He is currently working towards the
Ph.D. degree in electrical and computer en-
gineering at UIC. Since 2004, he has been a
Research Assistant for Professor O. Yu at the
Networking and Wireless Communications
Laboratory, UIC. His research interests in-
clude QoS provisioning (MAC and admission control) and perfor-
mance evaluation of future cellular wireless networks and cognitive
wireless networks.
Anfei Li received the B.S. degree from Uni-
versity of Science and Technology of China,
Hefei, 2002. He is currently a Ph.D. can-
didate in the Department of Electrical and
Computer Engineering, the University of
Illinois at Chicago. His research interests in-
clude medium access control (MAC) in 3G
and 4G wireless communication systems,
wireless ad hoc and sensor networks, all-
optical networks, and computing grids.
DOI 10.1155/WCN/2006/70903
Effective Radio Resource Management for Multimedia

Broadcast/Multicast Services in UMTS Networks
Nuno Souto,1, 2 Armando Soares,2 Patricia Eusébio,2 Américo Correia,1, 2 and João C. Silva1
1 Instituto de Telecomunicações, Avenue Rovisco Pais 1, 1049-001 Lisboa, Portugal
2 Associaçãopara o Desenvolvimento das Telecomunicações e Técnicas de Informática, Avenue das Forças Armadas,
Edifı́cio ISCTE, 1600-082 Lisboa, Portugal
Broadcast and multicast offer a significant improvement of spectrum utilization, and become particularly important where in-
formation channels are shared among several users. Mobile cellular environments are expected to evolve with the technological
approaches necessary to facilitate the deployment of multimedia services, such as streaming, file download, and carousel services.
The perspective that video streaming in wireless networks services is an attractive service to end-users has spurred the research
in this area. To provide for a video delivery platform in UMTS, the third generation partnership project (3GPP) addressed this
problem with the introduction of the multimedia broadcast and multicast services (MBMS) in 3GPP Release 6. In this document
we analyse several effective radio resource management techniques to provide MBMS, namely, use of nonuniform QAM constel-
lations, multicode, and macrodiversity to guarantee the optimal distribution of QoS depending on the location of mobiles.
Copyright © 2006 Nuno Souto et al. This is an open access article distributed under the Creative Commons Attribution License,
1. INTRODUCTION Multimedia broadcast and multicast services (MBMS)

introduced by 3GPP in Release 6 are intended to efficiently
In a mobile cellular network it is often necessary to transmit use network/radio resources (by transmitting data over a
the same information to all the users (broadcast transmis- common radio channel), both in the core network but most
sion) or to a selected group of users (multicast transmission). importantly in the air interface of UMTS terrestrial radio ac-
Depending on the communication link conditions some re- cess network (UTRAN), where the bottleneck is placed to a
ceivers will have better signal-to-noise ratios (SNR) than oth- large group of users. However, it should take additional ac-
ers and thus the capacity of the communication link for these count of these network/radio resources. MBMS is targeting
users is higher. Cover [1] showed that in broadcast transmis- high (variable) bit rate services over a common channel.
sions it is possible to exchange some of the capacity of the One of the most important properties of MBMS is re-
good communication links to the poor ones and the trade- source sharing among several user equipments (UEs), mean-
off can be worthwhile. A possible method to improve the ef- ing that these users should be able to listen to the same
ficiency of the network is to use nonuniform signal constel- MBMS channel at the same time. Sufficient amount of power
lations (also called hierarchical constellations) which are able should be allocated to these MBMS channels so that arbitrary
to provide unequal bit error protection. In this type of con- UEs in the cell can receive the MBMS service.
stellations there are two or more classes of bits with differ- One of the key issues in multicast transmission is the
ent error protection, to which different streams of informa- management of radio resources. The main requirement is to
tion can be mapped. Depending on the channel conditions, make an efficient overall usage of the radio resources. This
a given user can attempt to demodulate only the more pro- makes the use of a common channel the favourite choice,
tected bits or also the other bits that carry the additional in- since many users can access the same resource at the same
formation. An application of these techniques is in the trans- time, but this depends also on the number of users belong-
mission of coded voice or video signals. Several papers have ing to the multicast group, the type of service provided, and
studied the use of nonuniform constellations for this pur- the QoS that it can guarantee.
pose [1, 2]. Nonuniform 16-QAM and 64-QAM constella- In this paper we will analyse several effective radio re-
tions are already incorporated in the DVB-T (digital video source management techniques to provide MBMS, namely,
broadcasting-terrestrial) standard [3]. the use of non-uniform QAM constellations, multicode, and
macrodiversity. The objective is to guarantee the optimal dis- 2.1. System model
tribution of QoS depending on the location of the mobiles.
According to the proposed transmission method UEs will
In Section 2 the multicode packet scheduling model is
receive the service accordingly to their geographic position.
presented, Section 3 describes non-uniform QAM constella-
The RNC accounting for the differences in Node Bs ra-
tions, macrodiversity combining techniques are detailed in
dio resource availability divides MBMS data by its priorities
Section 4, and in Section 5 simulation results are presented.
and transmits them in a fashion that suits each Node B. In
Finally some conclusions are drawn in Section 6.
Figure 1 this approach is shown, where we can see the in-
formation scalability in two separate physical channels for
2. MULTICODE PACKET SCHEDULING one MBMS service (256 kbps). This corresponds to the trans-
(TWO QOS REGIONS) mission of two data streams, each of 128 kbps, where ba-
sic information providing the basic QoS is transmitted with
Up to today no special transport channel has been specified the power level needed to cover the whole cell, and the sec-
for the purpose of multicast, but some proposal and prelim- ond stream conveys additional information to users near the
inary studies have been provided. Therefore the driving con- Node B.
cept to support multicast on the UTRAN is to use the existing The model consists of two QoS regions, where the first
transport channels, with minor modifications. region receives all information while the second region re-
A flexible common channel suitable for point-to-multi- ceives the most important data. The QoS regions are associ-
point (PtM) transmissions is already available, namely, the ated with the geometry factor that reflects the distance of the
forward access channel (FACH), which is mapped onto the UE from the base station antenna. The geometry factor G is
secondary common control physical channel (S-CCPCH). defined as the ratio of interference generated in the own cell
In [4], it was shown that about 40% of the sector total to the interference generated in the other cells plus thermal
power has to be allocated to a single 64 kbps MBMS if full cell noise, that is,
coverage is required. This makes MBMS too expensive since Iown
the overall system capacity is limited by the power resource. G= . (1)
Iothers + PN
To make MBMS affordable for the UMTS system, its
power consumptions have to be reduced. If MBMS is car- Table 1 shows the G values chosen. For the first region the
ried on S-CCPCH, there is no inner-loop power control. Ex- geometry factor is G = 0 dB and for the second region G =
tra power budget has to be allocated to compensate for the −6 dB.
receiving power fluctuations. UE1 will receive the most important data (transmitted at
Since MBMS video streaming is scalable, one way to im- 128 kbps) to get a basic video quality service, whereas UE2
prove the power efficiency of MBMS carried over S-CCPCH will receive all the data to provide a higher quality reproduc-
is to split the MBMS video streaming into several streams tion of the input video.
with a different quality of service (QoS). The basic video layer
is coded by itself to provide the basic video quality and the 3. NONUNIFORM QAM CONSTELLATIONS
enhancement video layer is coded to enhance the basic layer.
The enhancement layer when added back to the basic layer Another transmission method which is based on the same
regenerates a higher quality reproduction of the input video. philosophy of the multi-code transmission method just de-
Only the most important stream is sent to all the users in the scribed is the use of nonuniform constellations. In this study
cell to provide the basic service. The less important streams we consider the use of 16-QAM non-uniform modulations
are sent with less amount of power or coding protection and for the transmission of broadcast and multicast services in
only the users who have better channel conditions are able WCDMA systems. For 16-QAM two classes of bits are used.
to receive that additional information to enhance the video Some modifications to the physical layer of the UMTS-
quality. This way, transmission power for the most impor- (universal-mobile-telecommunications-systems-) based sys-
tant MBMS stream can be reduced because the data rate is tem to incorporate these modulations were already proposed
reduced, and the transmission power for the less important in [5, 6].
streams can also be reduced because the coverage require-
ment is relaxed. 3.1. 16-QAM
Two possible MBMS multicode schemes will be consid-
ered. The first one uses a single rate stream (single spreading These constellations are constructed using a main QPSK con-
code), which is carried on a single 256 kbps channel and sent stellation where each symbol is in fact another QPSK constel-
to the whole area in the cell. The second one uses a dou- lation, as shown Figure 2.
ble streaming transmission, that is, two data streams (two The bits used for selecting the symbols inside the small
spreading codes), each of 128 kbps where basic information inner constellations are called weak bits and the bits corre-
for basic QoS is transmitted with the power level needed to sponding to the selection of the small QPSK constellation are
cover the whole cell, and a second stream conveys additional called stronger bits. The idea is that the constellation can be
information to users near the Node B (base station). This viewed as a 16-QAM constellation if the channel conditions
way, Node B power can be saved trading off with QoS of UEs are good or as a QPSK constellation otherwise. In the latter
at cell borders. situation the received bit rate is reduced to half. The main
Nuno Souto et al. 3
UMTS core network
Internet
SGSN
GGSN
Streaming server
RNS
RNC
UE1
Node B
UE2
QoS region 1
QoS region 2
Cell boundary
Enhancement layer
Basic layer
GGSN: gateway GPRS support node
SGSN: serving GPRS support node
RNC: radio network controller
UE: user equipment
Figure 1: Two QoS regions packet scheduling model.
Table 1: QoS regions parameters. Figure 3 shows a simplified transmission chain incorpo-
QoS region EU capacity Maximum bit rate G (dB)
rating 16-QAM non-uniform constellations. In this scheme
there are 2 parallel processing chains, one for the basic infor-
1 UE1 256 kbps 0
mation stream and the other for the enhancement informa-
2 UE2 128 kbps −6
tion.
parameter for defining one of these constellations is the ratio 4. MACRODIVERSITY COMBINING
between d1 and d2 as shown in Figure 2:
d1 Macrodiversity combining (MDC) is proposed as an en-
= k, where 0 < k ≤ 0.5. (2) hancement to the UMTS 3GPP Release 6 MBMS. In a point-
d2
to-multipoint (PtM) MBMS service the transmitted con-
Each symbol s of the constellation can be written as tent is expected to be network specific rather than cell spe-
cific, that is, the same content is expected to be multicas-
d2 d1 d d
s= ± ± + ± 2 ± 1 j. (3) ted/broadcasted through the entire network or through most
2 2 2 2
of it. Therefore, a natural way of improving the physical
If k = 0.5, the resulting constellation is a uniform 16-QAM. layer performance is to take advantage of macrodiversity. On
When k is lower than 0.5, the bit error rate (BER) of the the network side, this means ensuring sufficient time syn-
stronger bits improves but since the BER of the weaker sym- chronization of identical MBMS transmissions in different
bols decreases, the overall BER also decreases. cells; on the mobile station side, this means the capability to
combining them with different weights yields an optimum

I solution under an AWGN channel.
1000 1010 0010 0000
The output of the receiver can be represented as
1001 1011 0011 0001

N
2
y(t) = g j sm (t) + n j (t). (5)
j =1
d1
5. SIMULATION RESULTS
Q
Typically, radio network simulations can be classified as ei-
1101 1111 0111 0101
ther link level (radio link between the base station and the
user terminal) or radio network subsystem system level. A
1100 1110 0110 0100
single approach would be preferable, but the complexity
of such simulator—including everything from transmitted
d2 waveforms to multicell network—is far too high for the re-
quired simulation resolutions and simulation time. There-
fore, separate link and system level approaches are needed.
Figure 2: Signal constellation for 16-QAM nonuniform modula-
tion.
Link level simulations are necessary for building a re-
ceiver model in the system simulator that can predict the re-
ceiver block error rate (BLER) and BER performance, taking
receive and decode the same content from multiple transmit- into account channel estimation, interleaving, and decoding.
ters simultaneously. The system level simulator is needed to model a system with
Basically the diversity combining concept consists of re- a large number of mobiles and base stations, and algorithms
ceiving redundantly the same information bearing signal operating in such a system.
over two or more fading channels, and combine these mul- Table 2 presents some link level parameters which will
tiple replicas at the receiver in order to increase the overall be used in the following sections. The channel estimation is
received SNR. performed using the common pilot channel (CPICH) which
In macro diversity the received signals from different is transmitted in parallel to the data channels, using an or-
paths have to be processed using some sort of combining al- thogonal reserved code. At the receiver, the modulation is re-
gorithm. In this study two different combining procedures moved from the CPICH by multiplying it by its conjugate,
are considered, namely, selective combining (SC) and maxi- which results in a sequence of noisy channel estimates. These
mal ratio combining (MRC). noisy channel estimates are then passed through a moving
average filter and the filtered sequence can be interpolated
(or decimated) to match the rate of the data channels. 3GPP
4.1. Selective combining [4] refers to Vehicular A and Pedestrian B channel models as
Figure 4 shows a scheme of how selective combining oper- representative for the macrocellular environment and there-
ates at the receiver side. With SC the path/branch yielding fore results will be presented along this study for these two
the highest SNR is always selected. In order to guarantee that models. The velocities of 3 and 30 km/h were presented in
the receiver uses the path with the best quality a simultaneous 3GPP [4] for the Vehicular A channel, where 3 km/h has pro-
and continuous monitoring of all diversity paths is required. vided worst performance results.
The output of the diversity combiner will be Table 3 shows the system level assumptions used for the
simulations.

y(t) = gk · sm (t) + nk (t), with gk = max g1 , . . . , gN , The link performance results are used as an input by the
(4) system level simulator where several estimates for coverage
and throughput purposes can be made by populating the sce-
where gk is the maximum amplitude of the fading co- nario topology uniformly and giving users a random mobil-
efficients, and nk (t) is the additive Gaussian white noise ity. The estimates are made for every transmission time in-
(AGWN) which is independent from branch to branch. terval (TTI) being the packets that are received with a BLER
below 1% considered to be well received. The estimates for
4.2. Maximal ratio combining coverage purposes are made for an average of five consecu-
tive received packets; if the average received BLER of these
The maximal ratio combining (Figure 5), although being the packets is below the 1% BLER, the mobile user is considered
most complex combining technique presented, is the op- as being in coverage. For the throughput calculation the es-
timum way to combine the information from the differ- timation is made based on each individual packet received
ent paths/branches. The receiver corrects the phase rotation with a BLER lower than 1%.
caused by a fading channel and then combines the received Figure 6 shows the geometry CDF function values ob-
signals of different paths proportionally to the strength of tained for the macrocellular environment. The geometry fac-
each path. Since each path undergoes different attenuations, tor was previously defined in Section 2.1; a lower geometry
Nuno Souto et al. 5
Interleaver Modulation
Basic mapper
information Channel coding Physical channel Spreading and
Rate matching for physical
stream (turbo code) segmentation scrambling
Interleaver channel
1
Xk
2 parallel P physical
chains channels
Interleaver Modulation
Enhancement mapper Pilot
Channel coding Rate matching Physical channel Spreading and
information for physical channel
stream (turbo code) segmentation scrambling
Interleaver channel
P
Figure 3: Proposed transmitter chain.
SNR Select max

monitor SNR
g1 sm (t) + n1 (t)
Channel 1
Transmitter
g2 sm (t) + n2 (t)
y(t)
Channel 2 Receiver
..
.
gN sm (t) + nN (t)
Channel N
Figure 4: Selective combining.
factor is expected when user is located at the cell edge (the considering the VehA propagation channel. This means that
case where the interference received from the neighbouring we can only offer such a high bit rate for users located in
cells is higher than the interference experienced in its own the middle of the cell, not near the border. By using a mul-
cell). ticode transmission (2 spreading codes with SF = 16) with
The cumulative distribution function (CDF) of geometry two different transmission powers, each assuring a bit rate of
can be obtained through uniformly distributing a large num- 128 kbps, offering different QoS that depend on the location
ber of mobile users over the topology and calculating the G of the UEs, higher throughput is achieved with lower total
at each position. transmission power from the Node B.
From Figure 6 it is possible to notice that for the studied In Figures 8–10, the QPSK 1% BLER coverage versus
scenario about 95% of the users experience a geometry factor MBMS channel power (Node-B Tx Ec/Ior) is shown with se-
of −6 dB or better, 80% experience a geometry of −3 dB or lective combining or maximal ratio combining over 1 and 2
better, and about 62% of the users experience a geometry of radio links (RLs), respectively, for the studied path models
0 dB or better. and TTI of 40 ms and 80 ms. Due to the better operation of
Figure 7 presents the first results obtained with the link the turbo decoder with increasing TTI (increasing encoded
level simulator. The results are presented in terms of Ec/Ior block sizes) we observe a decrease in the required transmit-
(dB) representing the fraction of cell transmit power neces- ted power from the Node B when we use TTI = 80 ms instead
sary to achieve the corresponding BLER performance grad- of 40 ms. However, due to the limiting transport block size of
uated on the vertical axis. For the reference BLER = 10−2 5114 bits per block of the turbo encoder specified in 3GPP,
and bit rate of 256 kbps (use of a single spreading code with the bit rate of 256 kbps does not allow an increase in the en-
spreading factor SF = 8) we need to have a geometry fac- coded block size for TTI = 80 ms. As expected, the average
tor of 0 dB in order to achieve Ec/Ior less than 80% (−1 dB) coverage of maximal ratio combining is always better than

N
y(t) = g j 2 sm (t) + n j (t)
j =1
g1
Channel 1
Transmitter
g2
Channel 2 y(t)

Receiver
..
.
gN
Channel N
Figure 5: Maximal ratio combining.
Table 2: Link level simulation parameters.
Parameter Value
S-CCPCH slot format 12 (128 kbps)
Transport block size & number of transport blocks Varied according to information bit rate
per TTI (128 or 256 kbps) and TTI value
CRC 16 bits
Transmission time interval (TTI) 20 ms
CPICH Ec/Ior −10 dB (10%)
P-SCH (primary-synchronization channel) Ec/Ior −15 dB (3%)
S-SCH (secondary-synchronization channel) Ec/Ior −15 dB (3%)
Tx Ec/Ior Varied
OCNS (orthogonal channel noise simulation) Used to sum the total Tx Ec/Ior to 0 dB (100%)
Channel estimation Enabled
Power control Disabled
Channels Pedestrian B, 3 km/h, Vehicular A, 3 km/h
selective combining and increasing the number of received Table 3: System level assumptions.
radio links provides reduction in the transmitted power in- Parameter Value
dependently of the combining technique.
Cellular layout Hexagonal
In Figure 8, for the reference average coverage of 90%
the required Ec/Ior is about 60%–65% (PedB-VehA) for Sectorization Yes, 3 sector/cell
128 kbps. For 256 kbps the same values of Ec/Ior allow aver- Site-to-site distance 1000 m
age coverage around 52%–55%. There is the need of multi- Number of base stations 18
code or macrodiversity combining to allow an increase of bit Base station antenna gain 17.5 dBi
rate and average coverage and/or a reduction in transmitted Antenna beamwidth, −3 dB 70 degrees
power. With multi-code the bit rate of 256 kbps is achiev- Antenna front-to-back ratio 20 dB
able with two streams of 128 kbps, one of them requiring Propagation model Okamura-Hata
Ec/Ior1 = 30% (62% coverage in PedB environments) and Base station total Tx power (sector) 43 dBm
the other Ec/Ior2 = 50% (85% coverage in PedB). Thermal noise DL −103.3 dBm
However, macrodiversity offers better coverage and re- Orthogonally factor 0.4
duction of transmitted power than multicode. Tables 4 and
Std of shadow fading 10 dB
5 illustrate the required Ec/Ior for the reference BLER =
Cable losses 3 dB
1% using macrodiversity with Vehicular A and Pedestrian B
propagation channels, respectively. The performance of the
former is always a little bit worse. According to the results
of Tables 4 and 5 up to two MBMS channels with 256 kbps employed and considering that the maximum total available
could be transmitted at the same time if MRC with 2RL were Ec/Ior ≤ 83%.
Nuno Souto et al. 7
100 100
User locations with geometry < ascissa (%)
90 90
80 80
Average coverage (%)

70 70
60 60
50 50
40 40
30 30
20 20
10
10
0
0 0 10 20 30 40 50 60 70 80 90 100
10 8 6 4 2 0 2 4 6 8 10 12 14 16 18 20
S-CCPCH Ec/Ior (%)
Geometry (dB)
Urban macrocell VehA, 128 kbps, 80 ms TTI (1RL)

PedB, 128 kbps, 80 ms TTI (1RL)
VehA, 256 kbps, 40 ms TTI (1RL)
Figure 6: Geometry CDF, urban macrocell scenario. PedB, 256 kbps, 40 ms TTI (1RL)
100
Figure 8: QPSK average coverage versus Tx power (1RL).
100
90
10 1
80
70
BLER
60
10 2 50
40
30
20
10 3
12 10 8 6 4 2 0 10
Ec/Ior (dB)
0
0 10 20 30 40 50 60 70 80 90 100
PedB, 128 kbps, 80 ms TTI (G = 3 dB)
S-CCPCH Ec/Ior (%)
VehA, 128 kbps, 80 ms TTI (G = 3 dB)
PedB, 256 kbps, 40 ms TTI (G = 0 dB) VehA, 128 kbps, 80 ms TTI (2RL-SC)
VehA, 256 kbps, 40 ms TTI (G = 0 dB) PedB, 128 kbps, 80 ms TTI (2RL-SC)
VehA, 256 kbps, 40 ms TTI (2RL-SC)
Figure 7: BLER versus Tx power for QPSK, different bit rates and PedB, 256 kbps, 40 ms TTI (2RL-SC)
geometries (V = 3 km/h).
Figure 9: QPSK average coverage versus Tx power (2RL-SC).
Figure 11 presents an alternative way of offering the bit

rate of 256 kbps using nonuniform 16-QAM modulation and are only considering k = 0.5 (uniform 16-QAM constella-
a single spreading code with SF = 16 for G = 0 dB. This case tion).
is more spectral efficient than the previous one presented Figure 12 corresponds to Figure 11 with SF = 32 and ge-
in Figure 7 because it uses a higher SF but there is the dis- ometry G = −3 dB. In this case the maximum achievable
advantage of requiring a more complex receiver. An itera- bit rate is 128 kbps. For BLER = 10−2 the difference of total
tive receiver based on the one described in [5] is employed transmitted power between the strong and the weak blocks
for decoding both blocks of bits. For the reference value of also is 5.5 dB for either Vehicular A or Pedestrian B. The com-
BLER = 10−2 the difference of total transmitted power be- parison between Figures 11 and 12 indicates that we can de-
tween the strong and the weak blocks is about 5.5 dB for ei- crease the bit rate (increase of spreading factor) by decreas-
ther Vehicular A or Pedestrian B. Notice that in this study we ing the geometry (increasing of other cells interference). It
100 100
90
80
70 1
10
60
BLER
50
40
10 2
30
20
10
0 10 3
0 10 20 30 40 50 60 70 80 90 100 12 10 8 6 4 2 0
S-CCPCH Ec/Ior (%) Ec/Ior (dB)
VehA, 128 kbps, 80 ms TTI (2 RL-MRC) VehA, SF = 16, strong blocks (G = 0 dB)
PedB, 128 kbps, 80 ms TTI (2 RL-MRC) VehA, SF = 16, weak blocks (G = 0 dB)
VehA, 256 kbps, 40 ms TTI (2 RL-MRC) PedB, SF = 16, strong blocks (G = 0 dB)
PedB, 256 kbps, 40 ms TTI (2 RL-MRC) PedB, SF = 16, weak blocks (G = 0 dB)
Figure 10: QPSK average coverage versus Tx power (2RL-MRC). Figure 11: BLER versus Tx power for 16-QAM strong and weak
blocks of bits (SF = 16), k = 0.5.
Table 4: Vehicular A, 3 km/h, 90% coverage, 1% BLER. 100
Bit rate TTI length 1RL SC (2RL) MRC (2RL)

−1.87 dB −4.61 dB −7.59 dB
128 kbps 80 ms
64.9% 34.6% 17.4% 1
10
— — −3.89 dB
256 kbps 40 ms
BLER
— — 40.75%
10 2
Table 5: Pedestrian B, 3 km/h, 90% coverage, 1% BLER.
Bit rate TTI length 1RL SC (2RL) MRC (2RL)

−2.39 dB −4.92 dB −8.09 dB
128 kbps 80 ms 10 3
57.6% 32.2% 15.5% 12 10 8 6 4 2 0
— — −4.10 dB Ec/Ior (dB)
256 kbps 40 ms
— — 38.9% VehA, SF = 32, strong blocks (G = 3 dB)
VehA, SF = 32, weak blocks (G = 3 dB)
PedB, SF = 32, strong blocks (G = 3 dB)
PedB, SF = 32, weak blocks (G = 3 dB)
means that we must decrease the bit rate if we intend to al-
low an increase of coverage. This is true independently of the
site-to-site distance between base stations (Node Bs). Figure 12: BLER versus Tx power for 16-QAM strong and weak
blocks of bits (SF = 32), k = 0.5.
In Figures 13–15, the 16-QAM 1% BLER coverage ver-
sus MBMS transmitted channel power (Node-B Tx Ec/Ior)
is shown with selective and maximal ratio combining over 1
and 2 radio links (RLs), for the studied path models and a for VehA (actually the 90% coverage for weak bocks is not
TTI of 40 ms. achievable for the later propagation channel with a single
In Figure 13, the performance of the conventional 1 ra- radio link). As expected the average coverage of the strong
dio link (RL) reception is illustrated for comparison with blocks is always much better than weak blocks. However, this
reception using macrodiversity combining. For the refer- difference tends to decrease as the number of radio links in-
ence average coverage of 90% and 1RL the difference of re- creases; for instance, in the 90% average coverage with 2RL
quired Ec/Ior between strong blocks and weak ones is about and MRC, the difference of required Ec/Ior is only 15% for
70% (PedB) and even higher percentage of Ec/Ior is required PedB (see Figure 15).
Nuno Souto et al. 9
100 100
90 90
80 80

70 70
60 60
50 50
40 40
30 30
20 20
10 10
0 0
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
S-CCPCH Ec/Ior (%) S-CCPCH Ec/Ior (%)
VehA, strong blocks (1RL) VehA, strong blocks (2RL-MRC)

VehA, weak blocks (1RL) VehA, weak blocks (2RL-MRC)
PedB, strong blocks (1RL) PedB, strong blocks (2RL-MRC)
PedB, weak blocks (1RL) PedB, weak blocks (2RL-MRC)
Figure 13: 16-QAM average coverage versus Tx power (1RL). Figure 15: 16-QAM average coverage versus Tx power (2RL-MRC).
100 256
90 224
80
Average throughput (kbps)
192
70
160
60
50 128
40 96
30
64
20
32
10
0 0
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
VehA, strong blocks (2RL-SC) VehA (1RL)

VehA, weak blocks (2RL-SC) PedB (1RL)
PedB, strong blocks (2RL-SC)
PedB, weak blocks (2RL-SC) Figure 16: 16-QAM average throughput versus Tx power (1RL).
Figure 14: 16-QAM average coverage versus Tx power (2RL-SC).

is 256 kbps, we need macrodiversity combining. With 2RL-
SC (Figure 17) we can observe a smooth step in the through-
Figures 16 and 17 show the 1% BLER throughput ver- put between 96 and 128 kbps, especially for the VehA chan-
sus MBMS transmitted channel power (Node-B Tx Ec/Ior) nel due to the way SC operates and the difference of required
with selective combining and maximal ratio combining over Ec/Ior between weak and strong blocks. We recall that for
1 and 2 radio links (RLs) for various channel models and 128 kbps only the strong blocks are correctly received. With
TTI = 40 ms. In Figure 16, the performance of the conven- 2RL-MRC there is no such behaviour around 128 kbps be-
tional 1 radio link (RL) reception is illustrated for compar- cause of the way this diversity combining operates. As ex-
ison. The maximum throughput of 256 kbps is not achiev- pected, the reference throughput is achieved with less Ec/Ior
able with 1RL, for both propagation channels, due to the low for MRC compared to SC.
coverage of weak blocks. To achieve the reference through- In Figures 18–20, the 1% BLER throughput versus
put between 95% and 99% of the maximum bit rate, which MBMS channel power (Node-B Tx Ec/Ior) is shown with
256 256
224 224

192 192
160 160
128 128
96 96
64 64
32 32
0 0
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
VehA (2RL-SC) VehA, 128 kbps, 80 ms TTI (2RL-SC)

PedB (2RL-SC) PedB, 128 kbps, 80 ms TTI (2RL-SC)
VehA (2RL-MRC) VehA, 256 kbps, 40 ms TTI (2RL-SC)
PedB (2RL-MRC) PedB, 256 kbps, 40 ms TTI (2RL-SC)
Figure 17: 16-QAM average throughput versus Tx power Figure 19: QPSK average throughput versus Tx power (2RL-SC).
(SC/MRC).
256 256
224 224
192
192
160 160
128 128
96 96
64 64
32 32
0 0
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
VehA, 128 kbps, 80 ms TTI (1RL) VehA, 128 kbps, 80 ms TTI (2RL-MRC)
PedB, 128 kbps, 80 ms TTI (1RL) PedB, 128 kbps, 80 ms TTI (2RL-MRC)
VehA, 256 kbps, 40 ms TTI (1RL)
VehA, 256 kbps, 40 ms TTI (2RL-MRC)
PedB, 256 kbps, 40 ms TTI (1RL)
PedB, 256 kbps, 40 ms TTI (2RL-MRC)
Figure 18: QPSK average throughput versus Tx power (1RL). Figure 20: QPSK average throughput versus Tx power (2RL-MRC).
maximal ratio combining and selective combining over 1 and 16-QAM performance results presented in Figure 16. How-
2 radio links, for the various channel models, TTI lengths, ever, this difference tends to decrease as the number of radio
and spreading factors based on Release 6 results [4] (named links increases. This means that the benefits of using macro-
QPSK in the caption). The performance of these R6 through- diversity combining are higher for QPSK than 16-QAM.
put results is illustrated for comparison, with the corre- Considering the reference bit rate of 256 kbps and refer-
sponding average throughput illustrated in Figures 16 and ence coverage of 95% with macrodiversity by maximal ra-
17. tio combining 2 radio links (2RL-MRC), the capacity gain
In Figure 18 we can check that for a 256 kbps bit rate of using nonuniform16-QAM is 0.2 dB + 3 dB = 3.2 dB. The
over 1RL the performance of QPSK is clearly worse than the 0.2 dB comes from the comparison of Figures 17 and 20 for
Nuno Souto et al. 11
the Vehicular A channel, and the last 3 dB is due to the use of Nuno Souto graduated in aerospace engi-
SF = 16 instead of SF = 8, which allows for using the double neering—avionics branch, in 2000 from In-
of the channels. stituto Superior Técnico, Lisbon, Portugal.
From November 2000 to January 2002 he
worked as a researcher in the field of au-
6. CONCLUSIONS tomatic speech recognition for Instituto de
Engenharia e Sistemas de Computadores,
In this paper we have analysed several effective radio resource Lisbon, Portugal. He is currently working
management techniques to provide MBMS, namely, use of for his Ph.D. degree in electrical engineering
nonuniform QAM constellations, multicode, and macrodi- in Instituto Superior Técnico. His research
versity to guarantee the optimal distribution of QoS depend- interests include wideband CDMA systems, OFDM, channel cod-
ing on the location of mobiles. In this study we have also ing, channel estimation, and MIMO systems.
presented the expected capacity gains that multicode and
nonuniform 16-QAM modulations with more complex re- Armando Soares graduated in telecommu-
ceivers can provide to reduce the PtM MBMS channel power. nication and computer science engineering
The latter receivers are more power efficient than current at Instituto Superior de Ciências do Tra-
receivers based on QPSK modulation. We have shown that balho e da Empresa, Lisbon, Portugal. He
macrodiversity combining offers better capacity gains than is currently finishing his M.S. degree at
multi-code for broadcast/multicast services. The use of both the same university. He has been working
techniques at the same time is suggested. Non-uniform 16- as a researcher in the fields of radio re-
QAM receivers should be built in the near future with or sources optimisation and efficient alloca-
without the macrodiversity combining already specified by tion for 3G UMTS networks and beyond.
These research activities are being devel-
3GPP, as an effective mean to increase not only the through-
oped in ADETTI/ISCTE. Since 2003, he has been working in In-
put, but also the number of simultaneous simulcast services. formation Society Technologies EU funded telecommunications
research projects, namely, SEACORN, B-BONE, and currently C-
ACKNOWLEDGMENT MOBILE.
The authors would like to thank the European Commission Patricia Eusébio graduated in telecommu-
project IST-2003-507607 Broadcasting and Multicasting over nication and computer science engineering,
Enhanced UMTS Mobile Broadband Networks, B-BONE, in 2003 from Instituto Superior de Ciências
which has partially funded this work. do Trabalho e da Empresa (ISCTE), Lis-
bon, Portugal. She received the M.S. degree
in computer science and telecommunica-
REFERENCES tion engineering at the same university in
2005. From February 2004 to January 2005,
[1] T. Cover, “Broadcast channels,” IEEE Transactions on Informa- she was with ISCTE as an Assistant Teacher
tion Theory, vol. 18, no. 1, pp. 2–14, 1972. and she was enrolled in some EU funded
[2] M. B. Pursley and J. M. Shea, “Nonuniform phase-shift-key telecommunications research projects, namely, SEACORN and B-
modulation for multimedia multicast transmission in mobile BONE. She is currently working as a Project Manager in a telecom-
wireless networks,” IEEE Journal on Selected Areas in Communi- munication and energy company—Netplan, in RF planning and
cations, vol. 17, no. 5, pp. 774–783, 1999. optimisation area.
[3] “Digital video broadcasting (DVB) framing structure, channel
coding and modulation for digital terrestrial television (DVB-
T),” March 1997. ETSI, European Telecommunication Standard Américo Correia received the B.Sc. degree
ETS 300 744. in electrical engineering from the University
of Angola in 1983, the M.S. and Ph.D. de-
[4] 3GPP25.803, “S-CCPCH Performance for MBMS”. grees from Istituto Superior Técnico (IST),
[5] N. Souto, J. C. Silva, R. Dinis, and F. Cercas, “Iterative turbo Lisbon, Portugal, in 1990 and 1994, respec-
multipath interference cancellation for WCDMA systems with tively. From 1991 to 1999 he was with IST
non-uniform modulations,” in Proceedings of IEEE 61st Vehicu- as an Assistant Professor. He is currently
lar Technology Conference (VTC ’05), vol. 2, pp. 811–815, Stock- with Instituto Superior de Ciências do Tra-
holm, Sweden, May-June 2005. balho e da Empresa (ISCTE), Lisbon, Portu-
[6] N. Souto, J. C. Silva, F. Cercas, and R. Dinis, “Non-uniform gal. He visited Nokia Research Center from
constellations for broadcasting and multicasting services in September to December 1998 as a visiting scientist. From Septem-
WCDMA systems,” in Proceedings of IEEE 14th IST Mobile ber 2000 to August 2001 he joined Ericsson Eurolab Netherlands.
& Wireless Communications Summit, Dresden, Germany, June His current research topics include wideband CDMA, MIMO, ra-
2005. dio resource management, and multimedia broadcast/multicast
services.
João C. Silva received the B.S. degree in

aerospace engineering from Instituto Supe-
rior Técnico (IST), Lisbon Technical Uni-
versity, in 2000. From 2000 to 2002 he
worked as a business consultant on McKin-
sey&Company. From 2002 onwards, he has
been working on his Ph.D. thesis at IST,
focusing on spread spectrum techniques,
multiuser detection schemes, and MIMO
systems.
DOI 10.1155/WCN/2006/24616
Impact of Video Coding on Delay and Jitter in 3G Wireless

Video Multicast Services
Kostas E. Psannis and Yutaka Ishibashi
Department of Computer Science and Engineering, Graduate School of Engineering, Nagoya Institute of Technology,
Nagoya 466-8555, Japan
We present an efficient method for supporting wireless video multicast services. One of the main goals of wireless video multicast
services is to provide priority including dedicated bandwidth, controlled jitter (required by some real-time and interactive traf-
fic), and improved loss characteristics. The proposed method is based on storing multiple differently encoded versions of the video
stream at the server. The corresponding video streams are obtained by encoding the original uncompressed video file as a sequence
of I-P(I)-frames using a different GOP pattern. Mechanisms for controlling the multicast service request are also presented and
their effectiveness is assessed through extensive simulations. Wireless multicast video services are supported with considerably
reduced additional delay and acceptable visual quality at the wireless client-end.
Copyright © 2006 K. E. Psannis and Y. Ishibashi. This is an open access article distributed under the Creative Commons
Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is
properly cited.
1. INTRODUCTION H.264 [3–6]. Given a bandwidth of several hundred of kilo-

bits per second, the recent codecs, such as MPEG-4, can effi-
Multimedia transport typically requires stringent QoS met-
ciently transmit quality video.
rics (bandwidth and delay and jitter guarantees). However, in
addition to unreliable wireless channel effects, it is very hard An MPEG video stream comprises intra-frames (I), pre-
to maintain an end-to-end route which is both stable and has dicted frames (P), and interpolated frames (B) [3–5]. Ac-
enough bandwidth in an ad hoc network. The rapid growth cording to MPEG coding standards, I-frames are coded such
of wireless communications and networking protocols will that they are independent of any other frames in the se-
ultimately bring video to our lives anytime, anywhere, and quence; P-frames are coded using motion estimation and
on any device. each one has a dependency on the preceding I- or P-frame;
Until this goal is achieved, wireless video delivery faces finally the coding of B-frames depends on the two “an-
numerous challenges, among them highly dynamic network chor” frames—the preceding I/P-frame and the following
topology, high error rates, limited and unpredictably vary- I/P-frame. An MPEG coded video sequence is typically par-
ing bit rates, and scarcity of battery power. Most emerg- titioned into small intervals called GOP (group of pictures).
ing and future mobile client devices will significantly differ Streaming of live or stored video content to group of mo-
from those used for speech communications only; handheld bile devices comes under the scope of multimedia broad-
devices will be equipped with color display and a camera, cast/multicast service (MBMS) standard [7]. MBMS stan-
and have sufficient processing power to allow presentation, dardization is still in process. It seems that its pure commer-
recording, and encoding/decoding of video sequences. In ad- cialization will need at least three more years. Some of the
dition, emerging and future wireless systems will provide suf- typical applications are subscription to live sporting, events,
ficient bit rates to support video communication applica- news, music, videos, traffic and weather reports, and live TV
tions. Nevertheless, bit rates will always be scarce in wire- content. MBMS has two modes in practice: broadcast mode
less transmission environments due to physical bandwidth and multicast mode. The difference between broadcast and
and power limitations; thus, efficient video compression is multicast modes is that the user does not need to subscribe in
required [1, 2]. each broadcast service separately, whereas in multicast mode,
In the last decade, video compression technologies have the services can be ordered separately. The subscription and
evolved in the series of MPEG-1, MPEG-2, MPEG-4, and group joining for the multicast mode services could be done
by the mobile network operator, the user him/herself, or a Moreover, video resilience techniques using hierarchical al-
separate service provider. The current understanding about gorithms are proposed where transmission of I-, P-, and B-
the broadcast mode is that the services are not charged, frames is sent with varying levels of FEC protection. Some
whereas the multicast mode can provide services that are of the prior research works on error resilience for broadcast
billed. Specifically MBMS standard specifies transmission of terminals focus on increasing FEC based on the feedback
data packets from single entity to multiple recipients. The statistics for the user [14]. A comparison of different error
multimedia broadband-multicast service center should be resilience algorithms for wireless video multicasting on wire-
able to accept and retrieve content from external sources and less local area networks is reported in [15]. However, in the
transmit it using error resilient schemes. literature survey none of the methods applied error resilience
In recent years several error resilience techniques have techniques at the video coding level to support multicasting
been devised [8–15]. In [8], an error resilience entropy cod- services.
ing (EREC) has been proposed. In this method the incoming Error resilient (re-) encoding is a technique that enables
bitstream is reordered without adding redundancy such that robust streaming of stored video content over noisy channels.
longer VLC blocks fill up the spaces left by shorter blocks in a It is particularly useful when content has been produced in-
number of VLC blocks that form a fixed-length EREC frame. dependent of the transmission network conditions or under
The drawback of this method is that the codes between two dynamically changing network conditions.
synchronization markers are dropped, results in that any This paper focuses on signaling aspects of mobile clients,
VLC code in the EREC frame be corrupted due to trans- such as joining or leaving a multicast session of multimedia
mission errors. A rate-distortion frame work with analytical delivery. Developing error resilience technique which pro-
models that characterize the error propagation of the cor- vides high quality of experience to the end mobile user is a
rupted video bitstream subjected to bit errors was proposed challenging issue. In this paper we propose a very efficient er-
[9]. One drawback of this method is that it assumes that ror resilience technique for MBMS. Similar to [16] by encod-
the actual rate-distortion characteristics are known, which ing separate copies of the video, the multicast video stream
makes the optimization difficult to be realized practically. is supported with minimum additional resources. The corre-
In addition the error concealment is not considered. Error sponding version is obtained by encoding every (i.e., uncom-
concealment has been available since H.261 and MPEG-2 pressed) frame of the original movie as a sequence of I-P(I)-
[4]. The easiest and most practical approach is to hold the frames using a different GOP pattern.
last frame that was successfully decoded. The best known The paper is organized as follows. In Section 2 the mul-
approach is to use motion vectors that can adjust the im- timedia broadcast/multicast service standard is briefly dis-
age more naturally when holding the previous frame. More cussed. In Section 3 the problem of supporting multimedia
complicated error concealment techniques consist of a com- broadcast/multicast service over wireless networks is formu-
bination of spatial, spectral, and temporal interpolations lated. In Section 4 the preprocessing steps required to sup-
with motion vector estimation. In [10] an error resilience port efficient multicast streaming services over wireless net-
transcoder for general packet radio service (GPRS) mobile works are detailed. Section 5 presents the extensive simula-
accesses networks is presented. In this approach the bit allo- tions results. Finally conclusions are discussed in Section 6.
cation between insertion error resilience and the video cod-
ing is not optimized. In [11] optimal error resilience inser- 2. MULTIMEDIA BROADCAST/MULTICAST SERVICE
tion is divided into two subproblems: optimal mode selec-
tion for macroblocks and optimal resynchronization marker Third generation partnership project (3GPP) has standard-
insertion. Moreover, in [12] an approach to recursively com- ized four types of visual content delivery services and tech-
pute the expected decoder distortion with pixel-level preci- nologies.
sion to account for spatial and temporal error propagation
(i) Circuit-switched multimedia telephony [17].
in a packet loss environment is proposed. In both meth-
(ii) End-to-end packet-switched streaming (PSS) [18].
ods [11, 12], interframe dependencies are not considered. In
(iii) Multimedia messaging service (MMS) [19].
MPEG-4 video standard [5], application layer error resilient
(iv) Multimedia broadcast/multicast service (MBMS) [7].
tools were developed. At the source coder layer, these tools
provide synchronization and error recovery functionalities. The first three mobile applications assume the point-to-
Efficient tools are resynchronization marker and adaptive point model, where two single end-points (e.g., client-server)
intra-frame refresh (AIR). The marker localizes transmission communicate one another. As its name indicates, MBMS
error by inserting code to mitigate errors. AIR prevents error has two modes in practice: broadcast mode and multicast
propagation by frequently performing intra-frame coding to mode.
motion domains. However, AIR is not effective in combating A broadcast service can be generalized to mean a unidi-
error propagation when I-frames are less frequent. rectional point-to-multipoint service in which data is trans-
A survey of error resilient techniques for multicast appli- mitted from a single source to multiple terminals in the as-
cations for IP-based networks is reported in [13]. It presents sociated broadcast service area. On the other hand, a mul-
algorithms that combine ARQ, FEC, and local recovery tech- ticast service can be defined as a unidirectional point-to-
niques where the retransmissions are conducted by multicast multipoint service in which data is transmitted from a sin-
group members or intermediate nodes in the multicast tree. gle source to a multicast group in the associated multicast
K. E. Psannis and Y. Ishibashi 3
service area. Only the users that are subscribed to the spe- Streaming applications Download applications
cific multicast service and have jointed the multicast group
3GPP file Service
associated with the service can receive the multicast services. RTP playload (codec)
As a difference a broadcast service can be received without download announcement
separate indication from the customers. In practice multicast FLUTE
users need a return channel for the interaction procedures in RTP ALC/FEC
order to be able to subscribe to the desired services.
Similar to (PSS) and (MMS), two type applications of LCT
(MBMS) standard are anticipated. UDP
(i) MBMS download: to push a multimedia message to IP-multicast
clients. Multimedia broadcast/multicast service (MBMS) bearer(s)
(ii) MBMS streaming: continuous media stream transmis-
sion and immediate playout.
Figure 1: Protocol stack view of MBMS.
The protocol stack is designed to accommodate the above ap-
plications as illustrated in Figure 1.
The streaming stack is very similar to PSS [18]. On the When a mobile joints an existing multicast session, there
other hand, the download stack is unique in terms of its is a delay before which it can be synchronized. This delay is
adoption of IETF reliable multicast/broadcast delivery in proportional to the frequency of I-frames as determined by
error-prone environments. As protocol, FLUTE is fully spec- the streaming server. Since I-frames require more bits than
ified and built on top of the asynchronous layered coding the P- and B-frames, the compression efficiency is inversely
(ALC) protocol of the layered coding transport (LCT) build- to the frequency of I-frames. Assume that Inumber is the fre-
ing block. File transfer is administrated by special-purpose quency of I-frames, Frate is the frame rate of the video com-
objects, file description table (FDT) instances, which provide pression. The worse case initial delay in seconds can be com-
a running index of files and their essential reception parame- puted as follows:
ters in-band of a FLUTE session. ALC is the adaption proto-
col to extend LCT for multicast. ALC combines the LCT and 1 1
delay = − , (1)
FEC building blocks. LCT is designed as a layered multicast Inumber Frate
transport protocol for massively scalable, reliable, and asyn-
chronous content delivery. An LCT session comprises multi- where
ple channels originating at a single sender that are used for Frate
some period of time to carry packets pertaining to the trans- Inumber = . (2)
N
mission of one or more objects that can be of interest to re-
ceivers. The FEC building block is optionally used together N is the distance between two successive I-frames defin-
with the LCT building block to provide reliability. The FEC ing a “group of pictures” (GoP). N can be defined as follows:
building block allows the choice of an appropriate FEC (e.g., ⎧
Reed-Solomon) code to be used with ALS, including using ⎪
⎪
⎪α × M, M > 0, α > 0, I- P- B-frames,
⎪
⎪
the no-code FEC scheme that simply sends the original data ⎪
⎪
⎨N = α M = 1, α > 0, I- P-frames,
using no FEC coding [7]. N =⎪ (3)
⎪N > 0, M = 0,
⎪
⎪
I-frames,
⎪
⎪
⎪
⎩M, N = M > 0, I- B-frames,
3. PROBLEM FORMULATION
The MBMS system introduces a new paradigm from the tra- where M is the distance between two successive P-frames.
ditional internet- or satellite-based multicasting system due (usually set to 3) and α is nonnegative constant (α ≥ 0).
to mobility. The system has to account for wide variety of Figure 2 depicts the worse delay in seconds for different
receiver conditions such as handover, speed of the receiver, combination of frame rate and the number of I-frames in a
interference, and fading. Moreover, the required bandwidth Group Of Pictures.
and power should be kept low for mobile devices. The graph in Figure 2 shows that the delay is propor-
Since mobility is expected during session there is typi- tional to the frequency of I-frames. The application would
cally significant packet loss during handover. If the packet also require more frequent transmission of I-frames so as
loss occurs on an I-frame, it would effect all the P- and B- to allow the use to joint the ongoing session. However, this
frames that predict from the I-frame. In the case of P-frames, would result in requiring more bandwidth.
the error concealment techniques could mitigate the loss; Assuming that the ratio of frame sizes for I-, P-, and B-
however, the distortion would continue to propagate until an frames is 5 : 3 : 2, the MPEG bitstream used for simulation
I-frame is found. These could also be managed using intra- is the “Mobile” sequence with 180 frames. The average band-
block refresh rate technique. On the other hand, loss of B- width is given by [20]
frames limits the loss to that particular frame and does not
result in error propagation. bandwidth = Frate × average (IP) size × 8 bits/byte, (4)
0.5 One other tool that is effective against error propagation

0.45 is intra-block refresh technique [5]. In this technique, a per-
centage of P- and B-frames block is intra-coded and crite-
0.4
rion for determining such intra-clock is dependent on the
0.35 algorithm. However, the intra-block refresh technique is not
0.3 effective in combating error propagation when I-frames are
Delay (s)
less frequent.
0.25
Apart from traditional broadcasting/multicasting tech-
0.2 niques, the MBMS system requires new technologies for er-
0.15 ror resilience. This is because MBMS does not allow re-
transmissions and the temporal fading conditions of wireless
0.1 channels could result in corruption of certain frames. Due to
0.05 the frame dependency within hybrid coding techniques the
errors propagate until an I-frame is decoded.
0
0 2 4 6 8 10 12 14 16
Group of pictures (N) 4. PROPOSED TECHNIQUE
Frate = 10 fps
In a typical video distribution scenario, video content is cap-
Frate = 20 fps
tured, then immediately compressed and stored on a local
Frate = 30 fps
network. At this stage, compression efficiency of the video
signal is most important as the content is usually encoded
Figure 2: Relative increase in the delay as a function of GOP (N). with relatively high quality and independent of any actual
channel characteristics. Note that heterogeneity of client net-
works makes it difficult for the encoder to adaptively encode
Mobile sequence (M = 3)
3.5 the video contents to a wide degree of different channel con-
ditions. This is especially true for wireless clients. It should
3 also be noted that the transcoding (decode-(re-) encode) of
stored video is as necessary as that for live video streaming.
2.5 For instance, pre-analysis may be performed on stored video
Bandwidth (Mbps)
to gather useful information. If the server only has the orig-

2
inal compressed bitstream (i.e., the original uncompressed
1.5 sequence is unavailable), we can decode the bitstream.
The problem addressed is that of transmitting a sequence
1 of frames of stored video using the minimum amount of en-
ergy subject to video quality and bandwidth constraints im-
0.5
pose by the wireless network.
Assume that I-frame is always the start point of a joining
0
0 5 10 15 20 multicast session. Since I-frames are decoded independently,
Group of pictures (N) switching from leaving to joining multicast session can been
done very efficiently The corresponding video streams are
Frate = 30 fps
obtained by encoding the original uncompressed video file
Frate = 20 fps
Frate = 10 fps
as a sequence of I-P(I)-frames using a different GOP pattern
(N = 5, M = 1).
P(I) are coded using motion estimation and each one
Figure 3: Relative increase in the bandwidth as a function of GOP
(N).
has a dependency only on the preceding I-frame. This re-
sults in that the corruption of P-frame does not affect the
next P-frame to be decoded. On the other hand, it increases
the P(I)-frame sizes.
where We consider a system where source coding decisions are
average (IPB) size made using the minimum amount of energy min Eq(i) {I,
P(I)} subject to minimum distortion (Dmin ) at the mobile
Iaverage 1 1 1
= + Paverage x − + Baverage x 1 − . client and the available channel rate (C Rate) required by
N M N M wireless network. Hence
(5)
min Eq(i) I, P(I) ≤ C Rate,
Figure 3 shows the increase in the network bandwidth as (6)
min Eq(i) I, P(I) ≤ Dmin .
a function of group of pictures. It can be seen from this graph
that more I-frames in a GOP results in increase in the net- It should be emphasized that a major limitation in wireless
work bandwidth. networks is that mobile users must rely on a battery with a
limited supply of energy. Effectively utilizing this energy is

a key consideration in the design of wireless networks. Our 80
goal is to properly select a quantizer q(i) in order to mini-
mize the energy required to transmit the sequence of I-P(I)-
frames subject to both distortion and channel constraints. 60
A common approach to control the size of an MPEG
frame is to vary the quantization factor on a per-frame ba-
40
sis [21]. The amount of quantization may be varied. This is
the mechanism that provides constant quality rate control.
The quantized coefficients QF[u, v] are computed from the 20
DCT coefficients F[u, v], the quantization scale, MQUANT,
and a quantization matrix, W[u, v], according to the follow-
0
ing equation:
16 × F[u, v]
QF[u, v] = . (7) (a) W[u, v] with low slope
MQUANT × W[u, v]
The normalized quantization factor w[u, v] is
MQUANT × W[u, v] 80
w[u, v] = . (8)
16
The quantization step makes many of the values in the coef-
60
ficient matrix zero, and it makes the rest smaller. The result is
a significant reduction in the number of coded bits with no
visually apparent difference between the decoded output and 40
the original source data [22]. The quantization factor may be
varied in two ways.
20
(i) Varying the quantization scale (MQUANT).
(ii) Varying the quantization matrix (W[u, v]).
0
To bound the size of predicted frames, an P(I)-frame is en-
coded such that its size fulfills the following constraints:
(b) W[u, v] with high slope

BitBudget I, P(I) ≤ C Rate,
(9)
BitBudget I, P(I) ≤ Dmin . Figure 4: Two normalized quantization matrices w[u, v] both with
The encoding algorithm in the first encoding attempt starts MQUANT = 8. (a) W[u, v] with low slope; (b) W[u, v] with high
slope.
with the nominal quantization value that was used to encode
the preceding I-frame. After the first encoding attempt, if the
resulting frame size fulfills the constraints (9), the encoder
proceeds to the next frame. Otherwise, the quantization fac- do subjective rating because it is not mathematically repeat-
tor (quantization matrix, W[u, v]) varies and the same frame able. For this reason we measure the visual quality of the in-
is re-encoded. teractive mode using the peak signal-to-noise ratio (PSNR).
The quantization matrix can be modified by maintaining We use the PSNR of the Y -component of a decoded frame.
the same value at the near-dc coefficients but with different The MPEG-4 bitstream used for simulation is the
slope towards the higher frequency coefficients. This proce- “Mobile” sequence with 180 frames, with a frame rate of
dure is repeated until the size of the compressed frame cor- 30 fps. The GOP format was N = 5, M = 1. We con-
responds to (9). The advantage of this scheme is that it tries sider a set of allowable channel rate, C Rate = (300 kbps,
to minimize the fluctuation in video quality while satisfying 200 kbps, 100 kbps). In order to illustrate the advantage of
channel condition. the proposed algorithm we consider a system where source
Figure 4 shows two matrices both with the same value coding decisions are mode without any constraints, using
at the near-dc coefficients but with different slope towards the same GOP format (N = 5, M = 1). Figure 5 shows the
the higher frequency coefficients. In other words, the quan- PSNR plot per frame obtained with the proposed algorithm
tization scale is fixed MQUANT and the quantization matrix and the reference scheme for the allowable channel rate.
W[u, v] varies. Clearly the proposed algorithm yields an advantage of
(PSNR) 1.42 dB, 1.39 dB, and 1.35 dB, for the allowable
5. SIMULATIONS RESULTS channel rates 300 kbps, 200 kbps, and 100 kbps, respectively.
Figure 6 depicts the performance of the proposed algorithm
There are two types of criteria that can be used for the evalua- compared with the performance of MPEG-4 simple profile
tion of video quality; subjective and objective. It is difficult to [5] codec during frame loss. The percentage of frames that
Mobile sequence overcome these additional resources we proposed the use of

35
differently encoded version of each video sequence. The dif-
34 ferently coded sequences are obtained by encoding frames of
the original (uncompressed) sequence as I-P(I)-frames us-
33
ing a different GOP pattern. The server responds to a mul-
PSNR (dB)
32 ticast request by switching from leaving to joining multicast

session very efficiently. By proper encoding versions of the
31 original video sequence, multicast video streaming services
30 can be supported with considerably reduced additional delay
and minimum jitter which implies acceptable visual quality
29 at the wireless client-end. Our future work includes develop-
0 20 40 60 80 100 120 140 160 180
ing a multilevel quality of services framework for wireless full
Frames
interactive multicast video services.
Proposed algorithm, C Rate = 4300 kbps
Proposed algorithm, C Rate = 200 kbps ACKNOWLEDGMENT
Proposed algorithm, C Rate = 100 kbps
Without constraints This paper was supported in part by International Informa-
tion Science Foundation (IISF), Japan (Grant no 2006.1.3.
Figure 5: PSNR for encoded frames in the multicast version. 916).
40 REFERENCES
35
[1] M. Etoh and T. Yoshimura, “Advances in wireless video deliv-
30 ery,” Proceedings of the IEEE, vol. 93, no. 1, pp. 111–122, 2005.
[2] K. E. Psannis and M. Hadjinicolaou, “MPEG based interactive
25
PSNR (dB)
video streaming: a review,” WSEAS Transaction on Communi-

20 cation, vol. 2, pp. 113–120, 2004.
[3] Coding of Moving Pictures and Associated Audio for Digital
15
Storage Media at up to About 1.5Mbits/s, (MPEG-1), October
10 1993.
5
[4] Generic Coding of Moving Pictures and Associated Audio.
ISO/IEC 13818-2, (MPEG-2), November 1993.
0 [5] Coding of Moving Pictures and Associated Audio.MPEG98/
0 20 40 60 80 W21994, (MPEG-4), March 1998.
Percentage of dropped frames (%) [6] T. Stockhammer and M. M. Hannuksela, “H.264/AVC video
Proposed algorithm
for wireless transmission,” IEEE Wireless Communications,
MPEG-4 simple profile, (I)-VOP period = 5
vol. 12, no. 4, pp. 6–13, 2005.
MPEG-4 simple profile, (I)-VOP period = 10
[7] 3GPP, “S-CCPCH Performance for MBMS,” Technical Speci-
MPEG-4 simple profile, (I)-VOP period = 15 fication Group Radio Access Network TR 25.803 v.1.3.0, 3rd
Generation Partnership Project (3GPP), Frankfurt, Germany,
March 2004.
Figure 6: PSNR as function of frames dropped. [8] R. Swann and N. Kingsbury, “Transcoding of MPEG-II for
enhanced resilience to transmission errors,” in Proceedings of
IEEE International Conference on Image Processing (ICIP ’96),
vol. 2, pp. 813–816, Lausanne, Switzerland, September 1996.
are dropped is varied and it is clearly seen that the proposed [9] G. De Los Reyes, A. R. Reibman, S.-F. Chang, and J. C.-I.
approach maintains the quality. On the other hand, in the Chuang, “Error-resilient transcoding for video over wireless
MPEG-4 simple profile codec, the quality degrades with the channels,” IEEE Journal on Selected Areas in Communications,
increase in frame loss percentage. The “Mobile” sequence vol. 18, no. 6, pp. 1063–1074, 2000.
was used for these experiments with bit rate 100 kbps and [10] S. Dogan, A. Cellatoglu, M. Uyguroglu, A. H. Sadka, and A.
frame rate 30 fps. The above figures show that the proposed M. Kondoz, “Error-resilient video transcoding for robust in-
algorithm minimizes jitter during multicast session. ternetwork communications using GPRS,” IEEE Transactions
on Circuits and Systems for Video Technology, vol. 12, no. 6, pp.
453–464, 2002.
6. CONCLUSIONS [11] G. Côté, S. Shirani, and F. Kossentini, “Optimal mode selec-
tion and synchronization for robust video communications
Error resilient (re-) encoding is a key technique that enables over error-prone networks,” IEEE Journal on Selected Areas in
robust streaming of stored video content over noisy chan- Communications, vol. 18, no. 6, pp. 952–965, 2000.
nels. It is particularly useful when content has been pro- [12] R. Zhang, S. L. Regunathan, and K. Rose, “Video coding with
duced independent of the transmission network conditions. optimal inter/intra-mode switching for packet loss resilience,”
In this paper, we investigated the constraints of supporting IEEE Journal on Selected Areas in Communications, vol. 18,
multimedia multicast services in wireless clients. In order to no. 6, pp. 966–976, 2000.
[13] G. Carle and E. W. Biersack, “Survey of error recovery tech- University, UK. From 2001 to 2002 he was awarded the British
niques for IP-based audio-visual multicast applications,” IEEE Chevening scholarship sponsored by the British Government. He
Network, vol. 11, no. 6, pp. 24–36, 1997. has more than 40 publications in Conferences and peer-reviewed
[14] P. Ge and P. K. McKinley, “Experimental evaluation of error journals. He received a degree in physics from Aristotle University
control for video multicast over wireless LANs,” in Proceed- of Thessaloniki (Greece), and the Ph.D. degree from the Depart-
ings of the 21st International Conference on Distributed Com- ment of Electronic and Computer Engineering of Brunel University
puting Systems Workshop (ICDCSW ’01), pp. 301–306, Mesa, (UK). He is a Member of the IEEE, ACM, IEE, and WSEAS.
Ariz, USA, April 2001.
Yutaka Ishibashi received the B.S., M.S.,
[15] P. Ge and P. K. McKinley, “Comparisons of error control tech-
and Ph.D. degrees from Nagoya Institute of
niques for wireless video multicasting,” in Proceedings of 21st
Technology, Nagoya, Japan, in 1981, 1983,
IEEE International Performance, Computing and Communica-
and 1990, respectively. From 1983 to 1993,
tions Conference, pp. 93–102, Phoenix, Ariz, USA, April 2002.
he was with NTT Laboratories. In 1993, as
[16] K. E. Psannis, M. G. Hadjinicolaou, and A. Krikelis, “MPEG-
an Associate Professor, he joined Nagoya In-
2 streaming of full interactive content,” IEEE Transactions on
stitute of Technology, in which he is now a
Circuits and Systems for Video Technology, vol. 16, no. 2, pp.
Professor in the Department of Computer
280–285, 2006.
Science and Engineering, Graduate School
[17] 3GPP, “Codec for Circuits Switched Multimedia Telephone
of Engineering. From June 2000 to March 2001, he was a Visiting
Service; General Description,” Technical Specification Group
Professor in the Department of Computer Science and Engineer-
Services and Systems Aspects TS 26.110, 3rd Generation Part-
ing at the University of South Florida. His research interests in-
nership Project (3GPP), Valbonne, France, June 2002.
clude networked multimedia applications, media synchronization
[18] 3GPP, “Transparent End-to-End Packet-switched stream-
algorithms, and QoS control. He is a Member of the IEEE, ACM,
ing service stage 1,” Technical Specification Group Services
Information Processing Society of Japan, the Institute of Image In-
and Systems Aspects TS 22.233, 3rd Generation Partnership
formation and Television Engineers, and the Virtual Reality Society
Project (3GPP), Valbonne, France, March 2002.
of Japan.
[19] 3GPP, “Multimedia Messaging Service (MMS); Stage 1,” Tech-
nical Specification Group Services and Systems Aspects TS
22.140, 3rd Generation Partnership Project (3GPP), Valbonne,
France, December 2002.
[20] K. E. Psannis and M. Hadjinicolaou, “Transmitting additional
data of MPEG-2 compressed video to support interactive op-
erations,” in Proceedings of International Symposium on Intelli-
gent Multimedia, Video and Speech Processing (ISIMP ’01), pp.
308–311, Hong Kong, May 2001.
[21] K. E. Psannis, Y. Ishibashi, and M. Hadjinicolaou, “A novel
method for supporting full interactive media stream over IP
network,” International Journal on Graphics, Vision and Image
Processing, vol. 5, no. 4, pp. 25–31, 2005.
[22] K. E. Psannis, M. Hadjinicolaou, and A. Krikelis, “Full interac-
tive functions in MPEG-based video on demand systems,” in
Recent Advances in Circuits, Systems and Signal Processing Book
Series, N. E. Mastorakis and G. Antoniou, Eds., pp. 419–424,
WSEAS Press, Rethimno, Greece, 2002.
Kostas E. Psannis was born in Thessaloniki,

Greece. He was awarded, in the year 2006,
a research grant by International Informa-
tion Science Foundation sponsored by Min-
istry of Education, Science, and Technol-
ogy, Japan. Since 2004 he has been a (Vis-
iting) Assistant Professor in the Depart-
ment of Technology Management, Univer-
sity of Macedonia, Greece. Since 2005 he
has been a (Visiting) Assistant Professor
in the Department of Computer Engineering and Telecommu-
nications, University of Western Macedonia, Greece. From 2002
to 2004, he was a Visiting Postdoctoral Research Fellow in the
Department of Computer Science and Engineering, Graduate
School of Engineering, Nagoya Institute of Technology, Japan.
He has extensive research, development, and consulting experi-
ence in the area of telecommunications technologies. Since 1999,
he has participated in several R & D funded projects as a Re-
search Assistant in the Department of Electronic and Com-
puter Engineering, School of Engineering and Design, Brunel
DOI 10.1155/WCN/2006/62657
Cross-Layer Quality-of-Service Analysis and Call Admission

Control in the Uplink of CDMA Cellular Networks
Chun Nie,1, 2 Yong Huat Chew,1 and David Tung Chong Wong1
1 Institute for Infocomm Research, Agency for Science, Technology, and Research, Singapore 119613
2 Department of Electrical and Computer Engineering, National University of Singapore, Singapore 117576
Received 26 September 2005; Revised 16 March 2006; Accepted 26 May 2006

This paper addresses cross-layer quality-of-service (QoS) provisioning in the uplink of CDMA cellular mobile networks. Each
mobile can take up to four UMTS traffic classes in our model. At the data link layer and the network layer, the QoS performances
are defined in terms of signal-to-interference-plus-noise ratio and outage probability, and packet loss rate and delay, respectively.
A call admission control scheme which fulfills these QoS metrics is developed to maximize the system capacity. The novelty of
this paper is that the effect of the lengthening of the on-periods of non-real-time traffic classes is investigated by using the Go-
Back-N automatic retransmission request mechanism with finite buffer size and limited number of retransmissions in the event of
transmission errors. Simulation results for a specific example demonstrate the reasonableness of the analytical formulation.
Copyright © 2006 Chun Nie et al. This is an open access article distributed under the Creative Commons Attribution License,
1. INTRODUCTION can be directly applied to the QoS provisioning of practical

systems. For example, only voice and data services in single-
The currently deployed universal mobile telecommunica- cell systems are considered in [7] and traffic sources are sim-
tions system (UMTS) network is characterized by its abil- ply modeled as exponential-on/exponential-off and Poisson
ity to support multimedia communications with different bit arrivals. Reference [8] investigated packet loss rate and delay
rates and quality-of-service (QoS) requirements. Four traf- performances in CDMA networks for voice, video, and data
fic classes, conversational, streaming, interactive, and back- services. However, analytical QoS formulation is given only
ground, are defined in the UMTS QoS architecture together for voice services, while video and data services are only ob-
with their respective QoS requirements [1]. Code division tained through computer simulations.
multiple access (CDMA) is the multiple access technology The main contribution of this paper is an analytical for-
used to support the transmissions of multiclass services. In mulation for the QoS performances of all of the four traffic
this paper, voice, video, web-browsing, and data are used as classes jointly at both the data link and network layers. We
typical applications of these four traffic classes. Their QoS adopt more realistic traffic models for both real-time (RT)
performances in the uplink are investigated and their QoS and non-real-time (NRT) traffic than those in the literature.
metrics are formulated at both the data link layer and the The effect of the lengthening of the on-periods of the NRT
packet level of the network layer. services is analyzed under Go-Back-N (GBN) automatic re-
In the literature, QoS provisioning in CDMA networks transmission request (ARQ) scheme. The QoS attributes are
has attracted a lot of research interests. At the data link layer, formulated in terms of the signal-to-interference-plus-noise
Gilhousen et al. [2] studied the outage probability for a sin- ratio (SINR) and outage probability at the data link layer, and
gle class on/off source in CDMA networks. Wong et al. [3–5] the average delay and packet loss rate at the network layer. A
extended the analysis of outage probability from a single class QoS-based call admission control (CAC) scheme is also pro-
sources to on/off multiclass sources, variable bit rate multi- posed. The maximum system capacity satisfying all QoS re-
class sources, and video multiclass sources. Recently, the out- quirements at both the data link and network layers is com-
age probabilities of multiclass multiconnection services are puted analytically.
investigated in [6]. At the network layer, packet loss rate and The subsequent sections of this paper are organized as
delay performances are studied for CDMA systems [7, 8]. follows. Section 2 develops a system model that describes a
However, [7, 8] do not provide an analytical platform which cellular mobile network and establishes appropriate traffic
models for the four traffic classes. In Section 3, an efficient corresponding location parameters for web-browsing (k = 3)
power control method is designed and the outage prob- and data (k = 4) services, respectively. The location and
abilities at the data link layer are formulated accordingly. shape parameters are defined in [12].
Section 4 deals with the packet level QoS performances. For a Pareto-on/Pareto-off process, the activity factors of
Section 5 presents analytical and simulation results to verify web-browsing and data traffic at their source can still be ap-
the reasonableness of the analysis. Section 6 develops a CAC proximately defined by pk , k ∈ {3, 4}, as
scheme with cross-layer QoS satisfactions. Finally, Section 7
concludes this paper.
t on,k
pk = , k ∈ {3, 4}, (4)
t on,k + t off,k
2. SYSTEM MODEL
A cellular mobile system with multiple square cells is consid- where t on,k and t off,k are the means of ton,k and toff,k , re-
ered. This model is commonly adopted and referred to as the spectively. The reasonableness of this assumption is verified
Manhattan model [9]. A base station (BS) is located at the through simulations in [13], at least for these parameters
center of each cell to serve a number of mobiles. Each mobile whose ranges are around the values specified in the 3GPP
supports multiconnection to transmit multiclass services. specification [1].
The type of traffic classes is denoted by an index k, where The assumptions and system parameters used are listed
k = 1 for voice, k = 2 for video, k = 3 for web-browsing, as follow.
and k = 4 for data, respectively. In order to evaluate the
(i) There exist N mobiles in each cell and they are uni-
QoS performances, appropriate traffic models are defined.
formly located in the cell.
Voice and video are, respectively, modeled as an exponential-
on/exponential-off process and a two-dimensional discrete- (ii) The area of a cell is denoted by A and the cellular net-
state continuous-time Markov chain, as shown in Figures work comprises of n square cells.
1(a) and 1(b). (iii) ni,k denotes the number of voice, video, web-browsing,
In Figure 1(a), a voice service is modeled as a two-state and data streams of the ith (1 ≤ i ≤ N) mobile, for
on/off birth-death process. In Figure 1(b), a video service k ∈ {1, 2, 3, 4}, respectively.
(k = 2) is a variable bit rate source and is described by (iv) Gk , γk∗ , and BER∗k , k ∈ {1, 2l, 2h, 3, 4}, denote the
the Sen’s model [10]. Each video service can be decomposed spreading gains, SINR, and bit-error-rate (BER) re-
into one high-bit-rate (HBR) and M low-bit-rate (LBR) min- quirements for voice, LBR video, HBR video, web-
isources. Hereafter, one HBR mini-source (k = 2h) and M browsing, and data services, respectively.
LBR minisources (k = 2l) will be used to replace a video ser- (v) Si,k and li,k , k ∈ {1, 2l, 2h, 3, 4}, 1 ≤ i ≤ N, denote
vice. The activity factors, which are the probabilities that the the received power and total number of active spread-
process stays in the on state, for the voice, LBR video, and ing codes used by voice, LBR video, HBR video, web-
HBR video are, respectively, given by browsing, and data services of the ith mobile, respec-
tively.
αk (vi) Perfect power control is implemented for each ser-
pk = , k ∈ {1, 2l, 2h}, (1) vice/minisource to ensure that the desired received
αk + βk
powers are achieved at the intracell BS.
(vii) All receivers have additive white Gaussian noise
where 1/βk and 1/αk are, respectively, the average on and off (AWGN) with power η.
periods, and k = 1 for voice, k = 2l for LBR video min-
(viii) Iintercell is the intercell interference from all neighboring
isources, and k = 2h for HBR video minisources, respec-
cells.
tively.
The source traffic of web-browsing and data services are (ix) GBN ARQ has limited number of retransmissions for
more accurately modeled as a Pareto-on/Pareto-off process web-browsing and data services.
[11]. Let us denote the on and off periods of web-browsing (x) Web-browsing and data services are equipped with fi-
and data by ton,k and toff,k , k ∈ {3, 4}, respectively. The prob- nite buffer of buffer sizes B3 and B4 , respectively, both
ability density functions (pdf) of ton,k and toff,k , k ∈ {3, 4}, in unit of packets.
denoted by uk (ton,k ) and vk (toff,k ), k ∈ {3, 4}, respectively, Voice and video services carry RT traffic and thus are
are given by [12] not very relevant to implement ARQ mechanism. Compara-
tively, web-browsing and data services carry NRT traffic and
−1 thus can initiate the GBN ARQ scheme in case of packet er-
uk ton,k = con,k aon,k con,k ton,k −con,k , ton,k ≥ aon,k , (2)
rors. Since GBN ARQ is a continuous retransmission scheme,
−1 web-browsing/data traffic observed in the channel is still an
vk toff,k = coff,k aoff,k coff,k toff,k −coff,k , toff,k ≥ aoff,k . (3)
on/off process except that the on-period observed in the
channels is lengthened as a result of retransmissions. This re-
In (2) and (3), con,k and coff,k represent the shape parameters sults in larger activity factors being observed in the channels
of the on and off periods, while aon,k and aoff,k represent the than those in the sources.
Chun Nie et al. 3
Mα2 (M 1)α2 α2
(0, 0) (0,1) (0, M)
β2 2β2 Mβ2
αk λ2 μ2 λ2 μ2 λ2 μ2
On Mα2 (M 1)α2 α2
Off
(1, 0) (1, 1) (1, M)
βk β2 2β2 Mβ2
(a) (b)
Figure 1: Traffic models: (a) 2-state Markov chain for a voice source, (b) 2-dimensional Markov chain for a video source.
Since each mobile experiences different amount of inter- large [2–6], with mean and variance given by
ference and retransmissions, the lengthened activity factors

N
rm dA
of each mobile can be different even for the same class of I intercell ≤ pi,k,c ni,k Si,k f ,
service. Let us denote the average on and off periods of web- i=1 k∈V
rd A
browsing and data services in the CDMA channel as t on,k,c
and t off,k,c , k ∈ {3, 4}, respectively, where the subscript c is Var Iintercell
used to represent the channel, obviously, t on,k,c > t on,k and

t off,k,c < t off,k . Let pi,k,c , 1 ≤ i ≤ N, k ∈ {1, 2l, 2h, 3, 4}, de-
N rm r dA
≤ S2i,k ni,k pi,k,c g 2
− pi,k,c f2 m
note the lengthened activity factors of voice, LBR video, HBR i=1 k∈V
rd rd A
video, web-browsing, and data services of the ith user in the
channel, respectively. pi,k,c = pk for k = 1, 2l, 2h as there is

+ S2i,2l ni,2 M pi,2l,c 1 + (M − 1)pi,2l,c
no retransmission scheme and pi,k,c > pk for k = 3, 4 as these
services use GBN ARQ scheme.
rm 2 rm dA
×g − M pi,2l,c f 2 ,
rd rd A
3. POWER CONTROL ALGORITHM AND QoS ANALYSIS (6)
AT DATA LINK LAYER
where
System capacity and QoS performance metrics in CDMA
r
networks are associated with the multiple access interference f m
rd
(MAI) contributed from the interfering mobiles. MAI in-
4
cludes both intracell and intercell interference resulting from rm (σ ln 10/10)2 40 log rm /rd ln 10
= e 1−Q √ − 2σ 2 ,
mobiles within and outside the reference cell. SINR is a func- rd 2σ 2 10
tion of the received powers, spreading gains and number of
active spreading codes, and is an important attribute at the rm
g
data link layer. It is necessary that the average SINR of each rd
service should be maintained at a required level. Denote set 8
rm 2 40 log rm /rd ln 10
V as {1, 2l, 2h, 3, 4}, V as {1, 2h, 3, 4}, ni,2l = Mni,2 , and = e(σ ln 10/5) 1 − Q √ − 2σ 2 .
ni,2h = ni,2 (1 ≤ i ≤ N) for the ith mobile, the average SINR rd 2σ 2 5
of the kth service stream can be expressed as [6] (7)
Si,k Gk σ2
In (7), is variance of the lognormal shadowing, rm and rd
= γk∗ , (5) denote the distance between a mobile and its own intracell
N
j =1; j
=i k∈V pi,k,c ni,k Si,k + I intercell + η BS, and the distance between the mobile and the intercell BS,
respectively. Let
where k ∈ V and i ∈ {1, 2, . . . , N }. I intercell denotes the mean
of the intercell interference. Our path loss model includes k∈V pi,k,c ni,k γk∗
Γi = ,
only path attenuation and lognormal shadowing which has Gk
been widely adopted [2–6]. Rayleigh and Ricean fading are N (8)
ignored. The total intercell interference is approximated by a 1− i=1 Γi 1 + f rm /rd dA/A
= .
Gaussian distribution if the number of mobiles is sufficiently 1 + Γi
According to the formulation that is presented in [6], the fol- infinite buffer is considered and thus is not realistic. In the
lowing power level is derived: following, we will investigate and provide the analytical plat-
form on the effect of a finite buffer on the packet loss rate and
ηγ∗j the average delay of a Pareto-on/Pareto-off distributed NRT
Si, j = , 1 ≤ i ≤ N, j = {1, 2l, 2h, 3, 4}. (9) traffic for CDMA systems.
1 + Γi G j
The data link layer QoS performance is analyzed in terms 4.1. Go-Back-N ARQ
of the outage probability, which refers to the probability that
the achieved SINR is below the SINR requirement or the Compared to the stop-and-wait ARQ, GBN is more efficient
achieved BER is above the BER requirement. Within the ith and easy to implement. Furthermore, it guarantees that the
mobile, the outage probabilities for voice, LBR video, HBR received packets are in sequence as compared to the selec-
video, web-browsing, and data services are formulated as tive repeat ARQ. Figure 2(a) gives a good illustration on the
Pout,i,k , 1 ≤ i ≤ N, k ∈ {1, 2l, 2h, 3, 4}, and given by [6] mechanism of GBN ARQ. At the source, the mobile has a fi-
nite buffer to accommodate the newly arrived packets. When
the first and subsequent few packets arrived, they are queued
→
→ δ − μi
Pout,i,k = ×Q
i,k
, (10) in the buffer and at the same time transmitted over the chan-
N V
σi nel. Upon reception, BS decodes the packet and sends an ac-
knowledgment (ACK if correctly decoded and NACK if is in
N error) back to the mobile. Only if ACK is received, the mobile
where σi2 = Var[Iintercell ], μi = j =1; j k∈V (l j,k S j,k ) +
∗
∞=i −t2 /2 √ will remove that packet from the buffer. In case if NACK is re-
I intercell , δi,k = Si,k Gk /γk − η, Q(x) = x e dt/ 2π, and ceived, both the particular packet and all its subsequent pack-
the notation ets are retransmitted sequentially. BS will ensure that NACK
is not sent for more than a given maximum number. In the
→
→ n1,2
Mn1,2
n1,2
n1,3
n1,4
process of retransmission, new packets continue to arrive and
= are queued in the buffer, as shown in Figure 2(b). There are
N V l1,1 =0 l1,2l =0 l1,2h =0 l1,3 =0 l1,4 =0
two situations where packets will be lost.
n j,1 Mn j,2 n j,2 n j,3 n j,4
(a) Since the buffer size is finite, when there are many re-
··· transmissions, buffer will overflow and newly arrived
l j,1 =0 l j,2l =0 l j,2h =0 l j,3 =0 l j,4 =0
j
=i j
=i j
=i j
=i j
=i
packet will be dropped.
(11) (b) A packet has been retransmitted for the allowable max-
nN,1 MnN,2 nN,2 nN,3 nN,4
imum number of times.
···
lN,1 =0 lN,2l =0 lN,2h =0 lN,3 =0 lN,4 =0 Assuming that k = {3, 4} represents web-browsing and data
services, respectively, the system parameters and assump-

N
n j,k l n −l tions are defined as follows.
× pi,k,c j,k 1 − Pi,k,c j,k j,k .
j =1 k∈V
l j,k (1) A finite buffer with a size of Bk packets, k ∈ {3, 4}, is
j=i
used by a sender.
(2) Each on-period contains lk packets of the same size,
Compared to the results in [6], the main contribution where the total length of the lk packets is a random
here is to calculate the outage probabilities in the environ- variable which follows a pdf that is defined in (2) or
ment with the GBN-ARQ scheme. The computation of the (3). Packets are generated continuously during the on-
lengthened activity factors will be discussed in the next sec- period with a fixed time duration, Tk , k ∈ {3, 4}.
tion. (3) When a packet is transmitted from a mobile to the BS,
the mobile waits for an acknowledgment within a time
4. PACKET LEVEL QoS ANALYSIS AT THE interval of Tk . The packet will be removed from the
NETWORK LAYER buffer upon the receipt of ACK. The ratio of Tk to Tk
is assumed to be an integer, sk , (e.g., sk = Tk /Tk = 2 in
In this section, our aim is to formulate the packet level QoS the example shown in Figure 2(a)) and Bk ≥ sk holds.
performance in the uplink of CDMA systems in terms of (4) Packet error probability is defined as pe,k , k ∈ {3, 4}.
packet loss rates and average delays. Packet level QoS at the (5) ACK and NACK are always received correctly.
network layer is directly associated with the outage proba- (6) Let the maximum number of retransmissions be Mre,k ,
bility. If outage occurs, the packets are assumed erroneous k ∈ {3, 4}.
due to excessive bit errors. For RT voice and video traffic,
these packets are discarded and treated as packet loss. For Next, the following variables, which are useful for our analy-
NRT web-browsing and data traffic, GBN ARQ mechanism is sis, are defined. For simplicity and ease of notations, the sub-
implemented to retransmit the erroneous packets which also script k, which is used to differentiate between the two NRT
result in longer packet delays. In previous works [7, 14, 15], services, will not be shown in the next few subsections. For
Chun Nie et al. 5
Maximum retransmission of
packet 2
3rd retransmission of packet 4
1st retransmission of
1st retransmission of packet 2 packet 4
1 2 3 4 2 3 4 2 3 4 5 6 4 5 6 4 5 6 4 5 6
Time
Mobile station
(sender)
Time
1 2 2 3 4
Base station
(receiver)
Discarded Accepted
Accepted Discarded Discarded
ACK
NACK
(a)
Packet arrivals
1 2 34 5 6
s
1+s
1 2 3 4 2 3 4 5 3 4 5 6 4 5 6 4 5 6 4 5 6 4 5 6 7 5 6 7 8 6 7 l
Transmission time Transmission time Time
of packet 2 of packet 4
Delay of packet 2
Delay of packet 4
Transmission finishing time Transmission finishing time
packet 2 packet 4
Packet removal time Packet removal time
packet 2 packet 4
(b)
9 overflow
A overflow
B overflow
Buffer size = 6
DC8765
EDC876
FEDC87
FEDC87
FEDC87
FEDC87
FEDC8
Buffer
654321
765432
876543
876543
876543
876543
C87654
FEDC
4321
54321
FED
Mre = 2
321
FE
status
1
21
s=2
1 2 3 1 2 3 4 5 3 4 5 6 7 8 C 7 8 C D E F Tx over air
1out
2out
3out
4out
5out
6out
7out
Dout
Eout
Fout
8out
Cout
Ton
Ton, c
(c)
Figure 2: (a) GBN ARQ mechanism, (b) definition of packet transmission and removal time in Go-Back-N ARQ, and (c) lengthening effect
of the on-period: an example.
example, Mre would mean Mre,k , T would mean Tk , and so been removed from the buffer and ω (ω ≤ χ) packets are cor-
on. rectly received. The finite buffer can store a maximum of B
packets, therefore, Nof (l) = max(l − χ − B, 0) denotes the
(1) v is the index used to represent packet sequence ap-
number of overflowed packets (if any) and χ − ω is the num-
pearing in the source, v = 1, . . . , l.
ber of unsuccessful packets which have attempted to retrans-
(2) tin,v denotes the initial transmission time of the vth
mit for Mre times. This is illustrated using Figure 2(c). In this
packet at the mobile at time, tarr,1 = 0.
example, when l = 15th packet arrival, χ = 6 packets (1 to 6)
(3) t f n,v denotes the finishing time of the vth packet at the
have been removed from the buffer. All of these packets have
mobile.
been correctly received eventually, and hence ω = 6. This
(4) trm,v denotes the time when the vth packet is removed
means that l − χ − B = 3 packets (9, A, and B) are lost. Note
from the buffer of the mobile. From definition, this
that after the lth packet, no packets will arrive and hence
will only happen if ACK is received, and hence trm,v =
there will not be any packet overflow.
t f n,v + sT.
Using the relationships that trm,χ ≤ (l − 1)T and trm,χ+1 ≥
(5) Ttr,v = t f n,v − tin,v is the transmission time before the
(l − 1)T, together with the fact that mq ranges from zero to
packet is successfully transmitted.
Mre , the range of χ can be found to be
(6) mv denotes the number of retransmissions for the vth
packet such that mv ≤ Mre .

The definitions of these variables can also be found in l −1−s
χ ≤ l − 1 − s, χ ≥ max − 1, 0 . (13)
Figures 2(b) and 2(c). There are a few interesting relation- 1 + (1 + s)Mre
ships which can be derived if the buffer size is infinite:
⎧
⎪
⎪(v − 1)T, v ≤ s + 1, Similarly, based on the fact that χ − ω packets have been re-
⎪
⎨
v−s−1 transmitted for Mre times, we can obtain
tin,v =
⎪
⎪ −
⎪
⎩ (v 1) + m q (1 + s) T, v > s + 1,
q=1
l −1−s−χ
⎧ (12) χ− ≤ ω ≤ χ. (14)
⎪ v (1 + s)Mre
⎪
⎪
⎪
⎪ 1 + (1 + s) mq T, v ≤ s,
⎨
q=1
Ttr,v =
⎪
⎪ v
⎪
⎪
⎪ s) m Although packet i is transmitted for iq=i−s (1+mq ) times,
⎩ 1 + (1 + q T, v > s.
q=v−s the first iq−=1i−s (1+mq ) is due to the erroneous transmissions
of its previous packets and only the final 1 + mi transmis-
The example shown in Figure 2(c) is used for illustration. sions will determine whether it will be successfully transmit-
Take tin,1 = 0 (referenced, v = 1), then tin,2 = T, tin,3 = 2T ted. Hence, out of ntr ≤ χ + (l − 1 − s − χ)/(1 + s) transmis-
(v = s + 1), packet 1 is not retransmitted, hence m1 = 0, sions associated to the χ packets only ω packets are success-
therefore tin,4 = 3 + 0 = 3, tin,5 = 7 since m2 = 1 and m3 = 0, fully received. The probability that ω packets are correctly re-
and so on. In the following, based on the above definition, ceived out of all the χ removed packets when the lth packet
we are going to derive a few results for finite buffer size. χ
arrive is given by Cω · (1 − peMre +1 )ω (peMre +1 )χ−ω , where Cω
χ
is the binomial coefficient. The probability that there are ω

4.2. The number of overflowed packets correct transmissions in all the ntr transmissions is given by
Cωntr · (1 − pe )ω pentr −ω . Averaging over all possible retransmis-
Assume there are l packets in an observed on-period. When sion and overflow scenarios, the average overflowed packets
the lth packet arrives at the buffer, we assume χ packets have conditional on l are given by
χmax χ ntr ω ntr −ω χ ω χ−ω

χ =χmin ω=ωmin Cω 1 − pe pe Cω 1 − peMre +1 peMre +1 max(l − χ − B, 0)
N of (l) = χmax χ ntr ω ntr −ω χ Mre +1 ω Mre +1 χ −ω , (15)
χ =χmin ω=ωmin Cω 1 − pe pe Cω 1 − p e pe
where ωmin = χ − (l − 1 − s − χ)/ (1 + s)Mre , χmin = max{(l − 4.3. The lengthened activity factor
1 − s)/ 1 + (1 + s)Mre − 1, 0}, χmax = l − 1 − s, and ntr ≤
χ + (l − 1 − s − χ)/(1 + s) can be derived using (13) and (14). Under the assumption of a small retransmission probability,
In (15), the denominator is the normalization factor. the lengthened activity factor in the GBN ARQ system, pon,c ,
Chun Nie et al. 7
can still be approximated by 4.4. Total packet loss

t on,c Packet losses result from both finite buffer overflow and
pon,c = , (16)
t on,c + t off,c retransmissions exceeding the maximum limit. The condi-
where t on + t off = t on,c + t off,c . We first illustrate how t on,c can tional average packet loss conditioned on l is given by
be obtained.
The lengthened on-period is given by t f n,l , that is, the N loss (l) = l − N of (l) peMre +1 + N of (l). (24)
time when it completed the transmission of the lth packet. Then, the mean of the packet loss rate over time is the prob-
Another variable k(l) is defined, where k(l) ≤ l is the abilistic summation of all possible instantaneous packet loss
number of packets transmitted over the channel. In the case rates based on (22) and (24), and thus is given by
when there are overflowed packets, k(l) will exclude these
packets. For the example shown in Figure 2, since there is 3 ∞

p(l)N loss (l)
overflowed packets, k(l = 15) = 12. Mathematically, the Ploss = . (25)
l =aon /T
l
on-period is given by
k(l)
4.5. Average buffer length and delay
ton,c|l = t f n,k − tin,1 = k(l) + (1 + s) mq T. (17)
q=1
The retransmissions are assumed to be minimal so that each
All retransmissions will follow the same statistics. Taking the new on-period arrives with an empty buffer. If an on-period
expectation of (17) with respect to k(l) and m, we have contains l packets, the buffer length shows the following be-
haviors: (a) increase by one if a retransmission is made, (b)
t on,c|l = E k(l) T + (1 + s)E[m]E k(l) T. (18) no change if a transmission or retransmission is successfully,
Using the packet error probability (outage probability) (c) the number of packets in the buffer may reach a max-
pe , the number of retransmissions m is a random variable imum value and stay at this state until the lth packet ar-
with probability given by rives, and (d) the number of packets in the buffer then de-
⎧ ρ−1 creases from the maximum value to zero. Figure 2(c) shows
⎪
⎨ 1 − pe pe , ρ < Mre , the buffer length from t = 0 to 23T which is given by
Pr(m = ρ) = ⎪ [012345666666666666654321] and illustrates this behavior.
⎩ 1 − p peMre + peMre +1 = peMre , ρ = Mre ,
e The buffer is empty after the last packet in the buffer is re-
(19) moved until the arrival of next on-period. In each on/off cy-
cle, the buffer length varies similarly.
and its mean is given by
Assume when the ξth packet arrives, the buffer is getting
pe − peMre +1 full, ξ ≤ l. If there is no overflow, the buffer length condi-
E[m] = . (20) tioned on l can be described by the following function:
1 − pe

Since k(l) = l − Nof (l), average over all retransmission and Qlength t | l
overflow scenarios, ⎧ !
⎪ t
⎪
⎪ − q, l − 1 ≥ trm,q+1 > t ≥ trm,q ,
E k(l) = k(l) = l − N of (l). (21) ⎪
⎪ T
⎪
⎪
⎨l − χ, trm,χ+1 > t ≥ l − 1,
As the on-period is Pareto distributed, the probability =
⎪
⎪
that an on-period has l packets, denoted by p(l), is approx- ⎪
⎪ l − χ − p, trm,χ+p+1 >t ≥ trm,χ+p , l − χ − 1 ≥ p ≥ 1,
⎪
⎪
⎪
⎩
imately given by 0, ton,c + toff,c > t ≥ trm,l ,

(l+1)T (26)
p(l) = Pr{t = lT } = con acon t −con −1 dt, t ≥ aon .
lT
(22) where x is the smallest integer greater than x. χ is the index
of the last removed packet when packet l arrived and defined
Based on (15), (20), and (21), the mean of the lengthened as trm,0 = 0. On the other hand, if there are Nof (l) overflow
on-period of a web-browsing/data service in the GBN ARQ packets, then
system given in (18) can be formulated by
∞ Qlength t | l
pe − peMre +1 (1 + s)
t on,c = p(l) × 1 + ⎧ !
1 − pe ⎪
⎪ t
l =aon /T ⎪
⎪ − q, ξ − 1 > trm,q+1 > t ≥ trm,q ,
(23) ⎪
⎪ T
⎪
⎪
⎪
⎪ trm,χ+Nof (l)+1 ≥ t ≥ ξ − 1,
× l − N of (l) × T , ⎨B,
⎪ (27)
= B − q, trm,χ+Nof (l)+q+1 > t > trm,χ+Nof (l)+q ,
⎪
⎪
where aon is the minimum length of Pareto on-period and ⎪
⎪
⎪
⎪
⎪
⎪ B − 1 ≥ q ≥ 1,
aon /T means the minimum number packets in each Pareto ⎪
⎪
⎪
⎩0,
on-period. ton + toff > t ≥ trm,χ+Nof (l)+B .
These expressions can be verified by looking at the queue is determined by the retransmission and overflow statistics.
length at time t, conditioned by l, in the example, where For example, in Q " length (t | l), N of , ξ, and ξ are used to re-
Trm,1 = 5, Trm,2 = 7, Trm,3 = 11, Trm,4 = 12, . . . , and place Nof , ξ, and χ, respectively. The value of ξ is estimated
Trm,7 = 18, . . . , as shown in Figure 2(c). using the average number of retransmissions as below:
However, there are many possible retransmissions and
packet overflow scenarios (ensemble space) that need to be # $
considered for a given ton,c and toff,c , denoted by ton,c (l) and ξ −s B 1 + E[m](1 + s) − s
ξ− = B =⇒ ξ = ,
toff,c (l). We approximate the ensemble average of Qlength (t | l) 1 + E[m](1 + s) E[m](1 + s)
under all of these scenarios by Q " length (t | l). In Q
" length (t | l), (28)
the transition time of each incremental increase in queue
length as in (27) is replaced by its statistical average, which and χ is estimated by
χmax χ ntr ω ntr −ω χ ω χ−ω

χ =χmin ω=ωmin Cω 1 − pe pe Cω 1 − peMre +1 peMre +1 χ
χ= χmax χ ntr ω ntr −ω χ Mre +1 ω Mre +1 χ −ω . (29)
χ =χmin ω=ωmin Cω 1 − pe pe Cω 1 − p e pe
The average queue length conditioned on l is given by lengthened activity factors of web-browsing and data services
within the ith mobile are supposed to be the summation of all
Q" length t | l dt −−→
Qlength (l) = . (30) probabilistic activity factors over a long time. Let AfFun(Ui,k )
ton,c (l) + toff,c (l) denote instantaneous lengthened activity factor using (16)
−−→
with respect to the parameter set Ui,k . Thus, the lengthened
Furthermore, if the on-period has l packets, the arrival rate
activity factors of web-browsing and data are given by
is assumed to be
→
→
−−→
l − N of (l) pi,k,c = × AfFun Ui,k . (33)
λ(l) = . (31)
ton,c (l) + toff,c (l) N V
Since l is random variable, we want to determine the average It is shown in (5), (9), (10), and (33) that the QoS per-
packet delay over time, denoted as D. Based on (22) and (30)- formances are intertwined across both the data link and net-
(31), D is given by work layers. That is, the outage probabilities, lengthened ac-
tivity factors, packet loss rates, and delays are interrelated
∞
l =aon /T p(l)Qlength (l) with each other. Therefore, an iteration process is developed
D= . (32) to obtain the stable outage probabilities (Pout,i,k , 1 ≤ i ≤ N,
∞
l =aon /T p(l)λ(l) k ∈ V) and the stable lengthened activity factors (pi,k,c ,
1 ≤ i ≤ N, k ∈ {3, 4}), satisfying (5), (9), (10), and (33).
In the discussion given above, one traffic class is con- The steps of the iteration are given as follows.
sidered, and the outage probability is assumed known. In
the following, a more practical situation is considered. The (1) Set initial pi,k,c to be pi,k,c = pk , 1 ≤ i ≤ N, k ∈ V.
fact that multiclass services are present and the performance (2) Calculate Si,k , Pout,i,k , 1 ≤ i ≤ N, k ∈ V, according to
metrics are interdependent, the computation becomes more (9) and (10).
complicated. In general, the computation needs to be per- (3) Based on (33), the new pi,k,c , k ∈ {3, 4}, are calculated.
formed iteratively. (4) With the new pi,k,c , k ∈ {3, 4}, iterate steps 2 and 3
until pi,k,c and Pout,i,k converge.
(5) If convergence occurs, the stable values of Pout,i,k , 1 ≤
4.6. Lengthened activity factor of non-real-time i ≤ N, k ∈ V, and pi,k,c , 1 ≤ i ≤ N, k ∈ {3, 4}, are
service obtained. If it does not converge, it means that there is
no feasible solution jointly satisfying (5), (9), (10), and
In order to facilitate further analysis, let us denote the pa-
(33).
rameter set vector [Tk , Tk , Bk , ck , ak , bk , Q{(δi,k − μi )/σi }, Mk ]
−−→
for the ith mobile as Ui,k , 1 ≤ i ≤ N, k ∈ {3, 4}, respectively.
−−→
Among the vector elements of Ui,k , 1 ≤ i ≤ N, k ∈ {3, 4}, 4.7. Packet level QoS performance at the network layer
Q{(δi,k − μi )/σi }, which is shown in (10), represents the in-
stantaneously outage probabilities of the web-browsing and Based on the above analytical work of the lengthened activity
data services for the ith mobile, respectively. The average factors, the packet loss rate and delay performances of the
Chun Nie et al. 9
Table 1: System parameters.
Parameter type Value Parameter type Value

Shadowing mean μ 0 Number of cells, n 9
Shadowing variance σ 2 σ = 6 dB Thermal noise power η −103.2 dBm (4.8 × 10−14 Watt)
Path loss constant 4
Table 2: Traffic parameter.
Real-time services Non-real-time service

Traffic parameter type
Voice Video Web-browsing Data
Average on-period (second) 1 0.418 (LBR) 1.5 (HBR) 1.6 2.937
Average off-period (second) 1.5 0.663 (LBR) 1.5 (HBR) 12 25.643
Activity factor (source traffic) 0.4 0.3867 (LBR) 0.5 (HBR) 0.1176 0.1028
Average rate (kbps) 24 122.3 14.1 22.8
Channel rate (kbps) 60 30 (LBR) 60 (HBR) 120 240
Spreading gain 64 128 (LBR) 64 (HBR) 32 16
Number of spreading codes 1 8 (LBR) 1 (HBR) 1 1
Buffer size (number of packets) 0 0 200 400
Convolutional rate 1/2 1/2 1/2 1/2
four classes are formulated. Within the ith mobile, let the given by
packet loss rates and delays for voice, LBR video, HBR video,
→
→
web-browsing, and data services be denoted by Ploss,i,k and −−→
Di,k = × DelayFun Ui,k , (37)
Di,k , 1 ≤ i ≤ N, k ∈ {1, 2l, 2h, 3, 4}, respectively. N V
As voice and video are NRT delay-sensitive services, no
ARQ mechanism is implemented in their packet transmis- where 1 ≤ i ≤ N, k ∈ {3, 4}, respectively.
sions. Thus, their packet loss rates are just equal to their
outage probability, which is given by
5. NUMERICAL RESULTS
Ploss,i,k = Pout,i,k , k ∈ {1, 2l, 2h}, (34)
In our analytical model, each mobile can support multicon-
and their average delays are simply their packet transmission nection multiclass traffic. In order to demonstrate the rea-
time, which is given by sonableness of our analytical formulation presented in previ-
ous sections, numerical results are presented in this section.
Di,k = Tk , k ∈ {1, 2l, 2h}. (35) A cellular mobile network with n square cells is considered.
We assume that the number of mobiles with heterogeneous
On the other hand, the lengthened activity factors, av- classes is identical in each cell and all mobiles are uniformly
erage packet loss rates, and average delays of web-browsing distributed. We simulate the network model with SMPL sim-
and data are based on both their instantaneous outage prob- ulation kernel, a type of discrete event simulator [16]. System
abilities and the GBN ARQ mechanism. Let us denote the parameters and traffic parameters are shown in Tables 1 and
average packet loss rates and average delay as Ploss,i,k and Di,k , 2.
1 ≤ i ≤ N, k ∈ {3, 4}, respectively, which are the average Each mobile in our analysis supports up to four diverse
−−→ −−→ classes simultaneously. Suppose that all mobiles in each cell
values over the time. Let PlossFun(Ui,k ) and DelayFun(Ui,k ) can be divided into four groups including different classes.
denote instantaneous packet loss rate and delay using (25) The class distribution and group size are given in Table 3. In
−−→
and (32), respectively, with respect to the parameter set Ui,k . practice, with 4 different traffic classes, there can be up to 15
Therefore, the average packet loss rates of web-browsing and different combinations and similar analytical approach can
data services are given by be applied. We vary the number of users in Group 1 and fix
→
→
the number of users in all the other groups. The numerical
−−→
Ploss,i,k = × PlossFun Ui,k , (36) results are plotted in Figures 3–14.
N V Firstly, we can clearly observe that all analytical results
show better agreements when the systems are in light and
and the average delays of web-browsing and data services are medium loads (less than 1.3 Mbps) than when they are in
Table 3: Number of services in each mobile user.

Group index Group 1 Group 2 Group 3 Group 4
Number of mobiles N1 = 5 ∼ 23 N2 = 2 N3 = 2 N4 = 5
Classes in each mobile 1 voice 1 video 1 voice + 1 video 1 web + 1 data
10 1 10 1
Packet loss rate/outage probability

10 2 10 2
10 3 10 3
10 4 4
10
5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Number of users in Group one Number of users in Group one
Simulation Simulation
Theory Theory
Figure 3: Packet loss rate/outage probability of voice services Figure 4: Packet loss rate/outage probability of video services
(Group 1). (Group 2).
heavy load. The deviation during heavy load, that is, when the analysis still takes much shorter time to work out the re-
there are more mobiles in the system, can be explained as sults than using simulation. For example, it takes more than
follows. The outage becomes more severe and thus retrans- 24 hours to obtain the simulation results, while the analyti-
missions occur more frequently during heavy load. Our GBN cal results can be computed in less than one hour. Therefore,
ARQ analysis is accurate assuming the retransmissions oc- the analytical solution proves to be much more efficient in
cur less frequently and the packet error rate is low. If a lot estimating the QoS performances.
of retransmissions happen under high load, the on-periods
of web-browsing or data services in the CDMA channel may
6. CALL ADMISSION CONTROL METHOD AND
overlap, which influences the computation of their length-
ADMISSION REGION
ened activity factors, outage probabilities, packet loss rates,
and delays. As all classes in CDMA systems are intertwined
In previous literatures, CAC is analyzed with many ap-
with each other, the QoS metrics therefore deviate from sim-
proaches in [17, 18]. But these works are not totally QoS-
ulation results. Therefore, our analytical formulation is more
based and do not address cross-layer CAC in CDMA net-
suitable for light and medium loads when the throughput
works. Our contribution is that the analytical formula-
of the system is below or around 1.3 Mbps. On the other
tion in this paper leads to the determination of the cross-
hand, under higher load, the packet loss rates and delay
layer admission region (AR) in the uplink of a CDMA sys-
performances have already exceeded their specific require-
tem. A QoS-based CAC scheme is given here. If the outage
ments. For example, the packet loss rates requirements of
probability, packet loss rate, and delay requirements are de-
these classes should be less than either 10−2 for voice and
fined as δout , δloss , and δd , the AR at the packet level in the
video or 10−3 for web-browsing and data, which are defined
uplink of CDMA systems, denoted by R, is given by
in [1]. #
Secondly, we also have some comments on the complex- R = (1, 2, 3, . . . , i, . . . , N) | Ploss,i,k
ity of the analysis. Our final analytical expressions are rela- $ (38)
≤ δloss , Di,k ≤ δd , Pout,i,k ≤ δout , SINRI,K = γK∗ ,
tively complex. This is due to the fact that we jointly con-
sider more realistic traffic models, GBN ARQ, multicell net- where 1 ≤ i ≤ N, k ∈ V.
work, and four traffic classes in order to approximate the real Figure 15 shows the CAC scheme in the uplink of CDMA
network. These factors complicate the analysis. Despite this, systems. This CAC scheme admits or rejects call admission
Chun Nie et al. 11
10 1 0.24
0.22
Lengthened activity factor

0.2
10 2
0.18
0.16
10 3
0.14
0.12
10 4 0.1
5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Theory Theory
Figure 5: Packet loss rate/outage probability of voice services Figure 7: Lengthened activity factor of web-browsing services
(Group 3). (Group 4).
10 1
100
10 1
10 2
Outage probability
10 2
10 3
10 3
10 4
10 4
5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Theory Theory
Figure 6: Packet loss rate/outage probability of video services Figure 8: Outage probability of web-browsing services (Group 4).
(Group 3).
requests based on the satisfaction of average SINR require- the outage probabilities of all mobiles and the lengthened
ments and outage probability performance at the data link activity factors of NRT services. Iterations are performed
layer and packet level QoS performances including packet to make both the outage probabilities and the lengthened
loss rate and delay at the network layer. In Figure 15, when activity factors converge. If the iterations cannot reach con-
a specific set of mobile requests to be admitted into the net- vergence, feasible solutions are not available and thus this
work, the CAC process is initiated. The CAC first obtains the combination of mobile users should be rejected by CAC. If
power levels for all mobiles. If positive power solutions are the iterations converge, the stable outage probabilities for all
available, the SINR requirements of these mobiles are sat- services and the lengthened activity factors for NRT services
isfied at the data link layer. Otherwise, the CAC rejects this are obtained. Next, the packet loss rate and average delay of
set of mobiles directly due to their unsatisfactory average each service are calculated. If the obtained outage probabil-
SINR. With the positive power solutions, the CAC computes ity, packet loss rate, and delay requirements are simultane-
100 0.22
0.2
10 1
Lengthened activity factor

0.18
Packet loss rate
10 2
0.16
10 3
0.14
10 4
0.12
10 5 0.1
5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Theory Theory
Figure 9: Packet loss rate of web-browsing services (Group 4). Figure 11: Lengthened activity factor of data services (Group 4).
600 100
500
10 1
Delay (millisecond)
400
Outage probability
300 10 2
200
10 3
100
0 10 4
5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Theory Theory
Figure 10: Average delay of web-browsing services (Group 4). Figure 12: Outage probability of data services (Group 4).
ously satisfied, this set of mobiles can be assured of QoS re- given parameters in Table 1, an example of a 3-dimensional
quirements and thus can be admitted into the network by feasible AR is shown in terms of the number of mobiles in
the CAC scheme. Otherwise, this set of mobiles should be re- Figure 16.
jected by the CAC scheme. Compared to existing CAC meth- In a realistic CDMA system, the BS can utilize dedicated
ods in [17, 18], the main advantage of this CAC scheme is control channels to do fast power control and guarantee that
that it is totally based on the cross-layer QoS satisfaction of all each traffic stream is received with the desired power level.
admitted mobiles in terms of specific SINR, outage probabil- Based on the global information gathered from the network,
ity, packet loss rate, and delay requirements. That is, the QoS the CAC can find out admission region with our analytical
requirements of all admitted mobiles are completely satisfied model in advance and save as a table at the BS. During op-
at both the data link layer and packet level of the network eration, CAC at the BS can simply look up the table to make
layer, and the system capacity is thus maximized. Using the CAC decisions.
Chun Nie et al. 13
100
A set of mobile users
1 Is power No
10
distribution
feasible ?
Packet loss rate
10 2 Yes
Compute outage
3
probability/
10 No lengthened activity
factors
10 4
Not convergable ?
Iterate and
converge ?
10 5
5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 Yes
Number of users in Group one Compute outage
Simulation probability/packet
Theory loss rate/averege
delay
Figure 13: Packet loss rate of data services (Group 4).

Fulfill cross-layer No
QoS requirements ?
1000
Yes
900
To be included into Rejected by call
800
admission region admission control
700
Delay (millisecond)
600
Figure 15: Call admission control procedures.
500
400
300
Number of mobile users in Group four
200
100 15
0
5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
10
Number of users in Group one
Simulation
Theory 5
0
Figure 14: Average delay of data services (Group 4). 0 2 ile
0 4 mo reeb
10 20 6 of p th
30 er u
Numbe
r of mo 40 8
u m n Gro
b
bile use 50 N rs i
60 10
rs in Gro
up one use
7. CONCLUSION
We have presented an approximate analytical framework for Figure 16: Admission region with three groups of users (assume
the cross-layer QoS in CDMA networks. Four classes of ser- N2 = 0).
vices are served within the same mobile and GBN ARQ with
finite buffer size and limited retransmissions is implemented
for NRT traffic with Pareto-on/Pareto-off sources for the first method is proposed to maximize the system capacity and
time. In our analysis, the coupling of packet-level QoS at the leads to the determination of admission region in the up-
network layer and data-link-layer QoS is investigated. The link of CDMA systems. Our analytical work can be further
numerical results show that our analytical approach can ap- combined with the call level analysis of QoS performances
proximate the simulation results quite well up to medium to provide a joint capacity evaluation at both call and packet
traffic load. Based on the cross-layer QoS constraints, a CAC levels in CDMA networks.
REFERENCES [18] L. Zhang, X. Cheng, and K. R. Subramanian, “Admission

control for integrated video-conferencing/voice/data services
[1] 3GPP TS 23.107, “QoS Concept and Architecture,” v5.7.0, De- in broadband CDMA networks,” Computer Communications,
cember 2002. vol. 23, no. 5, pp. 499–510, 2000.
[2] K. S. Gilhousen, I. M. Jacobs, R. Padovani, A. J. Viterbi, L. A.
Weaver Jr., and C. E. Wheatley III, “On the capacity of a cellu-
lar CDMA system,” IEEE Transactions on Vehicular Technology, Chun Nie received the B.Eng. degree
vol. 40, no. 2, pp. 303–312, 1991. from Northwestern Polytechnic University,
[3] D. T. C. Wong, J. W. Mark, K. C. Chua, J. Yao, and Y. H. Chew, China, and the M.Eng. degree from the Na-
“Performance analysis of multiclass services in the uplink of tional University of Singapore, Singapore,
wideband CDMA,” in Proceedings of the 8th International Con- in 2000 and 2005, respectively, all in electri-
ference on Communication Systems (ICCS ’02), vol. 2, pp. 692– cal engineering. He is currently working to-
696, Singapore, November 2002. wards his Ph.D. degree at the Department of
[4] D. T. C. Wong, J. W. Mark, K. C. Chua, and B. Kannan, “Per- Electrical and Computer Engineering, Uni-
formance analysis of variable bit rate multiclass services in the versity of North Carolina, Charlotte, NC,
uplink of wideband CDMA,” in Proceedings of IEEE Interna- USA. His research interests include medium
tional Conference on Communications, vol. 1, pp. 363–367, An- access control, quality-of-service, cross-layer protocol design, and
chorage, Alaska, USA, 2003. resource management in wireless networks.
[5] D. T. C. Wong, J. W. Mark, and K. C. Chua, “Performance eval-
uation of video services in a multirate DS-CDMA system,” in Yong Huat Chew received the B.Eng.,
Proceedings of the 14th IEEE International Symposium on Per- M.Eng., and Ph.D. degrees in electrical en-
sonal, Indoor and Mobile Radio Communications (PIMRC ’03), gineering from the National University of
vol. 2, pp. 1490–1495, Beijing, China, September 2003. Singapore (NUS), Singapore. He has been
[6] C. Nie, D. T. C. Wong, and Y. H. Chew, “Outage analysis for with the Institute for Infocomm Research
multi-connection multiclass services in the uplink of wide- (formerly also known as Centre for Wire-
band CDMA cellular mobile networks,” in Proceedings of the less Communications, NUS, and Institute
3rd IFIP-TC6 Networking Conference, pp. 1426–1432, Athens, for Communications Research), an institute
Greece, May 2004. under the Agency for Science, Technology,
[7] J. B. Kim and M. L. Honig, “Resource allocation for multiple and Research, where he is currently a Senior
classes of DS-CDMA traffic,” IEEE Transactions on Vehicular Scientist, since 1996. He is also an adjunct Associate Professor in
Technology, vol. 49, no. 2, pp. 506–519, 2000. the Department of Electrical and Computer Engineering, National
[8] R. Fantacci and S. Nannicini, “Multiple access protocol University of Singapore. His research interests are in technologies
for integration of variable bit rate multimedia traffic in related to high spectrally efficient wireless communication systems
UMTS/IMT-2000 based on wideband CDMA,” IEEE Journal and radio resource management.
on Selected Areas in Communications, vol. 18, no. 8, pp. 1441–
1454, 2000. David Tung Chong Wong received the
[9] 3GPP TR 25.942, “RF system scenarios,” v.5.3.0, June 2004. B.Eng. and M.Eng. degrees from the Na-
[10] B. Maglaris, D. Anastassiou, P. Sen, G. Karlsson, and J. D. tional University of Singapore (NUS) in
Robbins, “Performance models of statistical multiplexing in 1992 and 1994, respectively, and the Ph.D.
packet video communications,” IEEE Transactions on Commu- degree from the University of Waterloo,
nications, vol. 36, no. 7, pp. 834–844, 1988. Canada, in 1999, all in electrical engineer-
[11] A. Jamalipour, The Wireless Mobile Internet: Architectures, Pro- ing. He is with the Institute for Infocomm
tocols and Services, John Wiley & Sons, Chichester, West Sus- Research, Singapore (formerly Centre for
sex, UK, 2003. Wireless Communications, NUS, and Insti-
[12] M. Evans, Statistical Distributions, John Wiley & Sons, New tute for Communications Research, NUS)
York, NY, USA, 3rd edition, 1998. first as a Research Engineer and currently as a Scientist, since 1994.
[13] C. Nie, “Packet level quality of service analysis of multiclass His research interests are in communications networks, 3 G/4 G,
services in a WCDMA mobile network,” M.S. thesis, National and ultra-wideband wireless mobile multimedia networks. His ar-
University of Singapore, Kent Ridge, Singapore, 2004. eas of research are in the medium access control, resource alloca-
[14] Y. Hayashida, N. Sugimachi, M. Komatsu, and Y. Yoshida, tion with quality-of-service constraints, traffic policing with het-
“Go-back-N system with limited retransmissions,” in Proceed- erogeneous traffic, and cross-layer design. He is a Senior Member
ings of the 8th IEEE Annual International Phoenix Conference of the IEEE. He was on the Technical Program Committee of the
on Computers and Communications, pp. 183–187, Scottsdale, IEEE WCNC 2003, IEEE WCNC 2005, IEEE GLOBECOM 2005,
Ariz, USA, March 1989. and IEEE ICCS 2006.
[15] Y. Hayashida, S. Fujii, and Y. Tezuka, “Delay performance of a
continuous ARQ system with copy-transmissions,” in Proceed-
ings of the 7th Annual Joint Conference of the IEEE Computer
and Communications Societies (INFOCOM ’88), pp. 714–721,
New Orleans, La, USA, 1988.
[16] M. H. MacDougall, Simulating Computer Systems: Techniques
and Tools, MIT Press, Cambridge, Mass, USA, 1987.
[17] B.-J. Hwang and J.-S. Wu, “Call admission control for a mul-
tiple rate CDMA system,” IEICE Transactions on Communica-
tions, vol. E84-B, no. 11, pp. 2932–2945, 2001.
DOI 10.1155/WCN/2006/21297
Cross-Layer Design and Analysis of Downlink Communications

in Cellular CDMA Systems
Jin Yuan Sun, Lian Zhao, and Alagan Anpalagan
Department of Electrical and Computer Engineering, Ryerson University, Toronto, ON, Canada M5B 2K3

A cellular CDMA network with voice and data communications is considered. Focusing on the downlink direction, we seek for
the overall performance improvement which can be achieved by cross-layer analysis and design, taking physical layer, link layer,
network layer, and transport layer into account. We are concerned with the role of each single layer as well as the interaction
among layers, and propose algorithms/schemes accordingly to improve the system performance. These proposals include adaptive
scheduling for link layer, priority-based handoff strategy for network admission control, and an algorithm for the avoidance
of TCP spurious timeouts at the transport layer. Numerical results show the performance gain of each proposed scheme over
independent performance of an individual layer in the wireless mobile network. We conclude that the system performance in terms
of capacity, throughput, dropping probability, outage, power efficiency, delay, and fairness can be enhanced by jointly considering
the interactions across layers.
Copyright © 2006 Jin Yuan Sun et al. This is an open access article distributed under the Creative Commons Attribution License,
1. INTRODUCTION and the wired backbone is of great importance today because

of the increasing data application requirement at the mo-
With the growing demand and popularity of high-speed data bile terminal (e.g., cellular phone and wireless laptop). While
applications in wireless networks, the system capacity and most previous research was on the performance optimization
bandwidth resource become increasingly stringent. Radio re- of individual layer, it often leads to performance degrada-
source management plays a key role in wireless system design tion of other layers or suboptimal system performance. The
and analysis. Research efforts are made to expand the capac- hierarchical structure of the wireless networks, as the wired
ity and to efficiently allocate the resource. Some of the efforts ones, facilitates us to design and study protocols for the sin-
include the evolution of the CDMA technology. For example, gle layer that is of particular interest since these layers (phys-
3G (third generation) and 3G+ (beyond 3G) CDMA systems ical layer, link layer, network layer, transport layer, and appli-
(i.e., WCDMA and CDMA2000) are superior to 2G CDMA cation layer) are transparent to one another. But this isola-
systems (IS-95) in that the former ones have higher carrier tion may cause suboptimal system performance. Recent re-
bandwidth and faster power control frequency (more precise search has shown that a well-designed cross-layer approach
channel feedback). On the other hand, some efforts focus on that supports multiple protocol layer adaptivity and opti-
the algorithm design which can be implemented in the soft- mization can yield significant performance gains [1]. Many
ware to optimize the system performance. As the research on researchers use the cross-layer approach for their designs.
3G and 3G+ CDMA systems emerges, and although count- However, these designs can be very different due to various
less works have contributed to this research area, there still combinations and interactions of multiple layers.
remain a great number of problems unsolved. In this paper, We have studied a number of cross-layer approaches for
we would like to share our ideas to approach some of the CDMA system optimization in the literature. In what fol-
problems in this field. We choose to focus on the cross-layer lows, we summarize some of these approaches. Authors of
design of the CDMA networks as this issue has recently been [2–8] propose cross-layer approaches to achieve system op-
capturing interests. timization in CDMA systems. In [2] a set of PHY-MAC
The traditional wireless networks mainly support voice (physical-MAC) mechanisms is proposed based on the rate
service without data service provided through the Internet adaptation provided by the MAC and the channel state from
backbone. However, the integration of the wireless network the PHY to improve spectrum efficiency and reduce power
consumption. Yu and Krishnamurthy [3] focus on cross- enhancement while retaining separate performance features
layer QoS (quality-of-service) guarantee by combining phys- without degradation. The uniformity of the proposed frame-
ical layer SIR (signal-to-interference ratio) and network layer work not only simplifies the implementation of the schedul-
blocking probability to reduce computational complexity ing algorithms at base stations, but also is verified to be ro-
and approximate the optimal solutions. Other works are also bust and resistant to various offered traffic load and variable
found to address physical/network cross-layer optimization service structure (voice/data proportion). At network layer,
issues [9, 10]. Price and Javidi [4] deal with the interaction we propose an adaptive prioritizing soft handoff algorithm
between congestion (transport layer) and interference (MAC for concurrent handoff requests aiming at a same cell. A pre-
layer), and integrate them into a single protocol by means of dicted set, an adaptive priority profile jointly exploiting the
rate assignment optimization. Friderikos et al. [5] interpret impact of required handoff power, and call holding time have
the rate adaptation as TCP-related since the rate in this paper been developed to realize the proposed algorithm. A link-
is defined as the ratio of the current congestion window and layer scheduler residing in each base station to ensure the
RTT (round-trip-time) of the connection, and jointly con- desired operation of the prioritizing procedure is also de-
siders it with physical layer (power). Hossain and Bhargava signed, with input information from network layer. At trans-
[6] model and analyze the link/PHY level influence on TCP port layer, we study the problem of TCP over wireless link
behavior and illustrate their dependency. Yao et al. [7] study and summarize the solutions for this problem from exten-
the reverse and forward link capacities balancing issue by sive research works. We design an algorithm to prevent the
covering link layer and the network layer to seek for optimal spurious timeouts at TCP sources caused by the stochastic
handoff probability. Chan et al. [8] propose a joint source intervals of wireless opportunistic scheduling.
coding power control and source channel coding, and inter- The rest of this paper is organized as follows. Section 2 il-
pret them as the MAC-layer power control and application- lustrates the system model based on which this research is
layer source coding, respectively, maximizing the delivered carried out. The main body consists of Sections 3, 4, and
service quality and minimizing the resource consumption. 5, where we propose strategies for LINK/PHY (link/physical
There are also additional attention on other aspects of cross- layer), NET/LINK/PHY (network/link/physical layer), and
layer design, such as to decrease the cross-layer interference TRANS/NET/LINK (transport/network/link layer), respec-
[11] instead of optimization. tively, and study their interactions. Section 6 gives out the
The above survey indicates that different interpretations simulation environment and the performance evaluation/
of “cross-layer” and resources belonging to these layers pro- analysis of the proposed cross-layer design. Finally Section 7
duce a variety of cross-layer studies. While existing works concludes this paper with primary contributions, open issues
address cross-layer issues based on two or three layers, we and certain limitations.
propose to fully address this issue by taking the four impor-
tant layers into account: physical layer, link layer, network 2. SYSTEM MODEL
layer, and transport layer. We design algorithms/protocols
for each of these layers by considering their communications We concentrate on a cellular WCDMA system with wrap-
and mutual impacts to prevent isolation thus improving the around cell structure, as shown in Figure 1, where mobile
overall performance. At the physical layer, two fundamental stations (MSs) select serving base stations (BSs, displayed as
techniques, power control and rate allocation, are studied. pentagrams) based on the measured strength of pilot signals
The proposed integrated power control and rate allocation (P p ) sent out periodically by BSs. Typically, 20% of the down-
is briefly introduced which is primarily used for the link- link total power will be assigned to pilot channel by each BS.
layer scheduling and is demonstrated in detail in the follow- The rest of the total power is shared by traffic channel and
ing scheduling schemes. At the link layer, a novel voice packet an other control channel (carries control information, power
scheduling scheme named modified adaptive priority queu- control symbols, etc.).
ing (MAPQ) and a unified framework (UF) for scheduling In forward link (downlink), BSs transmit packets to MSs
hybrid voice and data traffic are proposed. An adaptive prior- through traffic channels which consist of frames. Each frame
ity profile is defined in these schemes based on queuing delay is of a 10 ms length and is subdivided by 15 time slots within
and physical layer information such as required transmission one of which traffic destined for one specific MS is delivered.
power, and available transmission rate, which borrows the To guarantee that MSs receive packets correctly, the
idea of composite metric from wired systems. Estimation er- transmission power allotted by BSs has to overcome chan-
ror is considered when measuring received pilots at mobile nel impairments, which consist of path loss (dib /d0 )−α , shad-
stations. For MAPQ, this definition ensures system capacity owing e−(βXb ) , and Rayleigh (fast) fading using Jakes’ fading
improvement, packet dropping probability reduction, and model [12], where dib is the distance between the MS and the
fairness. Users are allocated resources according to their pri- BS, d0 is the reference distance, α is the path loss exponent,
orities in a modified PQ fashion constrained by total power β = ln 10/10 is a constant, and Xb is a Gaussian distributed
budget of base stations. For UF, we address the consistency shadowing (in cell b) random variable with zero mean and
of the framework as well as the distinctions of voice and data variance σX2 .
scheduling processes by discussing the common policy and Unlike the uplink case, downlink restriction is neither the
individual requirements of both classes. With this design, intracell (from adjacent MSs within home cell) nor inter-
the proposed algorithm accomplishes system performance cell (from neighboring other cells) interference, but the total
Jin Yuan Sun et al. 3
3. LINK-LAYER PACKET SCHEDULING (LINK/PHY)
Packet scheduling is a promising link-layer management

mechanism. More and more research [15, 16] on this tech-
nique has emerged. However, existing algorithms focus on
data scheduling with little or no concern on voice. The time-
varying nature of wireless channels determines the impor-
tance of voice scheduling. Voice dropping is less likely to
come from congestion than data dropping due to its higher
priority. Nevertheless, voice traffic is prone to suffer a deep-
fading channel and consumes a large amount of limited base
station power, thus must be scheduled properly. In the first
part of this section, we propose a novel voice scheduling
scheme named modified adaptive priority queuing (MAPQ)
Figure 1: A typical cellular structure with base stations.
for forward link in a wireless cellular WCDMA environment.
An adaptive priority profile is defined in the scheme based
on queuing delay and required transmission power, which
borrows the idea of composite metric imported by IGRP (in-
transmission power of BSs. Thus the Eb /I0 (bit energy to in- terior routing protocol) and EIGRP (enhanced IGRP) from
terference density ratio), received at each mobile i from its wired networks.
home BS b, is expressed as Next, we present a unified packet scheduling framework
(UF) to include data traffic and demonstrate the easy inte-
PG gration and general adaptability of MAPQ to the UF.
γi = Γ i bi , (1) The Eb /I0 used here is the same as (1) except that R in
PT − Pi Gbi + j =b PT G ji
Γ = W/R here, R ∈ {Rv , Rd }, is the required transmission
rate of voice or data packets. The actual received Eb /I0 for
where Γ = W/R is the spreading gain with W the spread spec- each MS takes the form of (2).
trum bandwidth and R the required packet transmission rate. For link-level scheduling implementation, each BS has a
Pi and PT (in Watt) denote the transmission power for user scheduler to regulate incoming hybrid traffic, and arrange
i and total downlink transmission power of BSs, respectively, packets aimed for MSs in a specified sequence depending
assuming BSs are transmitting at the maximum capacity al- on the logic of the scheduler. The scheduler implemented in
ways. Gi∗ denotes signal attenuation (reciprocal to link gain) BSs is sketched in Figure 2, where QV 1 and QV 2 denote voice
of the channel between user i and BS ∗. During each time queues while QD1 and QD3 denote data queues.
slot, MS i compares the received γi with target Eb /I0 γ∗ and
generates the power control command informing its serving
3.1. Integrated power control and rate allocation
BS to increase or decrease the transmission power. Note that
unlike [13] and many others of its kind, we consider mul-
We propose to use physical-layer information power/rate as
tipath fading as part of the channel impairments instead of
the input for link-layer scheduling (LINK/PHY).
using an orthogonality factor (OF) in (1). Authors of [13]
Fast closed-loop power control (CLPC) is applied to our
argue that OF is defined as the fraction of received downlink
scheme in the downlink to optimize the system performance.
power converted by multipath into multiaccess interference.
The merit of CLPC has two aspects compared to the alterna-
They consider only the path loss as the channel impairment
tive open-loop power control (OLPC).
and include an OF to reflect the multipath impairment. Since
(1) Under relatively good channel conditions, where fast
the multipath impairment is reflected by fast fading, we do
fading is not severe, the transmission power for mobiles from
not employ OF here.
the base station can be kept to the minimum required level
In reality, however, limited measuring ability of MSs can
to satisfy the SIR at all times since CLPC performs faster
introduce an erroneous estimation of received pilots, which
than channel fading rate. Thus, CLPC is able to compensate
results in errors in measuring Gi∗ (obtained from measuring
medium to fast fading and inaccuracies in OLPC [17]. As a
pilot signals). We approximate the sum of measurement er-
result, more transmission power of the base station remains
rors as log-normally distributed as indicated in [14]. Let eβYib
for voice users that are either far away from the base station
denote the measurement error of the measured value, where
or in a difficult environment, and data users, giving rise to
Yib is a Gaussian distributed random variable with zero mean
enhanced system capacity and data throughput in multime-
and a variance of σY2 . As a result, the actual received SIR for
dia networks.
each MS is derived from the following revised version of (1)
(2) Under poor channel conditions where users undergo
counting errors:
severe fast fading, OLPC may fail to adapt to the required
transmission power for each mobile due to the slow rate
P /P and inaccuracy of OLPC, resulting in an insufficient power
γi = Γ i−βYT . (2)
1 − Pi /PT + e ib j =b G ji /Gbi level to combat channel fading and fulfill QoS requirement.
Voice output queue

Voice queue QV 1 after sorting
Voice queue QV 2 Allocated first

Output queue
after allocation
Data queue QD1
. Allocated after
.
.
Data queue QD3 Data output queue
after sorting
Figure 2: Base station scheduler structure for scheduling framework.
Whereas our approach CLPC solves this problem because the for all packets in the simulation. Zero buffering delay means
transmission power is always adjusted accordingly to satisfy that the packet gets served in the first round upon arrival.
the QoS requirement. Let delayi Vthre denote the buffering delay of packet i, the
Single-code variable spreading gain (VSG) rate alloca- delay threshold of voice beyond which voice packets en-
tion technique is applied to our scheme, as demonstrated ter QV 1 . This step is implemented in the scheduler pro-
later, to adjust the transmission rate when the satisfaction gramming. It does not indicate the actual packet move-
cannot be achieved by power adaptation only. According to ment in the memory where the queues locate. Let poweri
the calculation of the transmission rate R after spreading, and powermean denote the required transmission power of
R = W/Γ, where W and Γ represent the spread spectrum packet i, and the mean downlink transmission power of ac-
bandwidth and the spreading gain, respectively, the actual tive users in one cell. Then we have delaynor i = delayi /Vthre ,
transmission rate that is obtained can be adjusted by differ- and powernori = power i /power mean . The way to normalize the
ent values of Γ, given the fact that W is a constant. Typically, delay and power components in the AP expression ensures
in the WCDMA standard, Γ ∈ {4, 8, 16, 32, 64, 128, 256, 512}. the two terms in AP (a ∗ delaynor nor
i , b/poweri ) comparable.
The actual value assigned to Γ depends on the transmission Parameters a and b are the adaptive factors determining the
rate demanded. Note that R is inversely proportional to Γ. weight of delay over power (wdp, in our case voice is delay
When we select a smaller Γ, we will obtain a larger R, and sensitive thus a larger wdp should be assigned). The smaller
vice versa. required power (the better channel condition), or the larger
The way that we combine power and rate for our schedul- delay, yields the higher priority (AP). The way to define the
ing scheme is to first adjust power with a fixed rate to the priority profile ensures users with better channels and larger
point that the change of power no longer produces any effect delays to get served first, thus increases the throughput, re-
on the scheme (see Section 3.2.2). Then the power is fixed duces the voice packet dropping probability, and guarantees
and the spreading gain will be adjusted to obtain the satis- the fairness which is measured by the mean delay in the net-
factory transmission rate and to fulfill the predefined goal. work.
More details will be given in the description of the voice/data MAPQ scheduler then allocates resources starting from
framework. QV 1 according to AP values and total transmission power
budget of BSs. The scheduler does not terminate even if the
user currently being served in QV 1 requires a power exceed-
3.2. MAPQ and UF
ing the remaining budget. Since each user’s priority depends
3.2.1. Voice only: MAPQ on both delay and power, higher priorities do not merely in-
dicate smaller powers. Also, after QV 1 has been fully checked
In general, MAPQ has two subprocesses: sorting and alloca- and if there is power remaining (Pr ), the scheduler will con-
tion. The operation of MAPQ scheme and its subprocesses is tinue to check QV 2 and serve available users in QV 2 using Pr .
demonstrated as follows. This is the major difference between MAPQ and classic PQ. A
MAPQ scheduler sorts incoming traffic into high- (QV 1 ) similar concept can be found in [18], where the authors pro-
and low- (QV 2 ) priority queues based on each packet’s cal- posed a modified FIFO scheme for power-constraint systems.
culated priority value, which is evaluated by jointly consid- This modified queuing fashion can lead to higher-power uti-
ering required power and buffering delay as APi = a ∗ lization efficiency, as will be shown later.
delaynor
i + b/powernor
i , where APi denotes the adaptive pri-
ority for packet i. delaynor
i and powernori are the normal- 3.2.2. Unified voice/data framework: UF
ized buffering delay and the normalized required power of
packet i, respectively. The buffering delay is defined as the The rest of this section is dedicated to the discussion of the
time interval of the arrival and departure of a packet. For UF. In the proposed framework, the allocation scheme is
simplicity, the calculation time of the scheduler is neglected applied to not only each queue within each class but also
among classes. It is apparent that voice class has higher prior- In the particular reallocation, process data scheduler does
ity than data class, which necessitates the employment of the not stop when the remaining power is not enough to sup-
proposed allocation algorithm (modified PQ) for each voice port the current packet, because data transmission rate is
queue to secure the scheduler not skipping to data class be- adjustable according to the amount of remaining power re-
fore completing checkup within voice class. However, classic source as indicated earlier. Hence, the remaining power (if
priority queuing scheduler will not jump to data users until any) can be further allocated to specially selected data users
all the voice users have been served. It may cause power waste with various transmission rate values. The idea behind is
if none of the users in voice queues but some of the users in summarized as follows. After every packet in the queue has
the data queue can be served. The proposed allocation algo- been checked, if there is remaining power for one more
rithm further improves system capacity and throughput by “normal” packet even at minimum rate, this packet gets
also performing an exhaustive search within data class after served (system capacity increased). If not, share the power
an exhaustive search within voice class. We do not repeat the left among “normal” packets since they require lesser power
similar part of scheduling in terms of sorting and allocation (data throughput improved). Only if neither of the above
for data traffic but would like to emphasize the discrepancy is true, we give the extra power to the first “moderate” or
of data scheduling (sorting and allocation) as follows. “urgent” packet in the sorted output queue. This packet has
Difference in sorting mainly lies in the expression of already been allocated resource after all packets in the data
adaptive priority profile. We have for data queue have been scheduled. Pr is not allotted to “moderate”
or “urgent” packets which are still in the queue since they
APi = a ∗ delaynor
i + b/powernor
i + c/ratenor
i , (3) consume relatively large powers. As queuing delay increases,
they will be assigned a larger AP until eventually get trans-
where APi denotes the adaptive priority for packet i, and a, mitted. This is the key role of delay taken into account in our
b, and c are adaptive constants. Voice and data have different design.
a, b, and c values. Note that “ratenor
i ” denotes the normalized
new required transmission power after decreasing the data
rate, using (1), where Γ = W/Rd . The normalization method 4. NETWORK LAYER ADMISSION CONTROL AND
is similar to that used for voice power normalization. Here SOFT HANDOFF (NET/LINK/PHY)
we use “ratenor
i ” in order to distinguish from poweri
nor
in the
nor
AP expression. Furthermore, “ratei ” is used to emphasize High mobility and universal access are enabled by handoff
that the new required transmission power is obtained by ad- mechanisms employed in next-generation cellular CDMA
justing the data rate. networks. Without handoff, forced termination would occur
Voice requires constant bit rate during transmission and frequently as mobile users traverse cell boundaries.
thus it is unlikely to change voice users’ priority through rate An apparent pair of contradictive parameters represent-
variation. On the other hand, data (best-effort traffic in this ing resource management efficiency and effectiveness in a
paper unless otherwise specified) transmission rate is vari- handoff system is the call blocking probability (Pb ) and the
able and can be raised if extra power is available or reduced handoff dropping probability (Pd ). Since fixed downlink to-
without enough power resource to support target rate. In tal traffic power is shared between newly accepted users and
data scheduling profile, the last term “c/ratei ” is not used ongoing handoff users, one’s being greedy will induce an-
until delay exceeds a predefined threshold Dthre , and the re- other’s being starved. It is therefore important to regulate
quired transmission power is still too large to increase AP. and optimize their behaviors by balancing the amount of re-
If the packet enters this stage (delay > Dthre ), it implies that sources distributed. Based on the fact that interrupting an
merely adjusting power does not get an opportunity for the ongoing call is more disagreeable than rejecting a new call,
packet to be transmitted. At this point, transmission rate is handoff users are issued higher priority to reduce Pd wher-
adjusted since the change of power only no longer produces ever competition arises. Among numerous prioritizing algo-
any effect. This is the only case where we decrease the re- rithms [19, 20], resource reservation has attracted an over-
quired transmission data rate in order to get a smaller re- whelming favor owing to lighter required communication
quired power while SIR target is maintained. Since data users overhead. However, existing algorithms prioritize handoff
are able to withstand some delay and do not have strict drop users as an entire category towards new users category. None
bound, in which case wdp should be less than 1 to serve of them concerns priority assignment among handoff users,
users under desirable channel conditions (small transmission which is necessary when several users attempt to handoff to a
power) with preference. same target cell simultaneously, under the constraint of lim-
For simplicity, we set a to 1 and adjust b in the range of ited available target cell power (guard power PG ) set aside
(0, 1) for voice. For data, a is set in the range of (0, 1) while for handoffs. By “simultaneously” we mean while the BS is
b (or c) is set to 1. We compare to show the sorting sub- still handling one handoff request, another one or several re-
processes and allocation/reallocation subprocesses of voice quests may emerge, and so forth over and over. This is es-
and data scheduling in Figures 3 and 4. We observe in Fig- pecially true in metropolitan cities where the mobile net-
ure 3 that the sorting processes for voice (Figure 3(a)) and works are mostly busy with cellular phone users. It does not
data (Figure 3(b)) scheduling are similar. While Figure 4 il- mean two or more requests emerging at exactly the same
lustrates more complex data allocation process (Figure 4(b)) time instant, which is not very likely in reality. Therefore,
because of the additional reallocation process. it is imperative to avoid signaling flood at the moment of
Begin Begin
For each existing packet, Yes For each existing packet, No

Yes No
buffering delay > Vthre ? buffering delay == 0?
For each existing packet,

Put it into QV 1 in the Put it into QV 2 in the Put it into queue “normal”
No buffering delay > Dthre ? Yes
order of arrival time order of arrival time in the order of arrival time
Put it into queue “moderate” Put it into queue “urgent”

in the order of arrival time in the order of arrival time
Calculate AP for each packet in each queue
Calculate AP for each packet in each queue
i = 1: rear, i + +
find the packet with max AP
i = 1: rear, i + +
find the packet with max AP
Schedule this packet

Schedule this packet
i == rear ?
No i == rear?
No
Yes Yes
Results of
Results of data sorting
voice sorting
(a) Sorting subprocess of voice scheduling (b) Sorting subprocess of data scheduling
Figure 3: Sorting processes.
simultaneous handoff requests, since predictions would not With wireless Internet applications growing, the forward
occur at the same time. Please refer to [21] for detailed expla- link becomes very critical to system capacity, in that the
nations about the motivation and necessity of the proposed bottleneck-like power capacity of BSs imposes stringency on
prioritizing algorithm. available power resources allotted to each sharing user. The
We propose a handoff prioritizing algorithm, employing CAC mechanism employed in this paper, thereby, is based
a link-layer scheduler with network and physical layer input on total downlink traffic power and the precedence of hand-
(NETWORK/LINK/PHY). For prioritizing handoff users, a offs over new calls, which is shown as a handoff request is
scheduler implementing link-layer scheduling with network- admitted if
layer inputs inheres in each base station to regulate incom-
c
Non
ing handoff requests and to arrange these messages desig-
nated to different target BSs in a specified sequence depend- Pi + Pho ≤ Pt , (4)
i=1
ing on the logic of the scheduler, which has been proposed
with physical-layer inputs in the previous section. The sched-
uler applied to BSs is sketched in Figure 5, assuming the cur- and a new connection request is admitted if
rent serving base station is BS0 . BS1 , BS2 , . . . , BSn denote tar- c
Non
get handoff BSs, with n the number of handoff base station
targets. Pi + Prsv + Pnew ≤ Pt , (5)
i=1
4.1. Connection admission control (CAC) where Non c is the number of ongoing calls in cell c, and P , P ,
i ho
Capacity of CDMA systems is “soft.” Acceptance of each new Prsv , Pnew , and Pt represent the required power of an ongoing
call increases the interference level of existing ongoing calls call i, the incoming handoff call, the reservation for future
and affects their quality [22]. Hence, CAC is deployed to con- handoffs (will be discussed later), the incoming new call, and
the downlink total traffic power of BSs, respectively. Both (4)
trol the access to such networks, complying with types of ser- N c
vice and quality-of-service (QoS) requirement, as well as cur- and (5) conform to the general admission criterion i=on1 Pi ≤
rent system load. Pt .
Results of
data sorting
Allocate resource for each data packet i with

power pw(i) and the required rate based on
priority and power budget (pwt) until pw(i) > pwt
Results of k = (i + 1): rear, k + +

voice sorting
Yes
k == rear?
Allocate resource for each packet i in QV 1 with No
No
power pw(i) and rate Rv based on priority pw(k) <= pwt?
and power budget (pwt) until pw(i) > pwt
Yes
k = (i + 1): rear, k + + Permit packet k to transmit at
Yes power pw(k), the required
k == rear? rate; pwt = pwt pw(k)
No
pw(t) == 0? Yes
pw(k) <= pwt? No
No
Yes
Permit packet k to transmit Yes All “normal” packets
at power pw(k), rate Rv ; served (allocated)?
pwt = pwt pw(k) Pr = pwt; share Pr among
them with equal rate No
Allocate resource for each packet j in QV 2 with
power pw( j) and rate Rv based on priority Pr enough for min
Yes
and power budget until pw( j) > pwt rate of 1st “normal” packet
in the queue?
k = ( j + 1): rear, k + + Assign Pr fully to this packet
with max available rate No
Yes
k == rear?
No Yes Any “normal” packets No
No served (allocated)?
pw(k) <= pwt?
Share Pr among them
Yes with equal rate Assign Pr to the served
Permit packet k to transmit (allocated) “moderate”/
at power pw(k), rate Rv ; “urgent” packet whichever
pwt = pwt pw(k) has the largest AP
End End
(a) Allocation subprocess of voice scheduling (b) Allocation and reallocation processes of data scheduling
Figure 4: Allocation processes.
Ordered by Ordered by
arrival time adaptive priority
Output to BS1
Handoff requests to BS1 after prioritizing
Output to BS2
Handoff requests to BS2
after prioritizing
. . .
.. .. ..
Output to BSn
Handoff requests to BSn
after prioritizing
Figure 5: Base station scheduler structure for handoff algorithm.

Initiate
Mobiles send PSMM predicting handoff, iff (9),

and transfer pilots to predicted set
BS prioritizes predictions destined to a same cell based on AP values
Any of the above mobiles withdrew, iff (8), No

or dynamic channel reservation employed?
Yes
Measurement BS updates PQ (removing withdrawals) or dynamic PG after Wt ,
quantity ΔT ΔT ΔT
CPICH1 and signals its neighbors with the updated PG (dynamic scheme only)
BSs update PQ (based on the changing power and Tc ) after Wt

AS Th + AS Th Hyst
Mobiles send PSMM requesting handoff, iff (7)
AS Th AS Th Hyst
AS Rep Hyst
Mobiles identified by BS? Yes
CPICH2
No
BS sends corresponding guard capacity requests to the target
CPICH3
BS allots resources based on mobiles’ order in PQ (if any), procures
guard capacity, an exhaustive search, and informs mobiles by HDM
Time
Cell 1 connected Event 1A Event 1C Event 1B Mobile transfers pilot to active set and sends back HCM
add cell 2 replace cell 1 remove cell 3
Terminate
with cell 3
(a) A typical soft handoff algorithm (b) Proposed soft handoff procedure
Figure 6: Soft handoff algorithm and procedure.
4.2. Soft handoff the deletion from the active set are based on dynamic thresh-
olds. Let Mps b and Bestactive be the measured pilot signal from
ps
One of the major benefits of a CDMA system is the ability of base station b, and best measured pilot from the active set, re-
a mobile to communicate with more than one base station spectively. All the variables appearing in the inequalities be-
at a time during the call [23]. This functionality allows the low have the unit of Watt. A base station b is added into the
CDMA network to perform soft handoff. In soft handoff a active set if
controlling primary base station coordinates with other base
stations as they are added or deleted for the call. This allows
the base stations to receive/transmit voice packets with a sin-
b
Mps > Bestactive
ps − AS Th + AS Th Hyst, (6)
gle mobile for a single call.
In forward link handoff procedure, a mobile receives pi- for a period of ΔT, and is removed from the active set if
lots from all the BSs in the active set through associated traffic
channels. All these channels carry the same traffic (with the
exception of power control subchannel [23]), which facili-
b
Mps < Bestactive
ps − AS Th − AS Th Hyst, (7)
tates the mobile to gain macroscopic diversity by combining
power received from the channels (i.e., maximal ratio com- for ΔT, where AS Th, AS Th Hyst, and ΔT are design pa-
bining [24]). Thus, less power is needed implying total inter- rameters.
ference lessening and system capacity raising. We briefly describe the mobile-assisted soft-handoff pro-
A basic soft handoff algorithm typically used in 3G cedure as follows: mobile detects pilot strength from its mon-
CDMA systems is illustrated in Figure 6(a) [25], with AS Th, itored set by (6) and sends a pilot strength measurement
AS Th Hyst, AS Rep Hyst, and ΔT defined as the threshold message (PSMM) to the serving BS. BS requests resources
of reporting for active set transfer, hysteresis of the former from the target handoff cell, allocates traffic channel, and
threshold, replacement hysteresis, and time to trigger, respec- sends a handoff direction message (HDM) to mobile. Mo-
tively. CPICH is the abbreviation of common pilot channel. bile transfers this pilot to the active set and transmits to BS a
The events, together with the hysteresis mechanism and time handoff completion message (HCM). Mobile starts handoff
to trigger mechanism are discussed in [25]. drop timer when the pilot strength in the active set meets (7)
We employ a similar basic algorithm with slight simplifi- and sends to BS a PSMM. Mobile removes the pilot from the
cation. The selection of a base station into the active set and active set to the monitored set as the above time expires.
Note that the monitoring mechanism enables us to per- computation complexity to the base station and reduce the
form the prediction for prioritizing without extra network base station’s handoff processing speed, all of the above rea-
resources or high cost, as will be discussed in the next sec- sons reinforce the need for prediction.
tion. A predicted set is proposed in our algorithm, which con-
sists of BSs satisfying the inequality beneath,

4.3. Adaptive prioritizing soft handoff algorithm b
Mps > λ Bestactive
ps − AS Th + AS Th Hyst . (8)
The parameters and performance measures of the proposed The prediction threshold PS Th obeys the dynamics of the
prioritizing algorithm are addressed in this section, together threshold for the active set switching, and is related by
with the description of the detailed implementation proce- PS Th = λ(Bestactive
ps − AS Th + AS Th Hyst), where λ, λ ∈
dure of the algorithm. We mentioned in Section 1 that the (0, 1) is a design constant affecting the prediction threshold
adaptive priority profile is designed by jointly considering above which the pilot is added into the predicted set, relative
several elements, which are critical to define a specific hand- to the active set threshold. The criterion (8) serves as a trigger
off user. for the execution of the prioritizing algorithm. When (8) is
satisfied, MS will report to BS of the prediction and the call
4.3.1. Prediction holding time Tc , and the request will be put into the priority
queue. As long as the queue is not empty, BS will perform the
First of all, user mobility and location information are algorithm at the end of Wt .
needed by prediction, which is the prerequisite of the pri-
oritizing algorithm. This information is utilized by predic- 4.3.2. Downlink transmission power
tions for reserving guard capacity in the literature to track
the speed and moving direction of mobiles. However, Wang Next, channel condition should be taken into account of the
et al. [19] claimed that such information procured from mo- profile, in that it is the indicator of required handoff power. A
bility models or GPS monitoring is generally costly and inac- user experiencing better link gain and hence demanding less
curate, and complicated as well. As an alternative, they pro- power is given a higher priority, in order to get more users
posed using measured pilot strength to predict handoff (in served with the same amount of scarce downlink power re-
IS-95 systems) since it is the origin of every handoff thus source. Assuming the maximum size of the active set is 2 (i.e.,
is accurate. Moreover, it is inexpensive since no additional at most 2 BSs co-serve a handoff user at the same time), we
network signaling is needed. We take advantage of this idea can apply the maximal ratio combining strategy in (1) to de-
for the prediction in our algorithm, but modified it for 3G rive the Eb /I0 of a mobile i within the soft handoff zone as
CDMA systems (i.e., WCDMA). It must be noted that the P G
prediction method introduced in this paper is not as complex γi = Γ bi bi , (9)
b=0,1
PT − Pbi G bi + =b PT G ji
j
and precise as the aforementioned one because our focus is
not on guard capacity reservation algorithm. However, with where 0 and 1 are in general the two coserving base station’s
elaborately designed prediction scheme the significance and identity numbers and Pbi is the transmission power to mo-
effectiveness of our algorithm will be more prominent. bile i from BS b (current BS 0 and target BS 1). The actual
Typically, in addition to the avoidance of signaling flood, received Eb /I0 takes the form of (2). Based on the straightfor-
prediction is updated at the end of every prediction win- .
ward power division strategy [27] (i.e., P0i = P1i ), under the
dow Wt to remove withdrawals (i.e., (7) holds) resulting .
presumption of j =0 G ji /G0i = j =1 G ji /G1i , the required
from incorrect predictions or call termination (Tc > Dth , see handoff power from BS1 to mobile i can be written as
Section 4.3.3 below). The output priority queue (PQ) is up-
dated accordingly based on the latest information procured . γi PT 1 + j =1 G ji /G1i
P1i = . (10)
through prediction notification from mobiles. When hand- 2Γ + γi
offs actually take place, mobiles which are in PQ are iden-
tified by BS and are allocated channels immediately if the 4.3.3. Call-holding time
guard power allows. On the other hand, if the handoff re-
quests are not identified as in the regular handoff procedure, The last term included in the profile is the call-holding
these requests have to be sent to the target cell first since time Tc . This information can be easily derived by the UE
the BS has to inform the target to reserve power resources, (user equipment) through monitoring the connection time
where there exists the uncertainty about whether these re- elapsed for the ongoing call. For the proposed profile, we im-
quests can be approved with sufficient resources. Hence with port a parameter Dth denoting the death threshold for on-
prediction, the availability of resource is assured to maintain going calls. The ongoing call is presumed to be terminated
dropping performance. The handoff execution delay is also by the user before the actual handoff takes place if its Tc
shortened which may cause power outage and fade margin is greater than Dth at the time the prediction is made. If
enlarging [26]. Note that it is wise to shorten this delay by all Tc < Dth holds at the time of prediction, higher priority is
means especially in our case. Since additional handoff execu- assigned to a longer Tc . Because it is more probable that this
tion time can be caused by queuing and sorting the handoff mobile will terminate its call soon and release the resource
predictions in the proposed algorithm, which may introduce for other mobiles’ use.
We finally conclude the adaptive priority profile for user [31]), and recalculating the retransmission timer with some
i as backoff according to Karn’s algorithm [32]. In wireless net-
1 works, when packet loss occurs for some reasons other than
APi = + Tcinor , (11) congestion, such as temporary blackout due to fading, or
μP1inor
when packets are correctly received but the corresponding
in which APi is the user i’s priority and μ is the adaptive fac- ACKs have not been returned which is the so-called spuri-
tor adjusting the proportion of power and time to be com- ous timeout, TCP will perform the same as for reacting to
parable quantitatively. P1inor and Tcinor denote the normalized congestion in wired networks because it is not able to iden-
downlink transmission power and the normalized call hold- tify these different types of losses. The spurious timeouts of
ing time of user i, respectively. P1inor = P1i /Pmean , where Pmean TCP in wireless communications eventually lead to unnec-
is the mean downlink transmission power of predicted hand- essary cwnd/throughput drop and inefficient bandwidth uti-
off users for the same target cell. Tcinor = Tci /Tscale , where lization, especially in the presence of the well-known stochas-
Tscale = 10 s is the scale of the calling time. We use a rough tic internals of wireless scheduling which is the focus of this
calling time measuring method. The available Tcinor values are section. We address this problem, present existing solutions,
0, 1, 2, . . . , Dth /Tscale . While the actual Tci values can be any and provide our algorithm.
number between 0 and Dth , for simplicity, we assume that the
Tci /Tscale values will be ceiled to one of the above Tcinor values Although there are difficulties implementing TCP in
for the AP calculation. This definition style is derived from wireless networks, so far no single research has proposed
the wired networks, where IGRP and EIGRP routing proto- to replace TCP with another transport layer protocol suit-
cols define a composite metric associated with each route in able for communications over wireless links. It is unwise
an alike fashion as mentioned in Section 3. Specifically, we to remove TCP since its hierarchical relationship with pop-
subdivide users into two classes, which are distinguished by ular application-layer protocols such as HTTP, FTP, TEL-
different priority profiles. According to Viterbi et al. [28], the NET, and SMTP has been well established. In order to fa-
maximum fade margin (max γd ) put apart for overcoming cilitate the seamless integration of mobile communications
shadowing correlation (with coefficient a2 ) is obtained at the through wireless networks with the wired Internet backbone,
cell boundary, subject to a certain outage probability target TCP over wireless techniques are proposed. In general, the
∗
(Pout ). Hence, we issue boundary users a lower μ since they proposals found in the literature can be categorized into
require a higher power for handoff (due to a higher γd ) to three classes: split-connection protocols (i.e., indirect-TCP
ensure fairness. For convenience, we set μ = 1 for ordinary (I-TCP) [33]), end-to-end protocols (i.e., explicit congestion
users and μ ∈ (0, 1) for marginal users. Dedicated surveys on notification (ECN) [34]), and link-layer proposals (i.e., for-
fade margin improvement and delicate relations among pa- ward error correction (FEC) [35]). One may refer to [36]
rameters such as γd , Pout ∗
, and a2 are present in [26, 28, 29]. for a detailed survey on different classifications of TCP-over-
The implementation procedure of the proposed soft wireless solutions.
handoff algorithm is drawn in Figure 6(b). Note that we pro- To the best of our knowledge, the impact of down-
vide the option of dynamic channel reservation mechanism link scheduling on the performance degradation of TCP in
in the flowchart, in spite of its absence in our algorithm. Ad- CDMA networks has not received much research attention.
ditionally, the exhaustive search allocation scheme incorpo- Two works regarding similar issues in time-slotted networks
rated in the flowchart can be traced in [30], where we pro- have been found in the existing literature. Authors of [37]
posed a modified queuing algorithm, considering that a user proposed a reservoir mechanism at the base station to store
with a smaller required power is possible to be at the back of some ACKs during scheduling midseason and release them
PQ, since the synthetic AP value is determinant when prior- in the offseason to avoid spurious timeouts at TCP sources.
itizing incoming users. While in the classic first-in-first-out It is a revised version or addition of the Snoop protocol [38]
queuing scheme, users behind will not be allocated until all (a special link-layer protocol). The problem of this algorithm
of the front users are served. is that they use ICMP packets to measure the round trip
time (RTT) for ACK release interval calculation. These ex-
5. TRANSPORT LAYER TCP PERFORMANCE tra ICMP packets can significantly increase the network traf-
(TRANS/NET/WIRELESS LINK) fic especially in a large network where there are lots of TCP
senders and receivers. In addition, they did not demonstrate
TCP congestion control is originated and well investigated in clearly what methodology they utilized to measure the idle
wired networks where congestion is the main cause of packet period and the scheduling cycle at the base station. Authors
loss, thus operates properly in such networks. But wire- of [39] proposed to use pure MAC layer information to cal-
less networks and mobile terminals feature a large amount culate a TCP-related metric for link-layer scheduling. Thus
of losses due to bit errors and handoffs, thus are in some TCP performance is maintained when they use this metric in
facets non-cooperative with traditional TCP congestion con- the link layer to schedule traffic from TCP sources. This al-
trol, resulting in end-to-end performance degradation. In gorithm can also be called TCP-aware link-layer algorithm.
wired networks, TCP assumes that packet loss is caused by A crucial part of this algorithm is to use MAC information to
congestions and reacts to it by decreasing the congestion approximately calculate the average RTT. However, this ap-
window (cwnd), retransmitting the missing packets, trigger- proach is very complicated since it requires heavy mathemat-
ing congestion control/avoidance mechanism (i.e., slow start ical calculations to obtain the new metric at the beginning of
FTP/TCP Node Mobile

source 3 terminals
Mobile host
Node
Wired network 0
RNC Base station Base
Internet station Node
4
Node
Mobile host 2
Server Base station Mobile host Node

PDSN Node 5
1
CBR/UDP Node
Mobile host source
6
Mobile host Node
Base station 7
Mobile host
RNC
Wireless network
(a) The integrated network topology for TCP proposal (b) Network topology for TCP-over-wireless
simulations
Figure 7: Network topologies for TCP proposal and simulations.
each scheduling cycle and updating of the information for Wireless CDMA networks consist of two parts: the uplink
recursive calculation afterwards. and the downlink. In the reverse direction (uplink, i.e., from
One of the innovations of our works is that we propose an the mobile station to the base station), the key restriction
algorithm to eliminate TCP performance degradation due to is the incremental interference in the system as communi-
wireless scheduling in CDMA downlinks. In this paper, we do cating mobiles increase, due to transmission power levels of
not give details on how wireless opportunistic scheduling im- other active users and imperfect orthogonality of channel
pacts TCP performance since it is well illustrated in [37, 39]. codes. Scheduling in this direction is not needed as long as
We assume it is the existent problem and are concerned with the system interference stays below the threshold. Here we
the solutions in a CDMA environment. As addressed in the assume that the simultaneously active users in the system are
preceding section that in CDMA downlinks, scheduling users not enough to cause the interference beyond the threshold.
under better channel condition first can improve overall Therefore, scheduling in this direction is of little importance
network performances. Wherever scheduling arises in wire- to be considered by TCP performance. Rather, we focus on
less networks, there are impacts on TCP sources. The pro- the downlink direction where we proposed novel scheduling
posal is based on an example of real-time video transmis- schemes and explained their necessity and effectiveness.
sions where the jitter is smoothed out at the receiver to en- Through the analysis above, downlink scheduling is
sure a constant rate playout. We apply this idea to CDMA the only affecting factor to degrade TCP performance in
downlink scheduling which introduces stochastic halt affect- our study. Specifically, when interscheduling cutoffs (inter-
ing TCP performance at the sender (typically a fixed sta- vals/halts) occur, there is a temporary silent period in the
tion in the wired part of the network from which data can wireless part of the network for the scheduler to collect the
be downloaded using HTTP or FTP application protocols). up-to-date channel information and to perform the new
Figure 7(a) depicts the network topology we are dealing with, scheduling at the base station. During this period, no traf-
where RNC (radio network controller) and PDSN (packet fic is in the wireless network and the mobile station will not
data serving node) connect the wireless network with the In- send back the expected ACK since it has not received the TCP
ternet backbone. packet queued at the base station. Consequently, the TCP
TCP mechanism is quite mature in wired networks. The source may undergo spurious timeouts without the ACK it
problem that we are facing now is because of the exclusion is expecting.
of potential wireless applications when TCP was proposed. What if we avoid this burst-and-silence traffic pattern to
Naturally, one solution would be emulating the behavior of smooth the traffic throughout the burst and the following
the wired network, so that the “wireless” effect on TCP could silence period, just as what we do to avoid annoying jitter
be eliminated. Wired scheduling is periodical and hence pre- in video playout? Then the TCP packets stored in the base
dictable since every user is equal in terms of channel con- station will arrive at the mobile station with steady rate and
dition (no time-varying fading over wired links). It can be the mobile station will return the ACK without huge gaps
prevented and will not be a cause of TCP spurious timeouts. for TCP to timeout. After the scheduler determines the order
Thus wired scheduling is not within the scope of our study. of the packets to be transmitted based on the channel con-
On the other hand, wireless scheduling is unpredictable and ditions of each mobile user, it calculates a new transmission
thus irregular due to time-varying wireless links (users have rate to send the packets in a steady pace instead of sending
to be rescheduled according to their instant channel fading). them out all at once. In this case, there is traffic flowing in the
Table 1: Key technical specifications of WCDMA.
Direct-spread code
Multiple access technique Number of slots/frame 15
division multiple access
Frequency reuse 1 Number of chips/slot 2560
Carrier bandwidth 4.4–5.2 MHz Intrasystem handoff Soft/softer handoff
Chip rate of spreading bits 3.84 Mcps Power control period Time slot = 1500 Hz rate
Maximum user data rate 2.3 Mbps Power control step size 0.5, 1, 1.5, 2 dB (variable)
4 · · · 256 (uplink)
Frame length 10 ms (38400 chips) Physical layer spreading factors
4 · · · 512 (downlink)
network at all times so that there is no cutoff any more. This interpretation of “cross-layer” yields a different concern on
is our Proposal 1. Let the queues at the base station be per- complex connections among layers. Some layers may inter-
TCP-flow and the ACK per-packet based (TCP Reno [40]). act in terms of one measure while others may be related
Let Ni be the number of TCP packets in queue i of the refer- in terms of another. In general, it is difficult to generate
ence base station, let Tbi and Tsi (in seconds) be the midsea- a method that conforms the performance measures across
son (burst) and offseason (silence) duration of the schedul- all the four layers involved in our research, as proposed in
ing cycle of queue i, respectively. The playout rate to smooth Sections 3–5; the interacting layers are LINK/PHY, NET-
out the “jitter” of the burst traffic R p is written as WORK/LINK/PHY, and TRANSPORT/NETWORK/LINK,
respectively, where they are associated based on currently
Ni prevalent and practical problems concerned. Thus for each
Rp = , (12)
Tbi + Tsi combination which is formulated by these layers’ featured
relationships/interactions, there are individual measures that
where Ni is known to the base station through queue moni- best exhibited the performance gain over noncombination.
toring, Tbi + Tsi is equivalent to one term Tci , the scheduling This is how we design the simulation senarios to exploit the
cycle of queue i, which can be obtained from the history as performance gain for each cross-layer combination.
The simulations of the following subsections are set up in
Tci (n) = (1 − ρ)T ci (n − 1) + ρTci (n − 1), (13)
a WCDMA environment. Some of the key technical specifica-
tions of WCDMA [41] used for our simulation environment
where Tci (n) and Tci (n − 1) denote the nth and its previous,
setup are listed in Table 1.
the (n − 1)th, scheduling cycles, respectively. T ci (n − 1) de-
notes the average duration of scheduling cycle of queue i up
to scheduling cycle (n − 1). ρ is a weighing parameter with a 6.1. MAPQ and UF
typical value of 1/1000 [39]. The initial value of the schedul-
ing cycle (i.e., Tci (1)) can be monitored by the base station Other relevant parameters are 19 wrap-around cells with ra-
through some timer setting. Having smoothed out the “jit- dius r = 500 m (macrocell). One BS is located in the center
ter” using the above algorithm, the base station can “play” of each cell with PT = 20 W and a portion of 70% of PT is
the traffic continuously and get the ACK back to the TCP dedicated to traffic channel [42]. Mobility speed in Rayleigh-
sender accordingly, without temporary blackout which is the fading model is 10 km/h (vehicular environment), α = 4,
∗ ∗
root of TCP spurious timeout and performance degradation. σX = 8 dB, σY = 2 dB, γvoice = 5 dB, and γdata = 3 dB.
The above proposal can be easily implemented and effec- Hybrid voice and data users are uniformly distributed,
tive, which is based on the fact that the ACK flow back to the approximately 30 users per cell on average. Voice traffic is
TCP source is continuous as long as the TCP packets waiting modeled as “ON-OFF” with 50% “ON” duration probabil-
at the base station get transmitted to the mobile destination ity, and best-effort data traffic is generated with exponentially
continuously. It applies to wireless part of the network with distributed arrival rate. Generally speaking, voice traffic has
both comparable and negligible delay compared with the de- lower transmission rate compared to data traffic. In the in-
lay in wired part of the network, because the timeout interval tegrated voice/data scheduling scheme, minimum voice rate
is updated by TCP through the measured variable round trip Rv is selected from one of the following values: {8, 16, 32,
delay. 64} kbps corresponding to a spreading gain of 512, 256, 128,
and 64, respectively, while Rv = 64 kbps is the fixed transmis-
6. PERFORMANCE ANALYSIS-SIMULATION sion rate in the voice-only scheduling (MAPQ). Data rate Rd
ENVIRONMENT & NUMERICAL RESULTS can be chosen from any available value allowed by the spread-
ing gain set of {4, 8, 16, 32, 64, 128, 256, 512}.
We address the performance measure and numerical results For voice-only scheduling, maximum tolerable delay is
in this section. As argued before (see Section 1), different dmax = 100 ms, and buffering delay (db = 60 ms) is used in
0.7 0.9
0.65 0.8
Normalized system capacity
0.6
Mean normalized delay

0.7
0.55
0.5 0.6
0.45 0.5
0.4
0.4
0.35
0.3
0.3
0.25 0.2
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Traffic load Traffic load
MAPQ MAPQ
FIFO FIFO
STPD STPD
(a) System capacity comparison for MAPQ (b) Mean normalized delay comparison for MAPQ
Figure 8: Performance evaluation of MAPQ.
the simulations to determine the unfairness criterion. The (2) UF for hybrid traffic.
power for calculating AP is procured from (1). For integrat-
(i) System capacity-NS /NC .
ing voice/data scheduling framework, the delay thresholds
(ii) Traffic throughput-ψA /NC .
for sorting voice and data packets into different queues are
(iii) Outage probability: fraction of time that a user’s
Vthre = 10 ms, and Dthre = 120 ms. The delay bounds of voice
received power is below the minimum acceptable
and data packets are 100 ms and 2 s, respectively, which will
power level to satisfy the target SIR.
be used in the mean normalized delay calculation in the fol-
(iv) Average power utilization (efficiency)—(total
lowing section.
power consumed)/(total traffic power budget of
We first define the performance measures used in the
BSs). It acts as the indicator of resource consump-
MAPQ and UF simulations for voice and data. Define NF ,
tion efficiency.
NA , NS , NC , ψF , and ψA as the total number of users in the
(v) Voice ratio—the proportion of voice traffic in the
network (in our case 30/cell), the number of active users in
hybrid traffic. It controls the variation of the hybrid
the network, the number of active users actually served, the
traffic structure.
number of cells in the system (in our case 19), the through-
put (kbps) of the system if all the active users can be served, (3) For both MAPQ and UF.
and the actual throughput of the system, respectively.
(i) Mean normalized delay:
(1) MAPQ for voice only. Ns
i=1 normalized delay of packeti
(i) Normalized system capacity (throughput)-ψA /ψF . , (14)
Ns
Note that voice packets have constant transmission
rate thus the capacity and throughput have similar where the numerator equals to
behaviors. buffering delay of packeti
(ii) Packet dropping probability—(number of packets . (15)
delay bound of packeti
dropped)/(number of packets transmitted). A voice
packet is dropped if its buffering delay exceeds the We measure the normalized delay only for success-
delay bound (100 ms). fully served users because there are other criteria,
(iii) Unfairness probability—we call it “unfair” if a voice dropping probability and data outage, to illus-
user’s buffering delay is greater than db yet not trate the behavior of each scheduling scheme with
served. Therefore, unfairness probability refers to service failures.
the probability that such unfair event happens. One
possible way is to use NUP /NA to measure it, where 6.1.1. Voice only: MAPQ
NUP is the number of users that experience unfair-
ness. The simulation runs over 100 000 times. Compared to sys-
(iv) Traffic load-NA /NF . tems where no sorting scheme nor modified PQ allocation
0.9 0.9
0.8 0.8
Voice dropping probability
Unfairness probability
0.7 0.7
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Traffic load Traffic load
MAPQ MAPQ
FIFO FIFO
STPD STPD
(a) Dropping probability comparison for MAPQ (b) Unfairness probability comparison for MAPQ
Figure 9: Performance evaluation of MAPQ.
scheme (FIFO) is deployed, system performances of the terms of voice dropping probability and unfairness probabil-
MAPQ scheme in terms of system capacity, voice packet ity, as shown in Figures 9(a) and 9(b). In Figure 9(b), MAPQ
dropping probability, mean normalized delay, and unfairness and STPD almost have the same performance but as the traf-
probability improve in various degrees, as shown in Figures fic load becomes heavier, the MAPQ shows the trend to out-
8 and 9, respectively. perform the STPD.
As the traffic load grows heavier, the performance gain Note that the fairness criterion in our simulation is im-
of the proposed scheme becomes more apparent. This can plied by both the mean normalized delay and the unfairness
be explained by the following observations. When the power probability measures. Smaller normalized delay and lower
budget is getting tighter and users are more competitive for unfairness probability indicate higher degree of fairness.
limited resources, the FIFO scheme is not capable of produc- We also testified the necessity of both sorting and the al-
ing satisfactory results due to its inadaptability to severe sys- location subschemes of MAPQ scheme by comparing MAPQ
tem environment. While the MAPQ scheme is generally sta- with two reference cases, namely, allocation (modified PQ)
ble and insensitive to throughout traffic variation, and able to without sorting and sorting without allocation (classic PQ).
produce acceptable outcomes even if experiencing stringent The proposed scheme outperforms both of the references
conditions, as a result of well designed adaptive features. in terms of system throughput, packet dropping probability,
Moreover, we compare the MAPQ scheme with a more and unfairness probability with 2%–10% performance gains
advanced scheduling scheme in the literature named STPD (not shown in this work).
(scheduling with transmission power and delay) [15]. In this
scheme, packets whose required transmission power is less 6.1.2. Unified voice/data framework (UF)
than a threshold Pth are classified into Group 1, otherwise Individual performance gain of voice under the proposed
are classified into Group 2. For real-time traffic like voice, if scheduling algorithm has been procured and illustrated
the maximum buffering delay of Group 1 is less than a delay above in terms of system capacity/throughput, packet drop-
threshold, Group 2 is transmitted first to avoid the exceed- ping probability, and unfairness probability. Note that the
ing of the tight delay bound. While for non-real-time traf- values of a, b, and c used in the simulation are obtained from
fic like data, Group 1 is always transmitted first since it is the estimation.
delay-tolerable. This algorithm is less complex in calculation In this section, we focus on measurable performance of
since it does not use the priority to sort each packet. How- hybrid voice/data traffic under the proposed unified frame-
ever, the simplicity may result in some degradation of the work. Three reference algorithms are compared with our
performance, as shown in Figures 8(a) and 8(b), where ob- algorithm, and evaluation is realized through several sig-
viously the more complex MAPQ scheme performs better in nificant criteria: system capacity, traffic throughput, outage
terms of both the system capacity and the mean normalized probability, average power utilization, and mean normalized
delay in the network. delay. the first reference algorithm employs SPS (static pri-
The MAPQ scheme not only outperforms the STPD ority scheduling) [18] algorithm for either class, the sec-
in these criteria, but also maintains other performances in ond reference algorithm employs STPD (scheduling with
30 103
3.5
Traffic throughput (kbps)

25
System capacity
2.5
20 2
1.5
15
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Voice ratio 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Voice ratio
UF
SPS UF RF
STPD SPS STPD
(a) System capacity comparison for UF (b) Traffic throughput comparison for UF
Figure 10: Performance evaluation of UF.
transmission power and delay), and the third reference algo- fails to combat channel fading and causes transmission fail-
rithm employs data scheduling only, which is the popular RF ure (SIR target unsatisfied). While the power baseline in SPS
(rate fairness) algorithm as in [43, 44]. They represent three is assured, because of the lack of sorting process, users are
typical simulation conditions. served in a first-in-first-out mode where a user occupying
Figures 10 and 11 show the comparison results in terms large power is possible to be allocated before users request-
of system capacity, traffic throughput, outage probability, ing smaller power. Also, due to our allocation subalgorithm
power utilization efficiency, and mean normalized delay, re- which is based on an exhaustive search mechanism, system
spectively. The results are obtained by averaging over differ- capacity and traffic throughput of the proposed algorithm
ent load situations. The proposed algorithm achieves better outperform those of SPS as displayed in Figures 10(a), 10(b)
performance in all cases in comparison with SPS and STPD. (solid versus dashed curves), respectively, without sacrificing
The proposed framework illustrates an appropriate combi- QoS satisfaction degree (here defined by outage probability
nation of individual performance sustentation as well as in- as in Figure 11(a), solid versus dashed curve).
tegrated optimization and robustness. Figure 10(a) exhibits Aside from the exploitation and maintenance of individ-
the capability of the proposed algorithm to explore the bursty ual behavior features, the capability of attaining overall op-
nature of the data traffic with a visible steep slope of the solid timization and robustness to various traffic structure, result-
square-marked curve under the variation of traffic ratio. This ing from the consistent infrastructure of the framework, is
capability benefits from discrepancy of voice/data schedul- depicted in Figures 11(a), 11(b), and 11(c), since in these fig-
ing (i.e., different adaptive priority profiles, data reallocation ures the fluctuation of the traffic structure does not affect UF
mechanism) in the framework. While SPS (dashed asterisk- much.
marked curve) appears mildest under the traffic structure Voice traffic has less burst, smaller range, and lower data
change, as a result of limited capacity and resource utilization rate than data traffic. Wang et al. [19] advocates that with
capability. Solid square-marked curves of Figures 10(a) and the same total offered load, a larger fraction of voice permits
10(b) confirm the results in [45] in which system through- better multiplexing and hence more efficient resource usage,
put (Figure 10(b)) decreases with the reduction of data por- giving rise to the fluctuation of average power utility with
tion, probably because of the fact that data traffic has much voice ratio, shown in their simulation results. On the con-
more burst and much higher bit rate, thus is more susceptible trary, Figure 11(b) convincingly illustrates that the UF guar-
to load change and affects more on throughput and capacity antees power utilization efficiency at above 97% at all times
(Figure 10(a)) behaviors. benefiting from data reallocation mechanism and is resistible
Although RF has much higher traffic throughput (ob- to a variety of offered traffic load and voice/data ratio. At the
served in Figure 10(b), circle-marked dash-dot curve) be- same time, the obvious fluctuation of the dashed curve and
cause it distributes rate equally among all active users regard- the dotted curve reflects the vulnerability of SPS and STPD
less of actual channel impairments, which comes at the ex- under the altered traffic structure.
pense of users being served with power below the minimum Further, the reallocation mechanism ensures “good”
acceptable target at most of the time (circle-marked dash-dot users to be better and “poor” users to be served, leading
curve in Figure 11(a)), this insufficient transmission power to fairness protection as shown in Figure 11(c), low and
1 1
0.9
0.8 0.95
Average power utilization

Outage probability
0.7
0.6 0.9
0.5
0.4 0.85
0.3
0.2 0.8
0.1
0 0.75
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Voice ratio Voice ratio
UF UF
SPS SPS
RF STPD
(a) Outage probability comparison for UF (b) Power utilization efficiency evaluation for UF
0.9
0.8
0.7
Mean normalized delay
0.6
0.5
0.4
0.3
0.2
0.1
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Voice ratio
UF
SPS
STPD
(c) Mean normalized delay (fairness) evaluation for UF
Figure 11: Performance evaluation of UF.
insensitive outage occurrence (solid curve in Figure 11(a), gain of the UF scheme comes largely from the desirable de-
reflecting high degree of QoS satisfaction) throughout var- lay performance of data users. Due to the dominant char-
ied voice ratio. acteristics of data packet delay in the network, this perfor-
Figure 11(c) shows the outstanding delay performance mance gain also compensates for the voice packet delay per-
of the UF scheme. In general, higher voice ratio yields bet- formance, observed as a nearly flat curve (the solid square-
ter delay performance since data packets may experience marked curve) in Figure 11(c). This nonfluctuated delay per-
much longer delay thus have larger impact on the delay be- formance curve further demonstrates the fairness assurance
havior in the network. In fact, the delay behavior of the among data users in the UF scheme. In addition, it implies
MAPQ scheme in Figure 8(b) does not show such striking that, in the MAPQ scheme, the improved performance of
performance enhancement. We mentioned before that data voice packets does not necessarily sacrifice the data packets
packet delay is usually much larger than voice packet de- behavior (i.e., data packet delay).
lay. The data packet delay performance determines the de- We believe all of the above gains are acquired from the
lay performance of the entire network. Therefore, we can unity of proposed framework and innovative sorting and al-
conclude that the result in Figure 11(c) suggests that the location (reallocation) mechanisms deployed within.
Table 2: Simulation parameters of the proposed handoff algorithm.
System parameters Value Handoff parameters Value

Bandwidth/chiprate 5 MHz/3.84 Mcps AS Th/AS Hyst/ΔT 2 dB/2 dB/0
Cell radius 1000 m (macrocell) Wt , Dth 1 s, 2 min
User locations/arrival type Uniformly/exp. distributed Tc (exp. distributed) Mean 1 min
PT , P t 20 W, 70% PT PG (for each neighbor) 4% Pt
Traffic model ON-OFF, 50% ON Prob. λ, μ 0.85, 0.8
User mobility/max speed 2-dimen. Rand. walk/100 Km/h a2 0.5
Rv set {64, 32, 16, 8} kbps Max. γd 6.2 dB
∗ ∗
α/σX /γ 4/8 dB/5 dB Pout 0.1
6.2. Prioritizing handoff algorithm We run over 10 000 trials and get the averages of the
above performance indicators as seen in Figures 12(a), 12(b),
Next, the performance of the network layer proposal is and 12(c).
evaluated. Table 2 displays the system and handoff param- The reason for the outperformance has two folds which
eters for our simulation senario. Handoff parameters such reflect the innovation of the proposed prioritizing algorithm.
as Wt /λ/PG , μ (derived from normalized Rs , the radius of (1) Handoff users demanding smaller powers are scheduled
∗
nonsoft handoff zone in [46]), and a2 /γd /Pout are obtained first in general, contributing to a larger number of successful
from [19, 28, 46], respectively. In [19], a fixed guard capacity handoff users with the same amount of guard power, thus Pd
reservation scheme is shown to have similar performance to is reduced (Figure 12(a)) and power efficiency is enhanced
their dynamic ones in terms of new call blocking probability, (Figure 12(b)).
handoff dropping probability, and average power utilization (2) Handoff users who have been connected for a longer
but with higher guard power waste at lighter load. Therefore, call time (Tc < Dth ) are more likely to cease and re-
we employ fixed reservation scheme for simplicity and gen- lease resources shortly, before other concurrent handoffs are
erate heavier traffic load (50%–100%) to diminish the guard dropped due to the lack of enough guard resources. It can be
power utilization discrepancy. Offered traffic load used in the interpreted as the resource borrowing (or reuse) mechanism
following figures is the actual amount of traffic normalized of earlier and faster handoffs from slower handoffs, giving
by the fully loaded amount of traffic in the network. Heavier rise to higher guard power utilization (Figure 12(c)).
traffic load imposes greater challenge on the success of the Note that an overlow λ triggers more frequent and false
proposed algorithm because of more severe dropping condi- predictions, because it extends the active set and may lead
tion. to too weak potential BSs. While an overhigh one hinders
Figures 12(a)–12(c) demonstrate that the proposed pri- potential true predictions. Consequently, this parameter has
oritizing algorithm outperforms the FG (fixed guard capac- to be designed carefully. The authors [19] addressed similar
ity) scheme [19] in terms of handoff dropping probabil- concerns on λ. We use λ = 0.85 as indicated in [19].
ity (Pd ), average guard power efficiency, and average guard From Figures 12(b) and 12(c) we observe that higher of-
power utilization, under the same prediction scheme and fered load yields larger difference between the proposed algo-
general handoff procedure proposed in Section 4.3. These rithm and the FG. Indeed, whether the guard capacity reser-
performance measures are defined below, where Ntho , Nd , vation scheme is fixed or adaptive produces no distinction in
Nssho , Prsv grd , and Psum ho denote total number of received terms of average guard power utilization, as shown in [19].
handoff calls, the number of blocked (dropped) handoff However, the fixed and adaptive reservation schemes can af-
calls, the number of successful handoff users, total reserved fect the handoff dropping probability to have very different
guard power, and total consumed guard power (sum of suc- performances, which is verified also in the above research
cessful handoff powers derived from (10)), respectively. work. Additionally, the authors show that the adaptive reser-
vation schemes exhibit greater difference between each other
(i) Pd -Nd /Ntho . as the offered traffic load grows. But since we employ fixed
(ii) The guard power efficiency-Nssho /Prsv grd . It indicates reservation schemes, the fixed guard power may fail to ex-
the number of successful handoff users that can be ploit or adapt to the dynamics of the handoff dropping prob-
supported by consuming certain Prsv grd , thus clarifies ability when the system is highly loaded and has poorer drop-
the efficiency of Prsv grd . ping behavior. This explains the reason for the nearly paral-
(iii) The guard power utilization-Psum ho /Prsv grd . It indi- lel curves in Figure 12(a). In the future research, more per-
cates the utility of the reserved guard power. If it is too formance gains of dropping probability are expected with
low, that implies the underutilization and a waste of the employment of the adaptive guard capacity reservation
system resources. scheme in the proposed handoff algorithm.
10 2 2.5
2.6
2.4
2.4
Average power efficiency

2.3
2.2
Dropping probability
2.2
2
2.1
1.8
1.6 2
1.4 1.9
1.2 1.8
1 1.7
0.5 0.6 0.7 0.8 0.9 1
0.8
0.5 0.6 0.7 0.8 0.9 1 Offered traffic load
Offered traffic load
Proposed handoff algorithm
Proposed handoff algorithm FG
FG
(a) Handoff dropping probability (b) Average power efficiency for handoff algorithm
0.65
0.6
Average power utilization
0.55
0.5
0.45
0.4
0.6 0.7 0.8 0.9 1
Offered traffic load
Proposed handoff algorithm

FG
(c) Average power utilization for handoff algorithm
Figure 12: Performance evaluation of the proposed handoff algorithm.
6.3. TCP cwnd behaviors there are a FTP source (node 0) sending TCP (data) traffic
and a CBR source (node 1) sending UDP (voice/video, re-
Finally, the performance analysis of transport-layer TCP/ ferred to as voice hereafter for simplicity) traffic. We create 5
wireless is presented. This section is dedicated to the nu- nodes (nodes 3–7) at the wireless terminal as the receivers of
merical results for the feasibility of the proposed algorithm both TCP and UDP packets. The base station is modeled by
(Proposal 1) for TCP performance improvement, in terms of node 2 connecting the bottleneck link. Note that the simula-
the evolution of TCP congestion window (cwnd) and TCP tion for Proposal 1 does not include the UDP sender thus
throughput. With the reasonable analysis and implementa- studies the TCP performance based on a pure TCP-traffic
tion details, success and effectiveness of the proposed al- network. However, the simulation for the extended analy-
gorithm is further confirmed in the simulation. While the sis of the design parameters employs exactly the topology as
network simulator-version 2 (NS-2) [47] models do not Figure 7(b).
support CDMA air interface or CDMA MAC implementa- The MSs act as the TCP sinks. Assume that the TCP
tions, there is difficulty to simulate the TCP performance sender always has data to transmit and can transmit as many
with the presence of the proposed scheduling schemes in NS- packets as its transmission window allows (bulk TCP data).
2. In this paper, we provide the experimental results using The TCP sinks receive TCP packets to deliver them to the user
MATLAB and leave the simulations in NS-2 for future work. and generate immediate ACKs for the TCP sender. The TCP
We set up a simple network as shown in Figure 7(b), where mechanism implemented for the experiment is only related
30 45
25
40
20
cwnd (packets)
cwnd (packets)
35
15
30
10
25
5
0 20
300 350 400 450 500 550 600 300 350 400 450 500 550 600
Nominal simulation time Nominal simulation time
(a) Evolution of TCP cwnd for standard TCP (b) Evolution of TCP cwnd for Proposal 1
Figure 13: Performance evaluation for TCP cwnd behaviors.
to dynamic congestion window evolution, according to the ACK (one ACK per TCP packet). The design parameter cho-
slow start and congestion avoidance algorithms [31]. Other sen for this purpose is the buffer size/queue limit BU. We set
mechanisms such as retransmission and recovery for error it to 40 (packets) based on the analysis in the extended sim-
control are not considered. ulations. Figures 13(a) and 13(b) illustrate the evolution of
The data is transmitted using variable bit rate in the TCP cwnd in terms of the nominal simulation time, for standard
packets with a length of 1000 bytes/packet. Voice (only sim- TCP and Proposal 1, respectively. Note that the nominal sim-
ulated later in the extended analysis for design parameters ulation time is used as the x-axis instead of the real simula-
analysis) is transmitted in the UDP packet with constant tion time (s) because of the lack of the simulator. We run the
bit rate selected from one of the following values: 8 kbps, simulation created by algorithms in MATLAB for over 60 000
16 kbps, 32 kbps, and 64 kbps, as used in the scheduling times, approximately every 10 000 times represents a 100 s in
schemes in Section 3. The packet length of a UDP packet a real simulator (i.e., NS-2). The curves in Figures 13(a) and
is a design parameter in our research and will be demon- 13(b) are plotted using the data obtained from every nominal
strated later. The buffers proposed in the base station and the time (0.01 s in NS-2).
TCP sender are per-TCP-connection and drop-tail, with the In order to distinguish the spurious timeout and the real-
size varied in the simulation. The bandwidth and propaga- loss-triggered timeout, we set the wireless link loss (due to
tion delay of the wired connections (node 0-node 2, node 1- shadowing and fading) to be 0. The irregular fluctuation of
node 2) are 20 Mbps and 20 ms, while they are set to 1 Mbps the cwnd in Figure 13(a) suggests only the spurious time-
and 1 ms for the wireless connections (node 2-node i, i = out due to wireless scheduling intervals and implies the in-
3, 4, 5, 6, 7). competence of the standard TCP in the presence of wire-
Since the scheduling cycle and interscheduling cutoff pa- less scheduling, which is implemented based on the MAPQ
rameters vary depending on the complexity and execution and the UF schemes proposed in Section 3. We also observe
time of different scheduling schemes, no typical parame- from Figure 13(b) that, compared to standard TCP, Proposal
ters are found in the literature. In reality, wireless service 1 performs much better in the cwnd evolution because it
providers deploy real-time monitoring systems to obtain use- avoids the spurious timeouts which largely degrade the cwnd
ful statistics such as link activity, bandwidth utilization, link performance. We acquire the desired cwnd behavior from
occupation time, throughput, for internal and external uses. Figure 13(b) that the congestion window keeps increasing
Link activity makes the scheduling cycle easy to be measured until the buffer overflows. Thus, the maximum window size
because the link will be idle if interscheduling cutoff hap- is always maintained which is approximate to the buffer size
pens. In our simulation, NS-2 provides real-time monitoring (in our simulation 40 packets). While the congestion window
mechanisms such as queue-monitor by which we can mea- behavior in Figure 13(a) does not obey the sawtooth shape,
sure the scheduling cycle using the packet arrival-departure and the achievable window size is even less than 20.
statistics in the queue-monitor output. Interscheduling cut- In addition, we analyze the TCP throughput performance
off happens at 0 departures. for the standard TCP and Proposal 1, as displayed in Table 3.
The throughput is calculated by
6.3.1. Proposed TCP algorithm ThruTCP = LTCP NTCP /Tnom , (16)
The congestion window behavior of Proposal 1 is compared where ThruTCP , LTCP , and NTCP denote the TCP through-
to that of the standard TCP, TCP Reno [40] with standard put (bps), the TCP packet length (8000 bits/packet), and the
Table 3: TCP throughput comparison.
TCP scheme Throughput (kbps) Number of TCP packets generated

Standard 54.216 89149
Proposal 1 70.217 90125
Table 4: Effects of design parameters on TCP cwnd behavior.
Achievable Achievable Achievable

PS(CBRR=64,BU=30) CBRR(PS=80,BU=30) BU(PS=80,CBRR=64)
cwnd cwnd cwnd
80 5.9 8 20.5 20 3.9
200 7.7 16 14 200 11.2
1000 16.8 32 8.2 400 21.1
total number of TCP packets generated (shown in Table 3), not configure queuing-related parameters (e.g., queuing dis-
respectively. Tnom denotes the total nominal simulation time. ciplines) since they are not considered in this research. By
We may not see striking throughput gain from this ex- default the queue is FIFO and drop-tail.
ample since the network we built is relatively simple thus not The CBRR impacts present the rationale explicitly: the
enough traffic is generated to test the throughput behavior. higher the voice transmission rate is, the more likely that
But from the tendency one can clearly tell that Proposal 1 has there is congestion in the network since the bandwidth re-
more desirable throughput performance than the standard source becomes more demanding. The increasing contention
TCP. We believe that with heavy loaded network simulated for the available bandwidth will induce more packet loss for
in the future, this throughput gain will be more apparent. both traffic, and the timeouts will occur inevitably which de-
creases the congestion window size.
6.3.2. Extended analysis of the design parameters
It makes sense from the BU effects that if the buffer is
Although the impacts of the network layer QoS profile, which designed to be small, the memory required for buffering is
manages the queuing disciplines and the bandwidth alloca- low trading off the maximum attainable cwnd because the
tion, are beyond the scope of this research, we analyze the ba- buffer will be full rapidly. In other words, optimizing the con-
sic role in the evolution of TCP congestion window since the gestion window with a large buffer challenges the excessive
queuing mechanism for Proposal 1 at the base station makes memory which can be costly. The buffer selection is hence
use of the network layer functions. We found through the important as well as complicated and should be designed
simulation by using NS-2 that, in general, the variations of carefully. Furthermore, we realize through the comparison of
PS (voice packet size), CBRR (voice transmission rate), and 16 kbps CBRR and 200 packets BU that a lower transmission
BU have great impacts on TCP cwnd. In the subsequent sim- rate (16 kbps versus 64 kbps) requires a considerably smaller
ulations, we employ standard TCP over the wired network buffer (30 versus 200), yielding similar or even better cwnd
(topology shown in Figure 7(b)) with 0 loss rate to study the performance. Analysis of the buffer size is the only issue in-
effects of the design parameters. Table 4 illustrates the change volved as the network-layer mechanism that we address in
of cwnd as a result of the varying PS (in the simulation 80, this work. Taking the comparison and analysis of this sub-
200, and 1000 bytes) with CBRR fixed at 64 kbps and BU 30 section into account, we designed the simulation parameter
packets; the varying CBRR (8, 16, and 32) with fixed PS and BU in the earlier simulation.
BU at 80 and 30, respectively; the varying BU (20, 200, and From the extended analysis of the design parameters, we
400) with fixed PS 80 and fixed CBRR 64. further conclude that for an algorithm or a protocol to work
It is noticed that a smaller PS results in greater reduction appropriately, several key parameters (i.e., PS, CBRR, BU)
of congestion window size, which will largely affect the im- need to be tuned carefully.
provement of TCP throughput. The reason maybe that since The experiment and simulation for this subsection are
the wired links in the network have the same bandwidth, primarily operated for verifying the feasibility and effective-
the smaller voice packets produced by the CBR source arrive ness of the proposed strategies. It is preparatory for our fu-
faster and have a higher generation rate. They are of higher ture research which includes in-depth study of the impacts
probability to be queued in front of the bigger data packets of varying TCP/UDP traffic ratio on TCP behaviors in the
thus get transmitted first. On the other hand, data packets integrated wired/wireless environment which were somehow
are more likely to be dropped if the queue overflows. This demonstrated in the above simulations; a more realistic net-
loss of TCP packets will further lead to the loss of the cor- work structure and traffic generation for overcoming prac-
responding ACK. It will eventually triggers the TCP timeout tical problems, such as the existent “wireless” effect on TCP
and then the congestion window reduction. Note that we did performance, using NS-2.
7. CONCLUSIONS AND OPEN ISSUES (2) In Section 3, further details pertaining to the choice
of the parameters (a, b, and c) will be provided.
In this work, we analyzed and designed cross-layer algo- (3) Another aspect which is not covered in this algorithm
rithms/schemes to improve overall performance across the is the additional handoff execution time caused by the intro-
entire cellular CDMA network, specialized in downlinks. We duction of the priority queuing. This can be important when
proposed a link-layer scheduling scheme MAPQ as a cross- signals from cells in the active set are dropping quite fast.
layer resource management issue for efficient resource alloca- (4) To import dynamic or adaptive guard capacity reser-
tion of underlying layer. Evaluation of the proposed scheme vation scheme to the proposed handoff prioritizing algo-
has been performed with a reference scheme and superiori- rithm would be an interesting topic.
ties are verified. It should be noted that this scheme can also (5) More results will be shown on the role of λ (the de-
be used for data scheduling with slight modification of wdp sign parameter that alters the prediction threshold in the pro-
(weight of delay over power) and QoS requirement of data posed handoff algorithm) and the capacity of base stations.
traffic. (6) For the TCP proposals, we will build up a larger scaled
Considering queuing delay as an important event trig- network where more subscribers and application sources will
ger and service indicator, together with required transmis- be generated for deeper understanding of the TCP behaviors.
sion power/rate, we also proposed an adaptive priority-based (7) In-depth study of the impacts of varying TCP/UDP
scheduling algorithm for unified voice/data frameworking traffic ratio on TCP behaviors in the integrated wired/wire-
and succeeded in fulfilling preset expectation through sim- less environment.
ulation.
By jointly considering downlink transmission power and
REFERENCES
call holding time, we proposed an adaptive prioritizing algo-
rithm to control concurrent handoff events to the same des- [1] A. J. Goldsmith and S. B. Wicker, “Design challenges for
tination. The performance improvement was obtained in the energy-constrained ad hoc wireless networks,” IEEE Wireless
simulation. Communications, vol. 9, no. 4, pp. 8–27, 2002.
At the transport layer, we studied its interaction with [2] L. Alonso and R. Agusti, “Automatic rate adaptation and
wireless link layer, particularly, wireless link scheduling. We energy-saving mechanisms based on cross-layer information
proposed an algorithm to avoid TCP spurious timeouts, to for packet-switched data networks,” IEEE Communications
regulate the behavior of TCP congestion window, and to en- Magazine, vol. 42, no. 3, pp. S15–S20, 2004.
[3] F. Yu and V. Krishnamurthy, “Cross-layer QoS provisioning in
hance the TCP throughput. These methods are claimed to
packet wireless CDMA networks,” in Proceedings of IEEE Inter-
be reasonable by theoretical analysis as well as the simula- national Conference on Communications (ICC ’05), vol. 5, pp.
tion verification. Extension of the simulation network and 3354–3358, Seoul, Korea, May 2005.
the traffic load scale to obtain greater performance gain of [4] J. Price and T. Javidi, “Cross-layer (MAC and transport) opti-
the proposed strategies is left to our future work. mal rate assignment in CDMA-based wireless broadband net-
Note that the complexity of our scheduling algorithms works,” in Proceedings of the 38th Asilomar Conference on Sig-
is determined fully by the real-time monitoring system. The nals, Systems and Computers, vol. 1, pp. 1044–1048, Pacific
computational part of the algorithms is simple. Once the pa- Grove, Calif, USA, November 2004.
rameters involved in the algorithms are procured from real- [5] V. Friderikos, L. Wang, and A. H. Aghvami, “TCP-aware power
time measuring, the rest of the work is just simple calcula- and rate adaptation in DS/CDMA networks,” IEE Proceedings:
tions. In our algorithms, power and delay are the key param-
[6] E. Hossain and V. K. Bhargava, “Cross-layer performance in
eters to be obtained. The downlink transmission power for cellular WCDMA/3G networks: modelling and analysis,” in
each subscriber is monitored and adjusted frequently to en- Proceedings of IEEE International Symposium on Personal, In-
sure the quality of service. The buffering delay is another im- door and Mobile Radio Communications (PIMRC ’04), vol. 1,
portant QoS measure, especially where the system resource pp. 437–443, Barcelona, Spain, September 2004.
needs to be allocated to support class-of-service. It is criti- [7] J. Yao, T. C. Wong, and Y. H. Chew, “Cross-layer design on
cal to ensure the delay bound for different classes. As men- the reverse and forward links capacities balancing in cellular
tioned above, wireless service providers are equipped with CDMA systems,” in Proceedings of IEEE Wireless Communi-
such monitoring system to make the procurement of these cations and Networking Conference (WCNC ’04), vol. 4, pp.
parameters in real time possible. 2004–2009, New Orleans, La, USA, March 2004.
Although the results obtained are what we have expected [8] Y. S. Chan, Y. Pei, Q. Qu, and J. W. Modestino, “On cross-layer
adaptivity and optimization for multimedia CDMA mobile
and are encouraging, there are some open issues and limita-
wireless networks,” in Proceedings of the 1st IEEE-EURASIP
tions of this work which call for deeper investigation. International Symposium on Control, Communications, and
Signal Processing (ISCCSP ’04), pp. 579–582, Hammamet,
Further Considerations Tunisia, March 2004.
[9] S. Singh, V. Krishnamurthy, and H. V. Poor, “Integrated
(1) The performance of the unified voice/data scheduling voice/data call admission control for wireless DS-CDMA sys-
tems with fading,” IEEE Transactions on Signal Processing,
framework was studied in a 19-cell layout. However, we can
vol. 50, no. 6, pp. 1483–1495, 2002.
further consider the effects of the load variations in the outer [10] C. Comaniciu and H. V. Poor, “Jointly optimal power and ad-
cells on the performance in a target cell. mission control for delay sensitive traffic in CDMA networks
with LMMSE receivers,” IEEE Transactions on Signal Process- [25] 3GPP TS 25.331, “RRC protocol specification,” 2000.
ing, vol. 51, no. 8, pp. 2031–2042, 2003. [26] D. Wong and T. J. Lim, “Soft handoffs in CDMA mobile sys-
[11] S. A. Ghorashi, L. Wang, F. Said, and A. H. Aghvami, “Im- tems,” IEEE Personal Communications, vol. 4, no. 6, pp. 6–17,
pact of macrocell-hotspot handover on cross-layer interfer- 1997.
ence in multi-layer W-CDMA networks,” in Proceedings of [27] 3GPP TS 25.214, “Physical layer procedures (FDD) v3.1.0,”
the 5th European Personal Mobile Communications Conference 1999.
(EPMCC ’03), pp. 580–584, Glasgow, UK, April 2003. [28] A. Viterbi, A. M. Viterbi, K. S. Gilhousen, and E. Zehavi, “Soft
[12] W. C. Jakes, Microwave Mobile Communications, John Wiley & handoff extends CDMA cell coverage and increases reverse
Sons, New York, NY, USA, 1993. link capacity,” IEEE Journal on Selected Areas in Communica-
[13] M. K. Karakayali, R. Yates, and L. Razumov, “Throughput tions, vol. 12, no. 8, pp. 1281–1288, 1994.
maximization on the downlink of a CDMA system,” in Pro- [29] K. M. Rege, S. Nanda, C. F. Weaver, and W.-C. Peng, “Analysis
ceedings of IEEE Wireless Communications and Networking of fade margins for soft and hard handoffs,” in Proceedings of
Conference (WCNC ’03), vol. 2, pp. 894–901, New Orleans, La, the 6th IEEE International Symposium on Personal, Indoor and
USA, March 2003. Mobile Radio Communications (PIMRC ’95), vol. 2, pp. 829–
[14] D. Zhao, X. Shen, and J. W. Mark, “Effect of soft handoff on 835, Toronto, Canada, September 1995.
packet transmissions in cellular CDMA downlinks,” in Pro- [30] J. Y. Sun, Y. F. Peng, and L. Zhao, “A novel packet scheduling
ceedings of the International Symposium on Parallel Architec- scheme based on adaptive power/delay for efficient resource
tures, Algorithms and Networks (I-SPAN ’04), pp. 42–47, Hong allocation in downlink CDMA systems,” in Proceedings of IEEE
Kong, May 2004. Canadian Conference on Electrical and Computer Engineering
[15] D. Kitazawa, L. Chen, H. Kayama, and N. Umeda, “Down- (CCECE ’05), Saskatoon, Canada, May 2005.
link packet-scheduling considering transmission power and [31] V. Jacobson, “Congestion avoidance and control,” in Proceed-
QoS in CDMA packet cellular systems,” in Proceedings of the ings of ACM SIGCOMM, pp. 273–288, Stanford, Calif, USA,
4th IEEE International Workshop on Mobile and Wireless Com- August 1988.
munications Network (MWCN ’02), pp. 183–187, Stockholm,
[32] P. Karn and C. Partridge, “Improving round-trip time esti-
Sweden, September 2002.
mates in reliable transport protocols,” ACM Transactions on
[16] W.-H. Sheen, I.-K. Fu, and K. Y. Lin, “New load-based re- Computer Systems, vol. 9, no. 4, pp. 364–373, 1991.
source allocation algorithms for packet scheduling in CDMA
[33] A. Bakre and B. R. Badrinath, “I-TCP: indirect TCP for mobile
uplink,” in Proceedings of IEEE Wireless Communications and
hosts,” in Proceedings of IEEE International Conference on Dis-
Networking Conference (WCNC ’04), vol. 4, pp. 2268–2273,
tributed Computing Systems (ICDCS ’95), pp. 136–146, Van-
New Orleans, La, USA, March 2004.
couver, BC, Canada, May-June 1995.
[17] D. M. Novakovic and M. L. Dukic, “Evolution of the power
[34] S. Floyd, “TCP and explicit congestion notification,” ACM
control techniques for DS-CDMA toward 3G wireless com-
Computer Communication Review, vol. 24, no. 5, pp. 10–23,
munication systems,” IEEE Communications Surveys and Tu-
1994.
torials, vol. 3, no. 4, fourth quarter, pp. 2–15, 2000.
[18] J.-H. Yoon, M.-J. Sheen, and S.-C. Park, “Scheduling methods [35] S. Lin and D. J. Costello, Error Control Coding: Fundamentals
with transmit power constraint for CDMA packet services,” in and Applications, Prentice-Hall, Englewood Cliffs, NJ, USA,
Proceedings of the 57th IEEE Semiannual Vehicular Technology 1983.
Conference (VTC ’03), vol. 2, pp. 1450–1453, Jeju, Korea, April [36] H. Balakrishnan, V. N. Padmanabhan, S. Seshan, and R. H.
2003. Katz, “A comparison of mechanisms for improving TCP per-
[19] X. Wang, R. Ramjee, and H. Viswanathan, “Adaptive and pre- formance over wireless links,” IEEE/ACM Transactions on Net-
dictive downlink resource management in next generation working, vol. 5, no. 6, pp. 756–769, 1997.
CDMA networks,” in Proceedings of the 23rd Annual Joint Con- [37] Y. Wu, Z. Niu, and J. Zheng, “A network-based solution for
ference of the IEEE Computer and Communications Societies TCP in wireless systems with opportunistic scheduling,” in
(INFOCOM ’04), vol. 4, pp. 2754–2765, Hong Kong, March Proceedings of IEEE International Symposium on Personal, In-
2004. door and Mobile Radio Communications (PIMRC ’04), vol. 2,
[20] J. W. Chang and D. K. Sung, “Adaptive channel reservation pp. 1241–1245, Barcelona, Spain, September 2004.
scheme for soft handoff in DS-CDMA cellular systems,” IEEE [38] H. Balakrishnan, S. Seshan, and R. H. Katz, “Improving re-
Transactions on Vehicular Technology, vol. 50, no. 2, pp. 341– liable transport and handoff performance in cellular wireless
353, 2001. networks,” ACM Wireless Networks, vol. 1, no. 4, pp. 469–482,
[21] J. Y. Sun, L. Zhao, and A. Anpalagan, “Soft handoff priori- 1995.
tizing algorithm for downlink call admission control of next- [39] T. E. Klein, K. K. Leung, and H. Zheng, “Improved TCP per-
generation cellular CDMA networks,” in Proceedings of 16th formance in wireless IP networks through enhanced oppor-
IEEE International Symposium on Personal, Indoor and Mo- tunistic scheduling algorithms,” in Proceedings of IEEE Global
bile Radio Communications (PIMRC ’05), Berlin, Germany, Telecommunications Conference (GLOBECOM ’04), vol. 5, pp.
September 2005. 2744–2748, Dallas, Tex, USA, November-December 2004.
[22] 3GPP TR 25.922 v3.6.0, “Radio resource management strate- [40] W. Stevens, “TCP Slow Start, Congestion Avoidance, Fast Re-
gies,” 2001. transmission, and Fast Recovery Algorithms,” RFC-2001, Jan-
[23] V. K. Garg, IS-95 CDMA and CDMA 2000: Cellular/PCs Sys- uary 1997.
tems Implementation, Prentice Hall, Englewood Cliffs, NJ, [41] “WCDMA (UMTS): FDD technical summary,” http://www.
USA, 1999. umtsworld.com/technology/wcdma.htm.
[24] A. Viterbi, CDMA Principles of Spread Spectrum Communica- [42] W.E.A 3GPP 2, “1xEV-DV evaluation methodology-adden-
tions, Addison-Wesley, Reading, Mass, USA, 1995. dum (v6),” 2001.
[43] F. Xu, M.-H. Ye, P. Zhao, and H.-M. Zhang, “The research characteristics, cross-layer resource optimization, and wireless sen-
on the service rate adaptation in mobile network,” in Proceed- sor networking. He has published more than 40 papers and articles
ings of International Conference on Communication Technology in international conferences and journals in his research areas. He
(ICCT ’03), vol. 2, pp. 970–976, Beijing, China, April 2003. currently serves as IEEE Toronto Section Chair, previously served
[44] M. Kazmi and N. Wiberg, “Power and rate assignment policies as Chair, Communications Chapter—IEEE Toronto Section (2004-
for best-effort services in WCDMA,” in Proceedings of the 13th 2005) and Technical Program Cochair of IEEE Canadian Confer-
IEEE International Symposium on Personal, Indoor and Mobile ence on Electrical and Computer Engineering (2004). His current
Radio Communications (PIMRC ’02), vol. 4, pp. 1601–1605, editorial duties include Guest Editor on special issue on Radio Re-
Lisbon, Portugal, September 2002. source Management in 3G+ Wireless Systems (2005-2006) and As-
[45] Y. Chen and L. G. Cuthbert, “Downlink performance of dif- sociate Editor in EURASIP Journal of Wireless Communications
ferent soft handover algorithms in 3G multi-service environ- and Networking. He is an IEEE Senior Member and a Registered
ments,” in Proceedings of the 4th IEEE International Workshop Professional Engineer in the province of Ontario, Canada.
on Mobile and Wireless Communications Network (MWCN
’02), pp. 406–410, Stockholm, Sweden, September 2002.
[46] Y. Chen and L. Cuthbert, “Optimum size of soft handover zone
in power-controlled UMTS downlink systems,” IEE Electronics
Letters, vol. 38, no. 2, pp. 89–90, 2002.
[47] “The network simulator - ns-2,” http://www.isi.edu/nsnam/ns.
Jin Yuan Sun received the M.A.S. degree

in computer networks from Ryerson Uni-
versity, Canada, in 2005. She received the
B.S. degree in computer information sys-
tems from Beijing Information Technology
Institute, China, in 2003. Since 2005, she has
been working as a Network Test Developer
at RuggedCom Inc., Ontario, Canada. Her
research interests are in wireless commu-
nications, computer networks, mobile net-
works, and sensor networks. She is a student Member of IEEE.
Lian Zhao received the Ph.D. degree in elec-

University of Waterloo, Canada, in 2002.
Before she joined Ryerson University in
2003, she worked as a postdoctoral fellow
with the Center for Wireless Communica-
tions, University of Waterloo. Since 2003,
she has been working as an Assistant Pro-
fessor at the Department of Electrical and
Computer Engineering, Ryerson University,
Toronto, Canada. She is a cofounder of the Optic Fiber Sensing
Wireless Network Laboratory in 2004. Her research interests are in
the areas of wireless communications, radio resource management,
power control, as well as design and applications of the energy ef-
ficient wireless sensor networks. She is an IEEE Senior Member
and a Registered Professional Engineer in the province of Ontario,
Canada.
Alagan Anpalagan received the B.A.S.,
M.A.S., and Ph.D. degrees in electrical en-
gineering from the University of Toronto,
Canada in 1995, 1997, and 2001, respec-
tively. Since August 2001, he has been with
the Ryerson University, Toronto, Canada,
where he cofounded WINCORE laboratory
in 2002 and has been leading the WAN
(wireless access and networking) R&D
Group. Currently, he is an Associate Profes-
sor and Program Director for Graduate Studies. His research inter-
ests are, in general, wireless communication, mobile networks, and
system performance analysis; and in particular, QoS-aware radio
resource management, joint study of wireless physical/link-layer
DOI 10.1155/WCN/2006/60349
On Cross-Layer Design for Streaming Video Delivery

in Multiuser Wireless Environments
Lai-U Choi,1 Wolfgang Kellerer,2 and Eckehard Steinbach1

1 Media Technology Group, Institute of Communication Networks, Department of Electrical Engineering and Information Technology,
Munich University of Technology, 80290 Munich, Germany
2 Future Networking Lab, DoCoMo Communications Laboratories Europe GmbH, 80687 Munich, Germany

We exploit the interlayer coupling of a cross-layer design concept for streaming video delivery in a multiuser wireless environment.
We propose a cross-layer optimization between application layer, data link layer, and physical layer. Our aim is to optimize the
end-to-end quality of the wireless streaming video application as well as efficiently utilizing the wireless resources. A possible
architecture for achieving this goal is proposed and formulated. This architecture consists of the process of parameter abstraction,
a cross-layer optimizer, and the process of decision distribution. In addition, numerical results obtained with different operating
modes are provided. The results demonstrate the potential of this proposed joint optimization.
Copyright © 2006 Lai-U Choi et al. This is an open access article distributed under the Creative Commons Attribution License,
1. INTRODUCTION On the other hand, the introduction of independent lay-

ers has proven to be a robust and efficient design approach,
Since the introduction of digital personal wireless networks and has served extremely well in the development and im-
around 1990, wireless communication has evolved from an plementation of both past and current communication sys-
add-on into the key business of large telecommunication tems. The interlayer dependencies which are introduced by
companies. At the beginning of the 21st century, personal the proposed cross-layer design should therefore be kept to a
wireless communication has become part of the daily life of minimum, to preserve the layered structure as much as pos-
most people in developed areas. Together with the daily life sible. It is important that cross-layer design does not run at
usage, the service provided by the telecommunication com- cross-purposes with sound and long-term architectural prin-
panies is evolving from voice-based telephony to more de- ciples of existing communication systems [1].
manding multimedia service, including email, web browsing, In this paper, we exploit the interlayer coupling of a cross-
database access, video on demand, video conferencing, re- layer design concept for streaming video delivery in a multi-
mote sensing, and medical applications. Multimedia services user wireless environment. We focus on a cross-layer opti-
require much higher data rates than voice-centered service mization between application layer, data link layer, and phys-
and they make the design of future wireless communication ical layer. Our aim is to optimize the end-to-end quality of
networks ever more challenging. the wireless streaming video application as well as efficiently
Cross-layer design was proposed to address those chal- utilizing the wireless resources. To achieve this aim, an ar-
lenges. The concept of cross-layer design introduces inter- chitecture for the joint layer optimization is proposed, which
layer coupling across the protocol stack and allows the ex- provides a potential solution for the implementation of the
change of necessary information between different layers. cross-layer optimization concept. This architecture does not
Although this concept can be employed in all communi- require a redesign of the existing protocols, but may require
cation networks, it is especially important in wireless net- extra modules to implement the function of the joint opti-
works because of the unique challenges of the wireless envi- mization.
ronment, like the time-varying and the fading nature of the The proposed architecture is general and consists of the
wireless channels. This wireless nature and user mobility lead process of parameter abstraction, a cross-layer optimizer, and
to random variation in network performance and connectiv- the process of decision distribution. It is designed with the
ity. goal of increasing compatibility and stability, and the goal of
Streaming
Base
client 2 Parameter abstraction
Streaming station
client 1 Application
n
layer
io
ut
r ib
ist
D
Cross-layer Transport layer
Streaming Decision
client 3 optimizer
Streaming Network layer
D
server
ist
r ib
ut
Radio link
io
n
layer
Parameter abstraction (MAC + PHY)
Figure 1: Streaming video server and mobile clients in a wireless Figure 2: Proposed system architecture: parameter abstraction,
multiuser environment. cross-layer optimization, and decision distribution.
reducing the signaling overhead. Every part in this architec- (3) simulation results which show the possible gains that
ture is formalized and its performance potential is demon- could be achieved with the proposed optimization architec-
strated by sample numerical results. ture and scheme.
An important issue in cross-layer design is the amount of The structure of this paper is as follows. In Section 2,
the required information exchange between the layers and the system architecture under consideration is introduced.
the time scale at which the optimization is performed. In Then, Sections 3, 4, and 5 present the formalism of the three
general, the lower the amount of information exchange and components in the proposed optimization architecture, re-
the longer the time-scales are, the more robust and imple- spectively. We provide numerical results in Section 6 which
mentable the design becomes. In this way, our proposed demonstrate the potential of the proposed joint optimiza-
cross-layer optimization interacts with the lower layers (data tion. Finally, we conclude our work and discuss some further
link layer and physical layer) on a long-term basis. This long- research in Section 7.
term approach can be extended also to the higher layers as
shown in [2, 3]. This long-term approach has recently been 2. SYSTEM ARCHITECTURE
successfully applied in [4].
We consider a video streaming server located at the base sta-
There is plenty of research activity currently going on in
tion1 and multiple streaming clients located in mobile de-
the field of cross-layer design focusing on the interaction be-
vices. As shown in Figure 1, streaming clients or users are
tween physical, data link, and higher layers, sometimes also
assumed to be sharing the same air interface and network
including the application layer. A review of some of these cur-
resources but requesting different video contents. Note that
rent research activities can be found in [5, 6].
only the protocol stack necessary for the wireless connection
In this paper, we focus on the joint optimization of three has to be considered since in our scenario the video stream-
layers in the protocol stack, namely the application layer ing server is located directly at the base station. Therefore, the
(layer 7), the data link layer (layer 2), and the physical layer transport layer and the network layer in the protocol stack
(layer 1). We include the application layer in the joint op- can be excluded from our optimization problem. We focus
timization because the end-to-end quality observed by the on the interaction between the application layer and the ra-
users directly depends on the application and the application dio link layer, which incorporates both the physical (PHY)
layer has firsthand information about the impact of each suc- layer and the data link layer.
cessfully decoded piece of media data on the perceived qual- At the base station, an architecture as shown in Figure 2 is
ity. We also include the physical layer and the data link layer proposed to provide end-to-end quality-of-service optimiza-
in our consideration because the unique challenge of mobile tion. This figure illustrates information flows and the tasks
wireless communication results from the nature of the wire- required for the joint optimization. The tasks can be split into
less channel, which these two layers have to cope with. The three main subtasks.
main contribution of this work includes the following: (1) Parameter abstraction: necessary state information is
(1) possible architecture for cross-layer optimization which collected from the application layer and the radio link layer
provides a potential solution of joint optimization of the
physical, data link, and application layer;
(2) mathematical description of the proposed architecture 1 Alternatively, we assume a proxy server is installed at the base station, in
and optimization; case the streaming server is remote.
Lai-U Choi et al. 3
through the process of parameter abstraction. The process of processing), while the data link layer is responsible for ra-
parameter abstraction results in the transformation of layer- dio resource allocation (multiuser scheduling or queuing)
specific parameters into parameters that are comprehensible and error control (by channel coding, usually a combina-
for the cross-layer optimizer, so-called cross-layer parame- tion of forward error-correction coding (FEC) and automatic
ters. retransmission (ARQ)). Since both of these two layers are
(2) Cross-layer optimization: the optimization is carried closely related to the unique characteristics of the wireless na-
out by the cross-layer optimizer with respect to a particular ture, it is useful to consider them together. In the following,
objective function. From a given set of possible cross-layer we refer to their combination as the radio link layer.
parameter tuples, the tuple optimizing the objective function The application layer is the layer where the media data
is selected. is compressed, packetized, and scheduled for transmission.
(3) Decision distribution: after the decision on a particular The key parameters to be abstracted for the cross-layer opti-
cross-layer parameter tuple is made, the optimizer distributes mization are related to the characteristics of the compressed
the decision information back to the corresponding layers. source data. This implies that these abstracted key param-
Note that an excellent discussion of other architec- eters may depend on the type of application or service be-
tures, so-called top-down and bottom-up approaches, can be cause the characteristics of the compressed source data may
found in [3]. In the following, the necessity and the details of depend on the application or service. In this paper, we con-
the parameter abstraction will be provided in Section 3, while sider a video streaming service application.
the cross-layer optimization and decision distribution are cov-
ered in Sections 4 and 5, respectively. 3.1. Data link layer and physical layer parameters
3. PARAMETER ABSTRACTION To formalize the data link layer and physical layer parameter
abstractions, we follow the approach proposed in [7, 8] and
define the set
In order to carry out the joint optimization, state informa-
tion or a set of key parameters have to be abstracted from
the selected layers and provided to the cross-layer optimizer. R = r1 , r2 , . . . (1)
This is necessary because the direct exchange of layer-specific
parameters may be difficult because of the following reasons. of tuples ri = (ri1 , ri2 , . . .) of radio-link-layer-specific parame-
j
(1) Compatibility: layer-specific parameters may easily be ters ri (e.g., modulation alphabets, code rate, airtime, trans-
incomprehensible or of no use for other layers. For instance, mit power, decorrelation time). Since these radio-link-layer-
a fading correlation matrix which is meaningful at the PHY specific parameters may be variable, the set R contains all
layer may well have no meaning at any of the higher layers. possible combinations of their values and each tuple ri repre-
Its influence on system performance therefore has to be ab- sents one possible combination. In this way, R can be an in-
stracted into a form which is meaningful for the other layers finite, countably infinite, or finite set, depending on the dis-
involved in the cross-layer optimization. crete or continuous nature of the parameter tuples. In order
(2) Signaling overhead: cross-layer design requires addi- to formalize parameter abstraction, we define the set
tional signaling between the layers, which produces access

delays. A reduction of the number of parameters which needs = r1 , r2 , . . .
R (2)
to be exchanged is therefore most welcomed. Abstraction
of layer parameters can help in achieving this reduction by of tuples ri = (ri1 , ri2 , . . .) of abstracted parameters. The re-
mapping several layer parameters into just a few abstracted lationship between the set R of all possible radio link layer
parameters. of all possible abstracted ra-
parameter tuples and the set R
(3) Stability: cross-layer design introduces coupling be-
dio link layer parameter tuples is established by the relation
tween otherwise independent layers. Because of the latency
time required in interlayer signaling, the system may become

G⊆R×R (3)
instable. Abstraction of layer parameters can facilitate stabil-
ity analysis as a consequence of the reduction of signaling
overhead and the increase of compatibility. The number of with domain R and codomain R. Here, the symbol × refers
the parameters is reduced and their influence on the individ- to the Cartesian product. The relation G is a subset of R × R
ual layer performance may be better understood than those
that defines the mapping between R and R. That is, only and
of the original layer parameters. all valid pairs (ri , r j ) are elements of G. We call this mapping
In wireless networks, the physical layer and the data link process the radio link layer parameter abstraction.
layer are dedicatedly designed for the dynamic variation of Let us look at an example. In a single-user scenario, we
the wireless channel during the provision of a particular ser- could, for example, abstract four key parameters: transmis-
vice. This is in contrast to wireline networks which experi- sion data rate d, transmission packet error ratio e, data packet
ence much less dynamic variation. The physical layer deals size s, and the channel decorrelation time t. This leads to
with the issues including transmit power (through transmit the abstracted parameter tuple ri = (di , ei , si , ti ). In a K user
power control), channel estimation, synchronization, signal scenario, one can extend the parameter abstraction for each
shaping, modulation and signal detection (through signal user. The parameter tuple ri then contains 4K parameters,
p 900
800
1 p G B 1 q
700
q
600
Distortion Di
500
Figure 3: A two-state Markov channel model.
400
(1) (1) (1) (1) (K) 300

, ei(K) , s(K)
ri = (di , ei , si , ti , . . . , di
(K)
i , ti ), in which a
group of four parameters belongs to one user. The trans- 200
mission data rate d is influenced by the modulation scheme,
100
the code rate of the used channel code, and the multi-user
scheduling. The transmission packet error ratio e is influ- 0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
enced by the transmit power, channel estimation, signal de-
Index i
tection, modulation scheme, channel coding, the current
user position, and so forth, The channel decorrelation time t Foreman
of a user is related to the user’s velocity and its surrounding Carphone
environment, while the data packet size s is normally defined Mother-daughter
by the wireless system standard. These interrelationships de-
fine the relation G from (3). A detailed discussion of the re- Figure 4: Measured loss distortion profile for a GOP in three video
lation G can be found in [2]. sequences.
Alternatively, it is possible to transform the transmission
packet error ratio e and the channel decorrelation time t into
the two parameters of a two-state Markov model as shown with domain A and codomain A. The relation F is a subset
in Figure 3, which are the transition probabilities (p and q)
of A × A that defines the mapping between A and A. That is,
from one state to another. In Figure 3, the states G and B rep- only and all valid pairs (ai , a j ) are elements of F. We call this
resent the good and bad states, respectively. The transforma- mapping process the application layer parameter abstraction.
tion is given by [9] as In this paper, we assume a streaming video service. The
abstracted parameters of this service include the source data
es (1 − e)s rate, the number of frames (or pictures) per second, size (in
p= , q= , (4)
td td terms of bytes), and maximum delay of each frame (or pic-
ture). Other important information for the optimizer is the
where p is the transition probability from the good to the bad distortion-rate function (encoding distortion) and the so-
state and q is the transition probability from the bad to the called loss distortion profile, which shows the distortion Di
good state. In this way, the abstracted parameter tuples take that is introduced in case the ith frame of the GOP is lost.
on the form ri = (di(1) , s(1) (1) (1) (K) (K) (K) (K)
i , pi , qi , . . . , di , si , pi , qi ). Figure 4 shows an example of the loss distortion profile of
One advantage of this transformation is that the resulting pa- lost frames for three different video sequences. This profile
rameter tuple is more comprehensible for high layers in the is generated from a group of picture (GOP) with 15 frames,
protocol stack. starting with an independently decodable intraframe and fol-
lowed by 14 interframes. The interframes can only be suc-
3.2. Application layer parameters cessfully decoded if all previous frames of the same GOP are
decoded error-free. The index in Figure 4 indicates the loss
Similar to the parameter abstraction in Section 3.1, for a for- of a particular frame, while the distortion Di is quantified
mal description, let us define the set by the mean-squared reconstruction error (MSE), which is
measured between the displayed and the transmission error-

A = a1 , a2 , . . . , (5) free decoded video sequence. It is assumed that as part of
the error concealment strategy, all the following frames of
of tuples ai = (a1i , a2i , . . . , ) of application-layer-specific pa- the GOP are not decodable and the most recent correctly de-
j coded frame is displayed instead of the nondecoded frames
rameters ai . Since these application-layer-specific parame- (copy the previous frame error concealment).
ters may be variable, the set A contains all possible combi-
nations of their values and each tuple ai represents one pos-
sible combination. We further define the set A = {a1 , a2 , . . .} 3.3. Cross-layer parameters
1 2 j
of tuples ai = (ai , ai , . . .) of abstracted parameters ai . The The abstracted parameter sets (R and A)
from both the ap-
relationship between A and A is established by the relation plication layer and the radio link layer form the input to
the cross-layer optimizer. Since any combination of the ab-

F⊆A×A (6) stracted parameter tuples from the two input sets is valid, it
Lai-U Choi et al. 5
is convenient to define the cross-layer parameter set where Pi is the probability that the ith frame is the first frame
lost during transmission of this GOP and Di is the mean-
X ×A
=R (7) square error that is introduced by this loss. Note that the
Di is taken from the measured distortion profile and is usu-
which combines the two input sets into one input set for ally different for each GOP. Figure 4 shows an example dis-
the optimizer. The set X = { x1 , x2 , . . .} consists of tuples tortion profile. The Pi can be computed from the 2-state
xn = (ri , a j ). Note that the cardinality of the set X grows Markov model as shown in Figure 3. For details, we refer to
exponentially with the number of cross-layer parameters.2 [2, 10, 11].
This means that the complexity of the cross-layer optimiza- For a multiuser situation, different extensions of the MSE
tion grows exponentially with the number of cross-layer pa- are possible. For example, the objective function can be the
rameters. sum of MSE of all the users. That is,
4. THE CROSS-LAYER OPTIMIZER

K
Γ(x) = MSEk (x), (14)
k=1
With the formalism introduced in Section 3, the operation of
the cross-layer optimizer Ω can now be described by where MSEk (x) is the MSE of user k for the cross-layer pa-
This objective function will optimize
rameter tuple x ∈ X.
Ω:X ⊂ X.
−→ Y (8)
the average performance in terms of MSE among all users.
of all possible ab- Another common definition of the objective function is
The optimizer gets as input the set X
stracted cross-layer parameter tuples and returns a true non- Γ(x) = max MSEk (x) (15)
empty subset Y as its output. In the following, we assume k∈{1,2,...,K }
| = 1, that is, the output of the optimizer is a single
that |Y which ensures that the MSE is minimized with the constraint
tuple and that all users obtain the same MSE.3 Yet another definition

= x
Y
opt ∈ X. (9)
K
Γ(x) = MSEk (x) (16)
The decision or output xopt of the cross-layer optimizer is k=1
made with respect to a particular objective function
leads to a maximization of the average PSNR among all users.
−→ R,
Γ:X (10)
5. DECISION DISTRIBUTION
where R is the set of real numbers. Therefore, the output of
the optimizer can be expressed as Once the output xopt = (ropt , aopt ) of the cross-layer optimizer
is obtained, the decisions ropt ∈ R and have to be
aopt ∈ A
xopt = arg min Γ x
. (11)

x∈X communicated back to the radio link layer and the applica-
tion layer, respectively. During this, the process of parameter
Notice that because X is a finite set, the optimization abstraction has to be reversed and the abstracted parameters
(11) is performed by exhaustive search guaranteeing the ropt and aopt have to be transformed back to the layer-specific
global optimal solution. The choice of a particular objective parameters ropt ∈ R and aopt ∈ A. This reverse transforma-
function Γ depends on the goal of the system design, and the tion is given by
output (or decision) of the optimizer might be different for
different objective functions. In the example application of ropt ∈ r | r, ropt ∈ G ,
(17)
streaming video, one possible objective function in a single- aopt ∈ a | a, aopt ∈ F .
user scenario is the MSE between the displayed and the orig-
inal video sequence, that is, the sum of loss distortion MSEL In case that {r | (r, ropt ) ∈ G} or {a | (a, aopt ) ∈ F} has more
and source distortion MSES : than one element, the choice of particular elements ropt and
aopt , respectively, can be made at the corresponding layers in-
MSE = MSES + MSEL , (12) dividually.
where MSEL can be computed from the distortion profile by
6. SAMPLE NUMERICAL RESULTS

15
MSEL = Di Pi , (13) In this section, we provide sample simulation results to eval-
i=1 uate the performance of the proposed joint optimization. We
2 For instance, assume that all, say n, cross-layer parameters are quantized 3 In practice, some or all of the cross-layer parameters may only take on
be-
to a fixed number, say q, of values. Then the cardinality of the set X values from a finite set. The resulting granularity in general leads to not
comes qn , which shows exponential growth in the number of cross-layer all users having the same quality of service as would be the case if all pa-
parameters. rameters were continuously adjustable.
Table 1: Multiuser scheduling: TDMA airtime assignment. Table 2: Resulting transmission data rates in kbps for each user.
Case → 1 2 3 4 5 6 7 Case → 1 2 3 4 5 6 7 Mod.
User 1 3/9 4/9 4/9 3/9 2/9 3/9 2/9 User 1 150 200 200 150 100 150 100
User 2 3/9 3/9 2/9 4/9 4/9 2/9 3/9 User 2 150 150 100 200 200 100 150 BPSK
User 3 3/9 2/9 3/9 2/9 3/9 4/9 4/9 User 3 150 100 150 100 150 200 200
Case → 8 9 10 11 12 13 14 Mod.
assume K = 3 users or clients (users 1, 2, and 3), each of User 1 300 400 400 300 200 300 200
which requests a different video sequence. We assume that User 2 300 300 200 400 400 200 300 QPSK
users 1, 2, and 3 request the carphone (CP), foreman (FM),
and mother & daughter (MD) video test sequence, respec- User 3 300 200 300 200 300 400 400
tively.4
6.1. Objective function
We choose the peak-signal-to-noise ratio (PSNR) as our per- of time arrangement in a time-division multiplexing-based
formance measure. The PSNR is defined as multiuser scheduling as shown in Table 1. A user’s transmis-
sion data rate is assumed to be equal to 100 kbps provided

that BPSK is used and 2/9 of the total transmission time is
2552 assigned to it. Therefore, if QPSK is used and 4/9 of the total
PSNR = 10 · log10 . (18)
MSE transmission time is assigned, the user can have a transmis-
sion data rate as high as 400 kbps. Table 2 shows the resulting
The larger the PSNR is, the smaller the MSE is, which is com- transmission rate for each user as a function of the time ar-
puted between the original video sequence and the recon- rangement and modulation scheme (BPSK or QPSK).
structed sequence at the client or user. Therefore, the larger The transmission error rate on the other hand depends
the PSNR is, the better the performance is. As an example, we on the transmission data rate, the average SNR, and the
use the objective function given in (15), which maximizes the error-correcting capability of the channel code. Usually, the
worst-case user’s performance. Therefore, the cross-layer op- performance of a channel code is evaluated in terms of the
timizer chooses the parameter tuple that minimizes the max- residual error rate (after channel decoding) for a given re-
imum of MSE (or equivalently maximizes the minimum of ceive SNR. In our simulation, we assume a convolutional
the PSNR) among the users. This leads to all users having the code of code rate 1/2 and a data packet size of 432 bits. The
same PSNR. However, the PSNR may nevertheless come out residual packet error ratio is shown in Figure 5(a) as a func-
different for each user because of granularity of the cross- tion of SNR [12]. However, in the wireless link, the receive
layer parameters (see footnote 3). SNR is not constant, but is fluctuating around the mean
value (long-term SNR), which is due to fast fading caused
by user mobility. In this way, the receive SNR can be mod-
6.2. Physical layer and data link layer parameters
eled as a random variable with a certain probability distri-
In the simulation, it is assumed that the data packet size s at bution, which is determined by the propagation property of
the radio link layer equals 432 bits, which is the same as the the physical channel (e.g., Rayleigh distribution, Rice distri-
specified packet size of the IEEE802.11a or HiperLAN2 stan- bution). The residual packet error rate in a fading wireless
dard [12]. The channel decorrelation time t is assumed to be link is computed by averaging this packet error ratio (e.g.,
50 milliseconds for all the three users, which corresponds to a taken from Figure 5(a)) with the fading statistics. Assum-
pedestrian speed (about 2 Km/h at 5 GHz carrier frequency). ing Rayleigh fading, the resulting average packet error rate
Since the transmission data rate d is influenced by the is given in Figure 5(b) as a function of the average signal-to-
modulation scheme, the channel coding, and the multiuser noise ratio SNR. This resulting average packet error ratio is
scheduling, two different modulations (BPSK and QPSK) used as the parameter e in (4) in our simulation.
are assumed. It is further assumed that there are 7 cases User’s position-dependent path loss and shadowing com-
monly observed in wireless links are taken into account by
choosing the long-term average SNR randomly and indepen-
4 We have chosen these particular video test sequences as they emphasize dently for each user uniformly within the range from 1 to 100
different situations in a real-world video sequence. FM contains a scene (0 dB to 20 dB).
change with rather quick camera movement, MD has no camera move- In summary, the abstracted parameters, namely date rate
ment or scene change, while CP has a quickly moving background ac- di , packet size si , and Markov model parameters (pi , qi ) for
companied by medium foreground movement. These situations typically
occur in real-life video sequences and lead to rather different properties
each user and each of the 7 or 14 cases of modulation and
of the encoded data streams, especially bit sizes of frames and sensitivity TDMA scheduling scheme (according to Table 1 or 2, resp.),
to frame losses. have to be communicated to the cross-layer optimizer.
Lai-U Choi et al. 7
100 100
Average packet error rate

10 1
Packet error rate
10 1
10 2
10 3 10 2
0 5 10 15 0 5 10 15 20
SNR (dB) Average SNR (dB)
BPSK BPSK
QPSK QPSK
(a) (b)
Figure 5: Example decoding error performances of a convolutional code with different modulations in an AWGN and a Rayleigh fading
channels: (a) packet error ratio after channel decoding as a function of the signal-to-noise ratio (SNR) in an AWGN channel [12]; (b) packet
error ratio after channel decoding as a function of the average signal-to-noise ratio SNR in a Rayleigh fading channel.
6.3. Application layer parameters shown Table 3 for each user, have to be communicated to the
cross-layer optimizer.
At the application layer, it is assumed that the video is en-
coded using the H.264/AV C [13] video compression stan-
6.4. Operating modes
dard with 30 frames per second and 15 frames per GOP (i.e.,
0.5-second GOP duration). Two different values of the source An operation mode without ARQ (referred to as forward
rate (100 kbps and 200 kbps) are considered. To this end, the mode) and an operation mode with ARQ (referred to as ARQ
video has been pre-encoded at these two different target rates mode) are investigated. We consider every GOP as a unit and
and both versions are stored on the streaming server. We can assume that each GOP has to be transmitted within the du-
switch from one source stream to the other at the beginning ration of 0.5 second.
of any GOP. In each GOP, the first frame is an I-frame and (i) Forward mode: we assume no acknowledgment from
the following 14 frames are P-frames. We use the measured the clients is available and the video frames of every GOP
distortion profile of a particular lost frame and the encod- for a particular client are repeatedly transmitted when the
ing distortion for the 3 requested videos. Figure 4 shows an transmission data rate is larger than the source data rate. For
example of a distortion profile in terms of MSE for a GOP instance, every GOP is transmitted twice if the transmission
at a source rate of 100 kbps. Also, note that since successful data rate is twice as large as the source data rate. If the trans-
decoding of P-frames depends on error-free reception of all mission data rate is 1.5 times the source data rate, a GOP is
previous frames of the same GOP, losing the first frame of transmitted once followed by retransmitting the I-frame, the
a GOP leads to the largest distortion, while losing the last first P-frame, the second P-frame, and so forth, until the pe-
frame of a GOP leads to the least distortion. Furthermore, riod of 0.5 second for the GOP is expired.
it is assumed that each video frame (or picture) is packe- (ii) ARQ mode: here we assume that instantaneous ac-
tized with maximum size of 432 bits and each packet only knowledgment of a transmitted packet is available from the
contains data from one frame. The size of each frame is de- clients and the data packets of every GOP for a particular
termined during the H.264/AVC encoding. These values are client are retransmitted in the way that the data packets in a
stored along with the bit stream and the distortion profile as GOP are received successfully in time order. That is, before
well as the value of the source distortion. Table 3 gives the transmitting a new packet, it is guaranteed that its previous
measured size (in terms of packets) for a GOP in the three packets in the GOP are received correctly.
video sequences at a source rate of 100 kbps, where I and Pn In the following, both modes of operation will be inves-
(n = 1, 2, . . . , 14) denote the I-frame and the nth P-frame, tigated.
respectively. We can see that the size of an I-frame is much
larger than that of a P-frame and the size of a P-frame varies 6.5. Simulation results and discussion
from frame to frame. This is related to the contents of a video.
In summary, the abstracted parameters, namely the loss Figures 6 and 7 provide simulation results of the following
distortion profile as shown in Figure 4 and the frame sizes as three scenarios.
Table 3: Measured sizes (in number of packets) of the encoded frames of a GOP for three different video sequences at 100 kbps.
Frame →
I P1 P2 P3 P4 P5 P6 P7 P8 P9 P10 P11 P12 P13 P14
Sequence ↓
Carphone 43 7 7 7 6 5 8 6 7 7 6 6 4 5 5
Foreman 47 5 6 7 5 6 7 6 5 5 6 5 5 3 4
Mother & daughter 50 1 2 3 3 3 4 4 4 5 6 8 10 12 14
1 1
Cumulative probability density function (CDF)

0.9 0.9
0.8 0.8
0.7 0.7
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
20 22 24 26 28 30 32 34 36 38 20 22 24 26 28 30 32 34 36 38
PSNR of the worst performing user (dB) PSNR of the worst performing user (dB)
Forward mode w/oJO ARQ mode w/oJO Forward mode w/oJO ARQ mode w/oJO
Forward mode w/JO ARQ mode w/JO Forward mode w/JO ARQ mode w/JO
(a) (b)
Figure 6: Cumulative probability density function (CDF) of the PSNR of the worst performing user: (a) results for scenario 1, BPSK
modulation, and source rate of 100 kbps; (b) results for scenario 2, BPSK/QPSK modulation, and source rate of 100 kbps.
(1) Scenario 1: we restrict ourselves that only BPSK mod- user’s position-dependent path loss and shadowing. In gen-
ulation is used at the radio link layer and only the source rate eral, fast fading takes place on a much smaller time scale
with 100 kbps is available at the application layer. Therefore, than the path loss and shadowing. In this paper, we eval-
only one constant abstracted parameter tuple (with 100 kbps uate the MSE averaged over fast fading by taking the ex-
for all 3 users) is provided by the application layer (i.e., pected value of the MSE with respect to the fast fading
|A | = 1) in this scenario, while the radio link layer provides for a particular position of the users or equivalently for a
7 abstracted parameter tuples (|R | = 7), which result from particular long-term SNR. Based on this value, the cross-
the 7 cases of time arrangement shown in Table 1. The cross- layer optimizer makes its decision. We also look at its sta-
layer optimizer selects one out of the 7 combinations of the tistical properties for an ensemble of user positions. There-
input parameter tuples (|X | = |R
| · |A
| = 7) such that our fore, the cumulative probability density function (CDF) of
objective function given in (15) is optimized. this average MSE is chosen to show the performance of
both modes (forward mode and ARQ mode). The perfor-
(2) Scenario 2: the same abstracted parameter tuple as in
mance of the worst performing user in the system with the
scenario 1 is assumed at the application layer but the radio
proposed joint optimization (w/JO) is compared with that
link layer provides 14 abstracted parameter tuples, which re-
in a system without joint optimization (w/oJO). A system
sult from the 7 cases of time arrangement with BPSK and
without joint optimization is assumed to assign the same
another 7 cases of time arrangement with QPSK.
amount of transmission time to all the users (i.e., Case 1 in
(3) Scenario 3: it is assumed that the two different source Table 1) and use BPSK modulation, while the source data
rates of 100 kbps and 200 kbps for each of the 3 users are pro- rate is fixed to 100 kbps. It can be seen from Figure 6(a)
vided by the application layer. This results in |A | = 23 = 8 that the PSNR of the worst performing user improves sig-
abstracted parameter tuples from the application layer. The nificantly in the system w/JO. For instance, there is about
same 14 abstracted parameter tuples as in scenario 2 are pro- 1 − 40% = 60% chance that the PSNR of the worst perform-
vided by the radio link layer. ing user is larger than 30 dB in the system w/JO in forward
The distortion MSE given in (12) is a random vari- mode, which improves to 2 dB when compared to the system
able controlled by the two factors, namely fast fading and w/oJO.
Lai-U Choi et al. 9
1 1

1
0.9 0.9
0.9
0.8 0.8
0.8
0.7 0.7
0.7
0.6 0.6
0.6
0.5 0.5
0.5
0.4 0.4
0.4
0.3 0.3
0.3
0.2 0.2
0.2
0.1 0.1
0.1 0 2 4 6 8 0 2 4 6 8
20 22 24 26 28 30 32 34 36 38
Δ PSNR (dB) Δ PSNR (dB)
PSNR of the worst performing user (dB)
Scenario 1 Scenario 1
Forward mode w/oJO ARQ mode w/oJO Scenario 2 Scenario 2
Forward mode w/JO ARQ mode w/JO Scenario 3 Scenario 3
(a) (b) (c)
Figure 7: (a) Cumulative probability density function (CDF) of the PSNR of the worst performing user for scenario 3, BPSK/QPSK mod-
ulation and source rate of 100 kbps/200 kbps; (b) performance improvement for the three scenarios in forward mode; (c) performance
improvement in ARQ mode.
A similar trend of improvement can be observed in optimization with three principle concepts, namely param-
Figure 6(b) and Figure 7(a) for scenarios 2 and 3, respec- eter abstraction, cross-layer optimization, and decision dis-
tively. The performance improves when more abstracted pa- tribution. Although we have focused on the application layer
rameter tuples are provided because more degrees of free- and radio link layer in a wireless system with a video stream-
dom can be obtained. This can be observed in Figure 7(b) ing service, this architecture can be easily generalized for dif-
and Figure 7(c) more clearly, where the performance im- ferent layers and different services. Our study reveals that
provement of the three investigated scenarios is shown. Here, this proposed architecture can provide a potential way to
ΔPSNR is defined as the difference between the PSNR of improve the performance and therefore help dealing with
the worst performing user in the system w/JO and that in the future challenges in wireless multimedia communica-
the system w/oJO. A close observation of Figure 7(b) reveals tion. Even when considering a small number of degrees of
that the amount of performance improvement of scenario freedom of the application layer and the radio link layer, we
2 is much larger than that of scenario 1 in forward mode, obtain significant improvements in user-perceived quality of
while the amount of performance improvement of scenario our streaming video application by joint optimization. Note
3 is only slightly larger than that of scenario 2. This indicates that we only consider the wireless hop in this study. Further
that the choice of higher transmission data rate (by using sophisticated research might be required in order to exploit
QPSK) provided by the radio link layer is favorable in for- this cross-layer design concept more completely. This work
ward mode, and the optimizer chooses it frequently. In con- has been partially presented at ICIP’04 [14].
trast, the choice of higher source rate (200 kbps) provided by
the application layer is not so favorable in this mode and the ACKNOWLEDGMENTS
optimizer seldom chooses it. On the other hand, this choice
The authors would like to thank the DoCoMo Communica-
of a higher source rate is favorable in ARQ mode, which can
tion Laboratories Europe GmbH, Munich, and the Alexan-
be seen from the graph in Figure 7(b), where the amount of
der von Humboldt Foundation (AvH) for kindly supporting
performance improvement of scenario 3 is fairly larger than
this research and thank Dr. Michel T. Ivrlač for very valuable
that of scenario 2. Therefore, choosing a suitable set of ab-
input and discussion.
stracted parameters tuples is important in order to obtain
large performance improvements while optimizing at low
complexity. REFERENCES
[1] V. Kawadia and P. R. Kumar, “A cautionary perspective on
7. CONCLUSION AND OUTLOOK cross-layer design,” IEEE Wireless Communications, vol. 12,
no. 1, pp. 3–11, 2005.
We have exploited the interlayer coupling of a cross-layer [2] L. Choi, M. T. Ivrlač, E. Steinbach, and J. A. Nossek, “Bottom-
design concept and proposed an architecture for the joint up approach to cross-layer design for video transmission over
wireless channels,” in Proceedings of the IEEE Vehicular Tech- at HKUST. After she obtained her Ph.D. degree in 2003, she has
nology Conference (VTC ’05), pp. 3019–3023, Stockholm, Swe- joined the Department of Electrical Engineering and Information
den, May 2005. Technology at Munich University of Technology, Germany. Her
[3] M. T. Ivrlač, Wireless MIMO Systems - Models, Performance, current research interests include the areas of smart/MIMO an-
Optimization, Shaker, Aachen, Germany, 2005. tenna systems, multiuser communications, signal processing for
[4] J. Brehmer and W. Utschick, “Modular cross-layer optimiza- wireless communications, multimedia communications, commu-
tion based on layer descriptions,” in Proceedings of the Wire- nication networks, resource allocation, and coding theory.
less Personal Multimedia Communications Symposium (WPMC
’05), Aalborg, Denmark, September 2005. Wolfgang Kellerer is a Senior Manager
[5] M. Van Der Schaar and S. Shankar N, “Cross-layer wire- at NTT DoCoMo’s European Research
less multimedia transmission: challenges, principles, and new Laboratories, Munich, Germany, heading
paradigms,” IEEE Wireless Communications, vol. 12, no. 4, pp. the Ubiquitous Services Platform Research
50–58, 2005. Unit. His current research interests are in
[6] S. Khan, Y. Peng, E. Steinbach, M. Sgroi, and W. Kellerer, “Ap- the area of mobile systems focusing on mo-
plication-driven cross-layer optimization for video stream- bile service platforms, peer-to-peer, sensor
ing over wireless networks,” IEEE Communications Magazine, networks, and cross-layer design. In 2004
vol. 44, no. 1, pp. 122–130, 2006. and 2005, he has served as the elected Vice
[7] M. T. Ivrlač and F. Antreich, “Cross OSI layer optimization - Chairman of the Working Group 2 (Service
an equivalence class approach,” Tech. Rep. TUM-LNS-TR-03- Architecture) of the Wireless World Research Forum (WWRF). He
09, Institute for Circuit Theory and Signal Processing, Munich is a Member of the editorial board of Elsevier’s International Jour-
University of Technology, Munich, Germany, May 2003. nal of Computer and Telecommunications Networking (COM-
[8] M. T. Ivrlač and J. A. Nossek, “Cross layer design - an equiva- NET) and serves as a Guest Editor for the IEEE Communications
lence class approach,” in Proceedings of the International Sym- Magazine in 2006. He has published over 60 papers in respective
posium on Signals, Systems, and Electronics (ISSSE ’04), Linz, journals, conferences, and workshops in the area of service plat-
Austria, August 2004. forms and mobile networking and he filed more than 20 patents.
[9] M. T. Ivrlač, “Parameter selection for the Gilbert-Elliott Before he joined DoCoMo Euro-Labs, he has been a Member of the
model,” Tech. Rep. TUM-LNS-TR-03-05, Institute for Circuit research and teaching staff at the Institute of Communication Net-
Theory and Signal Processing, Munich University of Technol- works at Munich University of Technology. In 2001, he was a Visit-
ogy, Munich, Germany, May 2003. ing Researcher at the Information Systems Laboratory of Stanford
[10] L. U. Choi, M. T. Ivrlač, E. Steinbach, and J. A. Nossek, “Anal- University. He received a Dipl.-Ing. degree (M.S.) and a Dr.-Ing.
ysis of distortion due to packet loss in streaming video trans- (Ph.D.) degree in electrical engineering and information technol-
mission over wireless communication links,” in Proceedings of ogy from Munich University of Technology, Germany, in Decem-
the International Conference on Image Processing (ICIP ’05), ber 1995 and in January 2002, respectively. He is a Member of IEEE
vol. 1, pp. 189–192, Genova, Italy, September 2005. ComSoc and the German VDE/ITG.
[11] Y. Peng, S. Khan, E. Steinbach, M. Sgroi, and W. Kellerer,
“Adaptive resource allocation and frame scheduling for wire-
Eckehard Steinbach studied electrical en-
less multi-user video streaming,” in Proceedings of the Inter-
gineering at the University of Karlsruhe
national Conference on Image Processing (ICIP ’05), vol. 3, pp.
(Germany), the University of Essex (Great
708–711, Genova, Italy, September 2005.
Britain), and Ecole Supérieme d’ Ingénieurs
[12] J. Khun-Jush, G. Malmgren, P. Schramm, and J. Torsner,
en Électronique et Électrotechnique (ES-
“HIPERLAN type 2 for broadband wireless communication,”
IEE) in Paris. From 1994 to 2000, he was a
Ericsson Review, vol. 77, no. 2, pp. 108–119, 2000.
Member of the research staff of the Image
[13] T. Wiegand, G. J. Sullivan, G. Bjontegaard, and A. Luthra,
Communication Group at the University of
“Overview of the H.264/AVC video coding standard,” IEEE
Erlangen-Nuremberg (Germany), where he
Transactions on Circuits and Systems for Video Technology,
received the Engineering Doctorate in 1999.
vol. 13, no. 7, pp. 560–576, 2003.
From February 2000 to December 2001, he was a postdoctoral fel-
[14] L. Choi, W. Kellerer, and E. Steinbach, “Cross layer optimiza-
low with the Information Systems Laboratory of Stanford Univer-
tion for wireless multi-user video streaming,” in Proceedings
sity. In February 2002, he joined the Department of Electrical En-
of the International Conference on Image Processing (ICIP ’04),
gineering and Information Technology of Munich University of
vol. 3, pp. 2047–2050, Singapore, Republic of Singapore, Oc-
Technology (Germany), where he is currently an Associate Profes-
tober 2004.
sor for Media Technology. His current research interests are in the
area of networked and interactive multimedia systems. He served as
Lai-U Choi received the B.Eng. degree from a Conference Cochair of “SPIE Visual Communications and Image
the University of Macau, Macau, in 1998. Processing (VCIP)” in San Jose, Calif, in 2001, and “Vision, Model-
She was educated in the Hong Kong Univer- ing and Visualization 2003 (VMV)” in Munich, in November 2003.
sity of Science and Technology (HKUST), He has been a Guest Editor of the Special Issue on Multimedia over
Hong Kong, for the M.Phil. and the Ph.D. IP and Wireless Networks of the EURASIP Journal on Applied Sig-
study from 1998 to 2003, all in electrical and nal Processing. He currently is a Guest Editor of the EURASIP Jour-
electronic engineering. During this period, nal on Applied Signal Processing, Special Issue on Advanced Video
she has also been a Research Assistant con- Technologies and Applications for H.264/AVC and Beyond. From
ducting research on MIMO signal process- 2006 to 2007, he serves as an Associate Editor for the IEEE Trans-
ing for downlink wireless communications actions on Circuits and Systems for Video Technology (CSVT).
DOI 10.1155/WCN/2006/25861
Multiservice Vertical Handoff Decision Algorithms
Fang Zhu and Janise McNair
Wireless & Mobile Systems Laboratory, Department of Electrical and Computer Engineering, University of Florida,
P.O. Box 116130, Gainesville, FL 32611, USA
Future wireless networks must be able to coordinate services within a diverse-network environment. One of the challenging prob-
lems for coordination is vertical handoff, which is the decision for a mobile node to handoff between different types of networks.
While traditional handoff is based on received signal strength comparisons, vertical handoff must evaluate additional factors, such
as monetary cost, offered services, network conditions, and user preferences. In this paper, several optimizations are proposed for
the execution of vertical handoff decision algorithms, with the goal of maximizing the quality of service experienced by each user.
First, the concept of policy-based handoffs is discussed. Then, a multiservice vertical handoff decision algorithm (MUSE-VDA)
and cost function are introduced to judge target networks based on a variety of user- and network-valued metrics. Finally, a per-
formance analysis demonstrates that significant gains in the ability to satisfy user requests for multiple simultaneous services and
a more efficient use of resources can be achieved from the MUSE-VDA optimizations.
Copyright © 2006 F. Zhu and J. McNair. This is an open access article distributed under the Creative Commons Attribution
cited.
1. INTRODUCTION (3GPP) is currently developing standards for the issue of

when, where, and how to initiate a vertical handoff between
Future wireless networks must be able to coordinate ser- the 3G cellular network and WLAN networks. Future wire-
vices within a diverse network environment. For example, a less integration must include still other relevant factors, such
widely deployed third generation (3G) cellular and data ser- as monetary cost, network conditions, mobile node condi-
vice, such as the general packet radio service (GPRS), may tions, and user preferences, as well as the capabilities of the
be supplemented by the local deployment of high bandwidth various networks in the vicinity of the user. Thus, a complex,
wireless local area networks (WLANs), such as IEEE 802.11 adaptive, and intelligent approach is needed to implement
and the European high performance radio LAN (HiperLAN). vertical handoff protocols to produce a satisfactory result for
Furthermore, as shown in Figure 1, existing networks, such both the user and the network.
as satellite, cellular, and WLAN, will need to integrate with
emerging networks and technologies, such as wireless mesh 1.1. Related work
networks and Wi-Max to allow a user to transparently and
seamlessly roam between systems. Related work on vertical handoff has been presented in re-
Seamless roaming involves handoff, which is the process cent research literature. Several papers have addressed de-
of maintaining a mobile users active connections as it moves signing an architecture for hybrid networks, such as the
within a wireless network [1]. Vertical handoff, or intersys- application-layer session initiation protocol (SIP) [4], the
tem handoff, involves handoff between different types of net- hierarchical mobility management architecture proposed in
works [2, 3]. Traditionally, handoff decisions have been based [5], and the P-handoff protocol [6], which complemented
on an evaluation of the received signal strength (RSS) be- classical vertical handoff by redirecting traffic to the best ad
tween the base station and the mobile node. However, tradi- hoc link, such as Bluetooth and 802.11b, on a peer-by-peer
tional RSS comparisons are not sufficient to make a vertical basis. However, these papers focused on architecture design
handoff decision, as they do not take into account the various and did not address the handoff decision point or the vertical
attachment options for the mobile user. More recently, band- handoff performance issues. Another work considered opti-
width and the type of network have been considered as fac- mizations after the vertical handoff decision has been made,
tors. For example, the third generation partnership project measuring performance with respect to handoff latency [7],
LEO or GEO
satellites
CSS
Global
Suburban
Urban
FES In-building
BS CSS
Picocell
CSS
CSS Microcell
Satellite cell
Macrocell
Figure 1: Diverse third- and fourth-generation (3G and 4G) wireless networks.
TCP timeout and throughput [8, 9], and packet loss [10]. posed along with a cost function that considers several hand-
However, the vertical handoff decision did not consider mul- off metrics. However, multi-service handoff was not fully dis-
tiple networks supporting multiple services for each user. cussed. However, the multiple active services case was not
The related papers that explored vertical handoff decision considered. The work in [22] adaptively adjusted the handoff
mainly focus on traditional issues, such as RSS and data rate. stability period based on a utility function to avoid unneces-
In [11], a fast-Fourier-transform- (FFT-) based signal decay sary handoffs and reduce decision time. Finally, the authors
detection scheme was used to reduce the ping-pong hand- have presented a tutorial on vertical handoffs in [3], and in
off effect, and an adaptive threshold configuration approach [23], introduce a cost function-based vertical handoff deci-
was proposed to prolong the time a user stays in WLAN. In sion algorithm for multiservices handoff. Preliminary results
[12, 13], a vertical handoff algorithm was proposed that took demonstrated significant gain in throughput. This paper ex-
into account RSS, data rate, and packet loss due to handoff tends the work to examine the system performance with re-
delay for a single service per user. A vertical handoff system spect to blocking probability and user satisfactions, that is,
based on computed background noise and signal strength the ability of the network to satisfy all of the users simultane-
was proposed in [14]. In [15], the WISE handoff decision ous requests.
algorithm was proposed to maximize energy-efficiency with- In this paper, several optimizations are proposed to en-
out sacrifice of overall network degradation. In [16], a QoS- hance the handoff decision process and to make the follow-
based handoff method between UMTS and WLAN was pro- ing contributions: (1) the development of a handoff cost
posed, but the definition of QoS was not defined in the pa- function that addresses an environment where users conduct
per. Finally, several papers have focused on mobility level multiple active sessions among a variety of wireless network
and user position in the network. In [17], mobility level was choices, (2) the design of a multiservice vertical handoff deci-
proposed as a proper metric for multi-tier handoffs. In [2], sion algorithm (MUSE-VDA), which incorporates a network
multi-network architectural issues were explored, and an ad- elimination process to potentially reduce delay and process-
vanced neural-network-based vertical handoff algorithm was ing in the handoff calculation, and (3) a constraint opti-
developed to satisfy user bandwidth requirements. In [18], mization analysis for the proposed handoff cost function for
a vertical handoff algorithm based on pattern recognition different types of user services spread among multiple net-
was presented. Although the above-mentioned research ad- works. In Section 2, the policy-based handoff approach is de-
dresses handoff decision, most research address 3G/WLAN scribed. Section 3 introduces the MUSE-VDA cost function
issues, and do not provide a way to incorporate a general, and algorithm to decide target networks based on a variety of
user-defined idea of quality of service, on which to base ver- user- and network-valued metrics. Finally, in Sections 4 and
tical handoff decisions. 5, the performance analysis and numerical results demon-
Several papers have created utility functions to better strate the load-balancing advantages of the proposed tech-
evaluate the choice for vertical handoff. In [19], the verti- nique, as well as the significant gains in satisfied user requests
cal handoff decision function was a measurement of network and a more efficient use of resources. Section 6 concludes the
quality. However, no performance analysis was provided. paper.
In [20], an active application-oriented handoff decision al-
gorithm was proposed for multi-interface mobile terminals 2. POLICY-BASED VERTICAL HANDOFFS
to reduce the power consumption caused by unnecessary
handoffs and other unnecessary interface activation, and in Vertical handoff performed on a policy-based networking
[21], a policy-enabled handoff decision algorithm was pro- architecture requires the coordination of a wide variety of
F. Zhu and J. McNair 3
network devices within a single administrative domain to

implement a set of quality-of-service- (QoS-) based services
[24]. Figure 2 shows two possible conceptual architectures of PDP
policy-based solutions that have been proposed by the IETF.
The two main architectural elements for policy control are Policy DB
the policy enforcement point (PEP) and the policy decision
point (PDP). These two elements may be located in the same
network node (as shown in Figure 2(a)) or in different nodes
PEP
(as shown in Figure 2(b)). The latter is especially convenient
to apply local policies.
PEP is a component that runs on a policy-aware node, Network node
such as an access point, and is the point at which the poli- (a) PEP and PDP located in the
cies are enforced. Policy decisions are made primarily at the same network node
PDP, based on the policies extracted from a network policy
database. The PDP as specified by the IETF may make use of
additional mechanisms and protocols to achieve additional Policy server
functionality such as user authentication, accounting, and
policy information storage.
PDP
In the case of vertical handoff, the policy database holds
information regarding the metrics to be considered for a Policy DB
vertical handoff, where handoff metrics are the measured
qualities that give an indication of whether or not a hand-
off is needed. As stated previously, in traditional handoffs,
only RSS and channel availability are considered. In the envi- PEP
sioned integrated wireless system, the following new metrics
are suggested [3].
Network node
(i) Service type. Different types of services require various (b) PEP and PDP located in different
combinations of reliability, latency, and data rate. network nodes
(ii) Monetary cost. A major consideration to users, as dif-
ferent networks may employ different billing strategies Figure 2: Two possible policy-based network architectures.
that may affect the user’s choice to handoff.
(iii) Network conditions. Network-related parameters such
as traffic, available bandwidth, network latency, and
congestion (packet loss) may need to be considered and more ambiguous. However, the use of an optimized cost
for effective network usage. Use of network informa- function can simplify the handoff process and speed up the
tion in the choice to handoff can also be useful for load handoff decision. Then, intelligent techniques can be devel-
balancing across different networks, possibly relieving oped to evaluate the effectiveness of new decision algorithms,
congestion in certain systems. balanced against user satisfaction and network efficiency.
(iv) System performance. To guarantee the system perfor-
mance, a variety of parameters can be employed in 2.1. Proposed vertical handoff interworking scenarios
the handoff decision, such as the channel propaga-
To demonstrate the operation of the policy-based architec-
tion characteristics, path loss, interchannel interfer-
tures, the following two scenarios are described: (1) net-
ence, signal-to-noise ratio (SNR), and the bit error rate
work-controlled handoff (NCHO)/mobile-assisted handoff
(BER). In addition, battery power may be another cru-
(MAHO), where the network generates a new connection
cial factor for certain users. For example, when the bat-
and finds new resources for the handoff, performing any ad-
tery level is low, the user may choose to switch to a
ditional routing operations, and (2) mobile-controlled hand-
network with lower power requirements, such as an ad
off (MCHO), where the mobile terminal must take its own
hoc Bluetooth network.
measurements and make the evaluations for the handoff de-
(v) Mobile terminal conditions. MT condition includes dy-
cision.
namic factors such as velocity, moving pattern, moving
NCHO/MAHO is shown in Figure 3(a). The handoff
histories, and location information.
decision procedure begins with the PEP. Upon receiving a
(vi) User preferences. User preference can be added to cater
handoff trigger, the PEP formulates a request for a policy de-
to special requests for users that favor one type of sys-
cision and sends it to the PDP. The request for policy control
tem over another.
from the PEP to the PDP may contain one or more policy
The use of new vertical handoff metrics and the policy- elements extracted from the mobile terminals that are neces-
based networking architecture increases the complexity of sary for handoff decision. The PDP then extracts other nec-
the handoff process and makes the handoff decision more essary information, for example, the users subscriber profile
ing for a vertical handoff controlled by the mobile terminal.

PDP However, if simple metrics are set, a combination of the two
techniques, that is, mobile-assisted handoff (MAHO), may
BS/AP REQ DEC be a viable option.
PEP Handoff trigger 3. MULTISERVICE VERTICAL HANDOFF DECISION

ALGORITHM COST FUNCTION
DEC
MT The MUSE-VDA vertical handoff cost function measures the
benefit obtained by handing off to a particular network. It is
(a) NCHO or MAHO handoff decision evaluated for each network n that covers the service area of
procedure a user. The network choice that results in the lowest calcu-
lated value of the cost function is the network that provides
Home agent
the most benefit, where the benefit is defined by the given
handoff policy.
The cost function evaluated for network n includes the
DEC cost of receiving each of the user’s requested services from
BS/AP
network n and is calculated:

Cn = Csn , (1)
PDP s
MT REQ DEC
Policy DB where s is the index representing the user-requested services,
and Csn is the per-service cost function for network n. Csn rep-
PEP Handoff trigger resents the QoS experienced by choosing to receive service s
from network n and is calculated as
(b) MCHO handoff decision proce-
dure Csn = Ws,nj Qs,n j , (2)
j
Figure 3: Two scenarios for policy-based architectures.

where Qs,n j is the normalized QoS provided by network n
for parameter j and service s. Ws,nj is the weight which in-
dicates the impact of the QoS parameter on the user or the
network. Csn includes both a normalized value for the QoS
and network conditions, from the database located in local or parameter and a weight for the impact of the parameter on
home network, makes the handoff decision, and returns the either the user or the network. For an example from the users
decision message to the PEP. The handoff decision is made perspective, suppose that a mobile terminal requests a ser-
using utility-function-based algorithms as proposed in [23]. vice with a specified minimum delay and minimum power
The PEP then informs the mobile terminal about the handoff consumption requirement. If the mobile terminal has a low
decision and enforces the policy decision by handing off to battery life, the power consumption takes on greater impor-
the target network. In NCHO/MAHO, we propose that the tance than meeting the delay constraints. For an example of
PDP point is represented by the base station (BS) or access a network-based QoS request and the corresponding impact,
point (AP). the availability of the services requested by the user in the
In MCHO, the mobile terminal finds new resources and target network impacts the network congestion in the tar-
the network approves the handoff decision. Thus, we propose get network. Using the impact factor, the network may direct
that the PDP is located at the mobile terminal. As shown in users toward a less desirable, but less congested network.
Figure 3(b), when the mobile terminal detects a severe QoS The handoff decision problem thus equals the following
degradation, its PEP module triggers the handoff decision constraint optimization problem:
process by sending a handoff decision request message to

the PDP. While some information is already available at local min C n = Csn Ws,nj Qs,n j s.t. Es,n j
= 0, ∀s, i, (3)
database, the PDP may also need other necessary informa- s j
tion, such as network conditions, from the network devices.
Other information may not be immediately available at the where Es,n j is the network elimination factor, indicating
BS or AP, and may need to be extracted from the network. whether the constraint i for service s can be met by network
Upon receiving all handoff metrics, the PDP makes the hand- n. It is equal to one if constraint i can be satisfied, and is
off decision and returns the decision to the PEP. The PEP equal to zero if constraint i cannot be satisfied. It is intro-
then informs the network the handoff decision by forwarding duced to reflect the inability of a network to guarantee the
the DEC message, along with enforced authentication infor- requested QoS constraints for a particular service s, and can
mation. A handoff will take place once the network approves. be implemented as a checklist at PDP. For example, an avail-
It may be a limiting factor to achieve the necessary process- able network may not be able to guarantee the minimum
Begin with a list of active services 3.3. Cost function example
As an example, consider a reporter in the field using wire-

Select the service with highest priority
less networks to send audio, video reports, and photographic
images to a home base, but whose equipment is running low
Evaluate (4) for each possible target network on battery power. There are three available networks, UMTS,
WLAN, and satellite. The cost function calculation from (3)
Handoff to network n based on the optimal result of (4) is formed as follows:
(i) n represents the three network choices, UMTS,
Yes
Any unassigned services left? WLAN, or a satellite network.
(ii) s represents the services needed, in this case, audio,
No
Update resource database video, and images.
(iii) j represents the constraint parameters: bandwidth,
End battery power consumption, and delay.
(iv) For collective handoff, a calculation of (3) is made for
Figure 4: Scenario 2: prioritized session handoff. each network.
(1) For example, for the UMTS network,
requested delay for a real-time service, and should be imme-

C UMTS = Wvideo,
UMTS
bandwidth Qvideo, bandwidth
UMTS
diately removed from consideration as a handoff target for
the requested service. + Wvideo, battery power Qvideo, battery power
UMTS UMTS
The application of the vertical handoff cost function is
flexible to allow for different vertical handoff policies. To + Wvideo,
UMTS
delay Qvideo, delay
UMTS
demonstrate the performance of the new cost function, two
different policy scenarios are explored. + Waudio,
UMTS
bandwidth Qaudio, bandwidth
UMTS
3.1. Collective session handoff + Waudio,

UMTS
battery power Qaudio, battery power
UMTS
(5)

It is assumed that a single user may conduct multiple com- + Waudio,
UMTS
delay Qaudio, delay
UMTS
munication sessions. In the first vertical handoff policy, the

vertical handoff decision is optimized for all sessions collec- + Wimage,
UMTS
bandwidth Qimage, bandwidth
UMTS
tively, that is, all of the users active sessions are handed off
to the same target network at the same time. The cost func- + Wimage,
UMTS
battery power Qimage, battery power
UMTS
tion, C n , is determined for all sessions going to a single net-

work. The optimal target network for handoff is determined + Wimage,
UMTS
delay Qimage, delay .
UMTS
by solving (3).
(2) Then, C WLAN and C Satellite are calculated similarly.
3.2. Prioritized session handoff (3) The lowest of the three costs C UMTS , C WLAN , and
C Satellite reveals the target network. If satellite cost
The second vertical handoff policy prioritizes each service is the lowest, then all sessions, video, audio, and
and then optimizes the vertical handoff decision individu- images, are sent via the satellite network.
ally for each session, that is, each of the users active sessions
may be independently handed off to a different target net- (v) For prioritized session handoff, a calculation of (4) is
work. In this scenario, the mobile terminal maintains a list of made for the highest priority session.
its current active sessions, arranged in priority order. Then, (1) For example, if the video feed has the highest pri-
the cost function Csn is evaluated for the highest priority ser- ority, then Cvideo
UMTS
is calculated first:
vice. The optimal target network is chosen by minimizing the
per-service cost:
Cvideo
UMTS
= Wvideo,
UMTS
bandwidth Qvideo, bandwidth
UMTS

min Csn = Ws,nj Qs,n j s.t. Es,n j
= 0, ∀i. (4) + Wvideo,
UMTS
battery power Qvideo, battery power
UMTS
(6)
s

+ Wvideo,
UMTS
delay Qvideo, delay .
UMTS
Then, the next highest priority service is selected, the cor-
responding cost function is evaluated, and the target network
is determined. The process continues to the last active ses- (2) Then CvideoWLAN
and Cvideo
Satellite
are calculated similarly.
sion. If the constraints for one session cannot be met, then (3) The lowest of the three costs Cvideo UMTS
, Cvideo
WLAN
, and
the user loses the individual session only. The process for the Cvideo reveals the target network for video service
Satellite
second scenario is outlined in Figure 4. only.

Network 3
D L E
SN3
SN3 N1 SN3 N2
H
O R
SN1 N3 SN2 N3
K C I S
A P B
Q
T
SN3 N1N2
Network 2
Network 1
SN3
G M F
SBOUND
Figure 5: 3G/WLAN overlay network scenario.
(4) The calculation is repeated for the next highest 4.1. Mobility model
priority service, say the audio feed. Thus, in the
prioritized session handoff it may be the case that User mobility trajectories are characterized by the widely
the video is sent via satellite for the bandwidth, used random waypoint (RWP) model [25]. Adjustments
but the audio is sent via UMTS. have been included to account for the shortcomings of the
waypoint model described in [12]. Each user chooses uni-
In the next section, the performance of the proposed formly at random a destination point (or waypoint) in the
MUSE-VDA algorithm and cost function is analyzed. First, dashed rectangle in Figure 5. A user moves to this destina-
a sample overlay network scenario is provided, along with tion with a velocity v, which is chosen uniformly in the inter-
a description of the mobility model, followed by calculations val (v min, v max 0). (The v min and v max are chosen to be
of the blocking probability and the average percentage of user 0.3 m/s and 12.5 m/s, resp.) When the user reaches the way-
requests that are satisfied by the network. point, it remains static for a predefined pause time, and then
moves again according to the same rule. Note that user tra-
4. MUSE-VDA PERFORMANCE ANALYSIS jectories characterized by the improved RWP model can be
assumed to be uniformly distributed at any given time.
For effective comparison with other techniques, the per- A user with active sessions that enters the overlay of all
formance analysis considers the case of 3G/WLAN hand- three networks must decide when and where to execute a ver-
off scenario, where received signal strength (RSS), chan- tical handoff request. If the request is accepted, the appropri-
nel availability, and bandwidth are the specified constraints. ate amount of bandwidth is assigned by the serving network.
However, note that any other network combination or any If the request is denied at one network, the request can be
other combination of the vertical handoff metrics listed in reassigned to another network, if resources are available at
Section 2 can just as easily be substituted in the evaluation. the second network. If the second (or third) network is not
The top view of a typical 3G network overlay environ- available, the request is blocked from the system. Next, we
ment is shown in Figure 5, where three networks of differ- formulate the calculation of the blocking probabilities.
ent maximum data rates coexist in the same wireless service
area. Network 1 (centered at A) and Network 2 (centered at 4.2. Blocking probability
B) each represent a WLAN, while Network 3 (centered at C)
represents a GPRS network. The shaded circles on the left Each of the three networks in Figure 5 is modeled as an
and right represent the area where RSS from Network 1 or M/M/1/Nn queue system [26], where Nn is the number of
Network 2 is stronger than that from Network 3. To high- available channels in Network n. Nn is calculated:
light the effects of the vertical handoff procedure among the
three networks, only the users within the overlapping areas Bn
Nn = , (7)
are considered, represented by the dashed square in Figure 5. D
where Bn is the total bandwidth of Network n, and D is the 4.3. Average percentage of satisfied user requests
average data rate of each user. The traffic load within the
overlay cells is ρ = λ/μ, where λ is the arrival rate of service Each user comes to the network overlay area with a certain
requests, μ is the departure rate, and arrivals and departures set of requests, including various services and data rates. As
are modeled as Poisson distributions. Handoff calls are given mentioned previously, the ability of the network to satisfy
a higher priority than new calls, and for simplicity, a buffer- user requests depends on whether the sessions are treated as
less handoff algorithm is used. a collective or as prioritized, individual sessions. In the col-
For the blocking probability of Network n, Pbn , we use lective MUSE-VDA and the RSS technique, all requests from
the blocking probability of an M/M/1/Nn queue when there one user are considered collectively. Thus, if a target network
are Nn users in system [26]: cannot satisfy all of the requests as a collective, then the user
is blocked from the system. In the prioritized MUSE-VDA

ρnNn 1 − ρn technique, each session is treated individually, and thus one
Pbn = , (8) user may have a subset of their requests satisfied, while other
1 − ρnNn +1
portions are blocked. The APUSR tracks the percentage of
where ρn is the effective load experienced by Network n: incoming requests that actually receive service at one of the
available networks.
ρn = rn ρ (9) The APUSR is calculated for the overlay network as fol-
lows:
and rn is the percentage of total requests that will go to Net-
work n, based on the vertical handoff decision metrics. To E AR = ARi P Ri , (12)
determine rn , both original handoff requests and the hand- i
off requests that arrive are included, to account for the times
where ARi is the APUSR for Region i, and where the regions
that the user has been rejected by another network. Since it is
are described in Table 1. ARi is calculated:
assumed that the users are uniformly distributed, the service

request load can be calculated according to the proportion of ARi = ti j P Ni j , (13)
the coverage area within the boundary region. The coverage j
areas are labeled in Figure 5, and the corresponding coverage,
the execution of the RSS and MUSE-VDA algorithms are de- where ti j is the maximum APUSR that can be received from
scribed in Table 1. Network Ni j in Region i, and P(Ni j ) is the probability that
For the RSS-based handoff algorithm, the values of rn for Network Ni j is available and chosen by a user. Finally, P(Ri )
n = 1, 2, 3 are calculated as follows: is the probability that a user is located in Region i:
SBOUND − SN1−N3 − SN2−N3 Si
r3 = , P Ri = . (14)
SBOUND SBOUND
SN1−N3 + SN3−N1 Pb3 In the next section, we implement the performance anal-
r1 = , (10)
SBOUND ysis and obtain results for several service request scenarios.
SN2−N3 + SN3−N2 Pb3
r2 = , 5. NUMERICAL RESULTS
SBOUND
where Pb3 is defined in (8), Si is the geometric area of region i The user mobility, user requests, network acceptances and
described in Table 1, and SBOUND is the geometric area of the denials for the 3G/WLAN overlay system in Figure 5 were
boundary region. modeled and simulated using MATLAB, based on the sys-
For the MUSE-VDA handoff algorithm, the values of rn tem parameters shown in Table 2. Each user can request a
for n = 1, 2, 3 are calculated: data rate up to a maximum of 500 kbps. To gauge the re-
sponse of the protocol to different traffic types, this data rate
SN1−N3 + SN3−N1 + SN3−N1N2 includes a combination of constant bit rate (CBR) services
r1 = ,
SBOUND and available bit rate (ABR) services, where the CBR request
SN2−N3 + SN3−N2 + SN3−N1N2 Pb1 per user is limited to a maximum of 50 kbps and the ABR
r1 = , request per user is limited to a maximum of 450 kbps. Note
SBOUND
that Network 1 or Network 2 can fully satisfy the maximum
r2 possible data rate request of 500 kbps. However, Network 3

SN3 + SN1−N3 +SN3−N1 Pb1+ SN2−N3+SN3−N2 +SN3−N1N2 Pb2 can only satisfy 30% of the maximum possible 500 kbps re-
= . quest. We note that the data rates for the networks listed in
SBOUND
(11) Table 2 can be considered as low estimates. However, the ob-
jective is to gauge the ability of a combination of networks
Finally, we develop a calculation for a measure of the ser- to satisfy as many user requests as possible. Thus, as data
vice obtained by each user, as compared to the services re- rates per network increase, the size of the data rate request
quested by each user. This is defined here as average percent- may also increase, but the resulting trends for the given algo-
age of users’ satisfied requests (APUSR). rithms would remain the same.
Table 1: RSS and MUSE-VDA algorithm description for 3G WLAN overlay network in Figure 5.
Region number Region area label Algorithm descriptions

1 SN3 (DEH,JFG) Network 3 provides the only coverage
(HIJK) Network 3 has the strongest RSS.
RSS Algorithm: if the request is denied by Network 3, the user
can try either Network 1 or Network 2 with equal probability.
2 SN3−N1N2
MUSE-VDA: the network order with respect to decreasing data rate is as follows:
Network 1 > Network 2 > Network 3.
The outcome of the cost function will be to choose Network 1, then Network 2
if Network 1 is denied, then Network 3, if Network 2 is denied.
(DHKJGP) Network 3 has the strongest RSS.
RSS Algorithm: Network 3 is chosen first. If the request is denied by
3 SN3−N1 Network 3, the user tries Network 1.
MUSE-VDA: according to the decreasing data rates, the selection made by
the cost function is first Network 1, then Network 3 if Network 1 is denied.
(EHIJFS) Network 3 has the strongest RSS.
(OPQA) Network 1 has the strongest RSS.
(RSTB) Network 2 has the strongest RSS.
Sbound Boundary region
Table 2: System parameters. 5.1. RSS-based algorithm results

Network (n) Network type Resource First, the RSS performance is examined to provide a base-
1 WLAN 2 Mbps [27] line for comparison with the MUSE-VDA results. Figure 6(a)
shows the APUSR with the increasing network load for
2 WLAN 1 Mbps [27] an RSS-based handoff algorithm. Since Network 3 has the
3 GPRS
Up to 8 slots, strongest transmit power, it is the preferred service provider.
21.4 Kbps per slot [27] Thus, at the low-load range, Network 3 must satisfy a large
portion of the total requests. With increasing network load,
the resources of Network 3 are used up earlier than the re-
sources of the other two networks. The affect is to separate
the APUSR into three regions.
As mentioned previously, the random waypoint model
is used to simulate user mobility, with the following param- (1) In the first region, 0.1 < ρ < 1, most of the requests
eters: vmin = 0.3 m/s (1 km/h), vmax = 12.5 m/s (45 km/h), go to GPRS (Network 3), while the WLANs are under-
and vthreshold = 5.5 m/s (20 km/h). used.
1 0.9
0.9 0.8
0.8 0.7
Blocking probabilities
0.7 0.6
0.6
0.5
APUSR
0.5
0.4
0.4
0.3
0.3
0.2
0.2
0.1 0.1
0 0
1 10 1 100 101
10 100 101
ρ ρ
Network 1 Network 3 Network 1

Network 2 Total Network 2
Network 3
(a) Average percentage of user satisfied requests (APUSR) (b) Blocking probability
Figure 6: Performance of the RSS-based algorithm.
(2) In the second region, 1 < ρ < 2, GPRS begins to deny performance. However, its achievable APUSR is lower than
users, and the WLANs begin to receive more requests. that of MUSE-VDA (which will be discussed in more detail
(3) In the third region, 2 < ρ, all three networks are satu- later in this section), that is, there remains a load-balancing
rated and the QoS degrades for all networks. issue for increasing requests.
We now examine the MUSE-VDA performance by con-
Thus, the problem with the RSS approach is that there is no
sidering two handoff scenarios: (1) collective handoff, where
load balancing according to the service requests of the users
all of the user’s active sessions are handed off to the same tar-
and the available networks.
get network at the same time, and (2) prioritized multinet-
Figure 6(b) demonstrates the corresponding blocking
work handoff, where each service is prioritized and optimal
probability of each network for the traditional RSS algo-
decision is made individually for each session.
rithm. An increase in blocking probability of Network 3 ear-
lier than Networks 1 and 2 can be observed. Mobile users
thus have a greater chance to select Network 1 and Network 5.3. MUSE-VDA
2 as service provider. Since they have a total APUSR that is
higher than Network 3 by itself, a “hump” can be observed. The MUSE-VDA cost functions, (3) and (4), are evaluated
The result that Network 3 is chosen more often as the tar- for each network based on the following parameters:
get handoff cell leads to two unsatisfactory effects: (1) unbal- (i) Network index n represents the two WLANs and one
anced load assignment and (2) low overall achievable data GPRS network, as shown in Table 2.
rate. Only when the resource in Network 3 is highly con- (ii) Two constraints are considered: available bandwidth
sumed, Networks 1 and 2 will have a greater chance to be the and RSS (R), where the limiting constraint for bandwidth is
service provider. Thus a more intelligent handoff algorithm Bsn − Breq ≥ 0 for some network n and service s, and the
that can balance the usage of overlay networks is needed, and limiting RSS contraint is Rn − Rth ≥ 0.
a higher overall APUSR is expected. (iii) The weights in the cost functions are normalized to
1, meaning that each service contraint is treated with equal
5.2. RSS with mobility metric weight.
(iv) The QoS factor is a normalized bandwidth calcula-
n n n
Next, we compare the RSS-only technique versus a mobility- tion, where QCBR, bandwidth = ln |1/BCBR |, and QABR, bandwidth =
n
level technique. Mobility level is a metric that can be com- ln |1/BABR |.
bined with RSS based to improve system performance. For (v) The target network is chosen according to the proce-
example, fast moving users (v > vthreshold ) are selected to dure described in Table 1.
receive service from the largest cell, while medium-to-slow Figure 8(a) shows MUSE-VDA results for the APUSR
users (v < vthreshold ) receive service from the small cells. provided by each of the three networks and overall achiev-
Figure 7 shows the APUSR and blocking probability com- able APUSR implementing the collective handoff algorithm,
parison of the pure RSS based algorithm and the RSS-based for comparison with Figure 6, the RSS-only case. Since either
algorithm combined with mobility level consideration. The Network 1 or Network 2 provides relatively larger data rate
mobility level algorithm demonstrates an improved APUSR than Network 3, they are the default service provider for the
1 1
0.9 0.9
0.8 0.8
0.7 0.7
0.6 0.6
APUSR
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
10 1 100 101 10 1 100 101
ρ ρ
RSS RSS
RSS + mobility RSS + mobility
MUSE-VDA
Figure 7: Performance of the RSS-based algorithm with added mobility considerations.
1 0.9
0.9 0.8
0.8 0.7
0.7 0.6
0.6
0.5
APUSR
0.5
0.4
0.4
0.3
0.3
0.2
0.2
0.1 0.1
0 0
1 10 1 100 101
10 100 101
ρ ρ
Network 1 Network 3 Network 1

Network 2 Total Network 2
Network 3
Figure 8: Performance of the MUSE-VDA algorithm.
mobile users, depending on their location. Thus, at the low- (1) In the first region, 0.1 < ρ < 1, most of the requests
load range, Network 1 and Network 2 satisfy the most por- go to the WLANs, which are able to handle the higher
tion of the total request. With the increasing network load, data rate requests.
the resource of Network 1 and Network 2 is consumed ear- (2) In the second region, 1 < ρ, WLANs begin to deny
lier than the resources of Network 3. Then mobile users start users, and the GPRS provides a useful alternative. All
to select Network 3 more frequently than in low-load range. three networks are being utilized and the performance
The portion of requests satisfied by Network 3 thus starts to degrades gradually.
increase when the portion satisfied by Network 1 and Net-
work 2 decreases. In this case, there are only two regions rep- Thus, in the MUSE-VDA case, the load balancing is im-
resented in the figure. proved for all networks.
1 1
0.9 0.9
0.8
0.8
0.7
0.7
0.6
APUSR
APUSR
0.6 0.5
0.5 0.4
0.3
0.4
0.2
0.3
0.1
0.2 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 10 1 100 101
CBR (kbps) ρ
RSS RSS
Collective MUSE-VDA Collective MUSE-VDA
Prioritized MUSE-VDA Prioritized MUSE-VDA
(a) Average percentage of user satisfied requests (APUSR)
Figure 9: APUSR with varying CBR data rate requests.
1
Figure 8(b) demonstrates the corresponding blocking 0.9
probability of each network. An increase in blocking prob- 0.8
ability of Networks 1 and 2 earlier than Network 3 can be
observed, which indicates that WLANs are favorite networks 0.7

due to their relative larger available bandwidth to each user. 0.6
0.5
5.4. MUSE-VDA results for more demanding
0.4
CBR services
0.3
In future wireless networks, users may request much higher
0.2
CBR service rates, as video and audio conferencing and other
real-time services become prevalent. Thus, the next set of 0.1
simulations study the impact of increasing the request of 0
1 100 101
CBR services. Results are now presented for APUSR and 10
blocking probability for three cases: the traditional handoff ρ
protocol based on the strongest RSS, the cost function with RSS
collective handoff, and the cost function with the prioritized Collective MUSE-VDA
handoff. Figure 9 shows APUSR versus CBR (per user) with Prioritized MUSE-VDA
zero blocking rates for all three algorithms. Here, each user (b) Blocking probability
may request variable CBR and up to 1 Mbps ABR services.
The figure demonstrates that APUSR decreases with an in- Figure 10: MUSE-VDA performance for more demanding CBR
crease in CBR request for all three algorithms. However, as data rate requests.
each user’s request increases, the ability of the large net-
work to support the variable data rate decreases more dra-
matically. By optimally spreading user’s services over several no longer join with Network 3. However, the APUSR in col-
networks using the prioritized MUSE-VDA technique, more lective MUSE-VDA decreases, since Networks 1 and 2 cannot
bandwidth can be assigned to ABR services, which results in cover the whole area. On the other hand, all 3 networks can
a higher overall APUSR per user. be used in prioritized MUSE-VDA, where the user’s two ses-
Figure 10 shows APUSR and blocking probability versions can be spread into multiple networks. If the bandwidth
sus user requests for the three handoff algorithms for the for one session cannot be satisfied, only one session will be
more demanding CBR requests. In this case, Network 3 is blocked. This results in a higher APUSR and a lower blocking
eliminated in RSS-based and collective MUSE-VDA handoff probability than the other two handoff schemes. Moreover,
algorithms, due to its limited data rate per user (less than since users moving out of the limited coverage of Networks 1
170 kbps). Thus, users will only be able to choose between and 2 cannot be served in RSS-based handoff algorithm and
Networks 1 and 2. This increases the APUSR in RSS-based collective MUSE-VDA, a nonzero blocking probability can
algorithm in light traffic (ρ < 1 in Figure 10(b)), since users be observed all the time.
6. CONCLUSION [10] M. Salamah, F. Tansu, and N. Khalil, “Buffering requirements

for lossless vertical handoffs in wireless overlay networks,”
Expanding services through the use and coordination of di- in Proceedings of 57th IEEE Semiannual Vehicular Technology
verse networks creates the challenge of developing a more Conference (VTC ’03), vol. 3, pp. 1984–1987, Jeju, South Ko-
complex, adaptive, and intelligent vertical handoff protocol. rea, April 2003.
In this paper, MUSE-VDA has been developed to maximize [11] Q. Zhang, C. Guo, Z. Guo, and W. Zhu, “Efficient mobil-
the benefit of the handoff for both the user and the network. ity management for vertical handoff between WWAN and
The optimizations incorporate a network elimination feature WLAN,” IEEE Communications Magazine, vol. 41, no. 11, pp.
to reduce the delay and processing required in the evaluation 102–108, 2003.
[12] M. Ylianttila, M. Pande, J. Makela, and P. Mahonen, “Opti-
of the cost function, and a multinetwork optimization is in-
mization scheme for mobile users performing vertical hand-
troduced to improve APUSR for mobile terminals with mul- offs between IEEE 802.11 and GPRS/EDGE networks,” in
tiple active sessions. A performance analysis demonstrated Proceedings of IEEE Global Telecommunications Conference
significant gains in the ability to satisfy user’s requests for (GLOBECOM ’01), vol. 6, pp. 3439–3443, San Antonio, Tex,
multiple simultaneous services and a more efficient use of USA, November 2001.
resources from the proposed optimizations. [13] M. Ylianttila, J. Makela, and P. Mahonen, “Supporting re-
In this treatment, a “proof-of-concept” has been pro- source allocation with vertical handoffs in multiple radio net-
vided based on the ability of a user or network to choose work environment,” in Proceedings of the 13th IEEE Interna-
among different network types, based on different service re- tional Symposium on Personal, Indoor, and Mobile Radio Com-
quirements. Future work is ongoing to look at a more communications (PIMRC ’02), vol. 1, pp. 64–68, Lisbon, Portugal,
plex treatment, including the study of optimal and reason- September 2002.
able weight selection, QoS factor normalizations, and policy [14] S. Sharma, I. Baek, Y. Dodia, and T. Chiueh, “OmniCon: a
mobile IP-based vertical handoff system for wireless LAN and
decision point architecture implementations.
GPRS links,” in Proceedings of IEEE International Conference on
Parallel Processing (ICPP ’04), pp. 330–337, Montreal, Quebec,
REFERENCES Canada, August 2004.
[15] M. Nam, N. Choi, Y. Seok, and Y. Choi, “Wise: energy-efficient
[1] I. F. Akyildiz, J. McNair, J. S. Ho, H. Uzunalioğlu, and W. interface selection on vertical handoff between 3G networks
Wang, “Mobility management in next-generation wireless sys- and WLANs,” in Proceedings of the 15th IEEE International
tems,” Proceedings of the IEEE, vol. 87, no. 8, pp. 1347–1384, Symposium on Personal, Indoor and Mobile Radio Communi-
1999. cations (PIMRC ’04), vol. 1, pp. 692–698, Barcelona, Spain,
[2] K. Pahlavan, P. Krishnamurthy, A. Hatami, et al., “Handoff in September 2004.
hybrid mobile data networks,” IEEE Personal Communications, [16] S. Jung, D.-H. Cho, and O. Song, “QoS based vertical hand-
vol. 7, no. 2, pp. 34–47, 2000. off method between UMTS systems and wireless LAN net-
[3] J. McNair and F. Zhu, “Vertical handoffs in fourth-generation works,” in Proceedings of 60th IEEE Semiannual Vehicular Tech-
multinetwork environments,” IEEE Wireless Communications, nology Conference (VTC ’04), vol. 6, pp. 4451–4455, Los Ange-
vol. 11, no. 3, pp. 8–15, 2004. les, Calif, USA, September 2004.
[4] W. Wu, N. Banerjee, K. Basu, and S. K. Das, “SIP-based vertical
[17] J. M. Holtzman and A. Sampath, “Adaptive averaging method-
handoff between WWANs and WLANs,” IEEE Wireless Com-
ology for handoffs in cellular systems,” IEEE Transactions on
munications, vol. 12, no. 3, pp. 66–72, 2005.
Vehicular Technology, vol. 44, no. 1, pp. 59–66, 1995.
[5] H. Badis and K. Al-Agha, “Fast and efficient vertical handoffs
[18] A. Mehbodniya and J. Chitizadeh, “An intelligent vertical
in wireless overlay networks,” in Proceedings of IEEE Interna-
handoff algorithm for next generation wireless networks,” in
tional Symposium on Personal, Indoor and Mobile Radio Com-
Proceedings of the 2nd IEEE/IFIP International Conference on
munications (PIMRC ’04), vol. 3, pp. 1968–1972, Barcelona,
Wireless and Optical Communications Networks (WOCN ’05),
Spain, September 2004.
pp. 244–249, Dubai, UAE, March 2005.
[6] J. Tourrilhes and C. Carter, “P-Handoff: a protocol for fine-
grained peer-to-peer vertical handoff,” in Proceedings of the [19] A. Hasswa, N. Nasser, and H. Hassanein, “Generic vertical
13th IEEE International Symposium on Personal, Indoor, and handoff decision function for heterogeneous wireless net-
Mobile Radio Communications (PIMRC ’02), vol. 2, pp. 966– works,” in Proceedings of the 2nd IEEE/IFIP International Con-
971, Lisbon, Portugal, September 2002. ference on Wireless and Optical Communications Networks
[7] M. Bernaschi, F. Cacace, and G. Iannello, “Vertical hand- (WOCN ’05), pp. 239–243, Dubai, UAE, March 2005.
off performance in heterogeneous networks,” in Proceedings [20] W.-T. Chen and Y.-Y. Shu, “Active application oriented vertical
of IEEE International Conference on Parallel Processing (ICPP handoff in next-generation wireless networks,” in Proceedings
’04), pp. 100–107, Montreal, Quebec, Canada, August 2004. of IEEE Wireless Communications and Networking Conference
[8] H. Huang and J. Cai, “Improving TCP performance during (WCNC ’05), vol. 3, pp. 1383–1388, New Orleans, La, USA,
soft vertical handoff,” in Proceedings of the 19th IEEE Inter- March 2005.
national Conference on Advanced Information Networking and [21] H. J. Wang, R. H. Katz, and J. Giese, “Policy-enabled hand-
Applications (AINA ’05), vol. 2, pp. 329–332, Taipei, Taiwan, offs across heterogeneous wireless networks,” in Proceedings of
March 2005. the 2nd IEEE Workshop on Mobile Computing Systems and Ap-
[9] S.-E. Kim and J. A. Copeland, “TCP for seamless verti- plications (WMCSA ’99), pp. 51–60, New Orleans, La, USA,
cal handoff in hybrid mobile data networks,” in Proceed- February 1999.
ings of IEEE Global Telecommunications Conference (GLOBE- [22] W.-T. Chen, J.-C. Liu, and H.-K. Huang, “An adaptive scheme
COM ’03), vol. 2, pp. 661–665, San Francisco, Calif, USA, De- for vertical handoff in wireless overlay networks,” in Pro-
cember 2003. ceedings of the 10th International Conference on Parallel and
Distributed Systems (ICPADS ’04), vol. 10, pp. 541–548, New-

port Beach, Calif, USA, July 2004.
[23] F. Zhu and J. McNair, “Optimizations for vertical handoff deci-
sion algorithms,” in Proceedings of IEEE Wireless Communica-
tions and Networking Conference (WCNC ’04), vol. 2, pp. 867–
872, Atlanta, Ga, USA, March 2004.
[24] R. Yavatakar, D. Pendarakis, and R. Guerin, A Framework
for Policy-Based Admission Control, Internet Engineering Task
Force, February 2006, http://www.ietf.org.
[25] D. B. Johnson and D. A. Maltz, Dynamic Source Routing in Ad
Hoc Wireless Networks, Kluwer Academic, Norwell, Mass, USA,
1996.
[26] D. Bertsekas and R. Gallager, Data Networks, Prentice Hall, En-
glewood Cliffs, NJ, USA, 1991.
[27] K. Pahlavan and P. Krishnamurthy, Principles of Wireless Net-
works, Prentice Hall, Upper Saddle River, NJ, USA, 2002.
Fang Zhu received the B.S. degree from

Beijing University of Posts and Telecom-
munications, Beijing, China, in 1999, the
M.S. degree from Colorado State Univer-
sity, Fort Collins, Colorado, in 2002, and
the Ph.D. degree from University of Florida,
Gainesville, Fla, in 2005. She is currently
with Verizon as a Software Systems Archi-
tect. Her research interests include mobil-
ity management, quality of service, and re-
source management for the next generation wireless systems.
Janise McNair holds the B.S. degree and the

M.S. degree in electrical engineering from
the University of Texas at Austin (1991 and
1993, resp.), and the Ph.D. degree in elec-
Georgia Institute of Technology, Atlanta,
Georgia, USA (2000). She is currently an
Assistant Professor in the Department of
Electrical and Computer Engineering at the
University of Florida. She is a Member of
the IEEE and the ACM, and she serves on the Editorial Board of
the Ad Hoc Networks Journal and the IEEE Transactions on Mo-
bile Computing. Her current research interests are vertical hand-
off management, mobile user security and authentication, and
medium access techniques for wireless sensor networks.

0002006003

Caricato da

Informazioni sul documento

Descrizione originale:

Copyright

Formati disponibili

Condividi questo documento

Condividi o incorpora il documento

Opzioni di condivisione

Hai trovato utile questo documento?

Questo contenuto è inappropriato?

Copyright:

Formati disponibili

0002006003

Caricato da

Copyright:

Formati disponibili

EURASIP Journal on Wireless Communications and Networking

Radio Resource Management

Guest Editors: Alagan Anpalagan, Rath Vannithamby,

Radio Resource Management

Radio Resource Management

Opportunistic Nonorthogonal Packet Scheduling in Fixed Broadband Wireless Access Networks,

Effective Radio Resource Management for Multimedia Broadcast/Multicast Services in UMTS

Cross-Layer Design and Analysis of Downlink Communications in Cellular CDMA Systems,

Alagan Anpalagan,1 Rath Vannithamby,2 Weihua Zhuang,3 and Sonia Aı̈ssa4

ON, Canada N2L 3G1

Received 9 July 2006; Accepted 9 July 2006

Space-Time Water-Filling for Composite MIMO

Received 1 September 2005; Revised 14 February 2006; Accepted 13 March 2006

1. INTRODUCTION capacity in MIMO Rayleigh fading channels with shadowing

User data Space- Space- User data

Figure 1: Point-to-point MIMO systems.

y = Hx + v, (1) The problem of spatial water-filling for MIMO Rayleigh fad-

p(λ)λ F(λ) = r(s)g(t)dt ds. (10)

where λk is the kth unordered eigenvalue of H† H, λ denotes ∞  

where m and Ω are coeﬃcients dependent on the MIMO sys-

as M → ∞. Simulations show that this approximation holds s

Space-time WF, numerical Spatial WF, simulated ρ = 16, simulated ρ = 8, numerical

Table 3: Comparison of space-time and spatial water-fillings.

Space-time water-filling Spatial water-filling

serves as an Associate Editor for the IEEE Transactions on Wire-

Brian L. Evans is the Mitchell Professor

Capacity Planning for Group-Mobility Users in OFDMA

Ki-Dong Lee and Victor C. M. Leung

Received 11 October 2005; Revised 28 April 2006; Accepted 26 May 2006

1. INTRODUCTION on each subcarrier. Also, they developed suboptimal solu-

where a ≈ −1.5/ log(5 BER) (BER denotes desired bit-error

minimize PO (y), This is supported by Proposition 2, namely, PS−1 (·) exists

Proof. 6. MINIMIZATION OF EXCESS CAPACITY RATIO

Proof. We have 6.1. Problem formulation: excess capacity

Here, α is a constant between 0 and 1, which denotes the

Table 1: Parameters used in experiments.

Item Value Description

Max. no. of connections, y

Max. no. of connections, y

Max. no. of connections, y

Figure 5: The maximum number of connections y versus γS with 1195

Figure 7: The maximum number of connections y versus γS with

Figure 6: The maximum number of connections y versus γS with

Max. no. of connections, y

1E 3 [3] M. Ergen, S. Coleri, and P. Varaiya, “Qos aware adaptive re-

1E 5 [5] Y. J. Zhang and K. B. Letaief, “Multiuser adaptive subcarrier-

ACKNOWLEDGMENTS Ki-Dong Lee received the B.S. and M.S.

Victor C. M. Leung received the B.A.S.

Joint Downlink Power Control and Multicode Receivers

Bessem Sayadi, Stefan Ataman, and Inbar Fijalkow

ETIS/ENSEA, University of Clergy-Pontoise/CNRS, 6 Avenue du Ponceau, 95014 Clergy-Pontoise, France

Received 30 September 2005; Revised 28 February 2006; Accepted 19 May 2006

1. INTRODUCTION codes in a synchronous multipath free channel. However,

Table 1: Terminology description. channel, corresponding to user k, is described as the follow-

Output SIR on each code, m = 1 . . . M1

From Figures 2 and 3, we observe the diﬃculty of the 4.5

of the MAI, ISI, and the intercode interferences. In the case

Transmit power’s ratio

After the convergence of the proposed strategy using a 74

In this paper, we have analyzed the benefits of combining 71

MIMO, OFDM, CDMA, etc.) and cross-layer optimization. Until

Adaptive Downlink Resource Allocation Strategies for

Navid Damji and Tho Le-Ngoc

where λk is the kth unordered eigenvalue of H† H, λ denotes ∞