Sei sulla pagina 1di 5

Implementation and Evaluation of Cooperative

Video Streaming for Mobile Devices


Marwan Ramadan, Layla El Zein, and Zaher Dawy
Department of Electrical and Computer Engineering
American University of Beirut
Beirut, Lebanon
Email: {mjr03, lmz07, zd03}@aub.edu.lb

Abstract— A promising approach for power reduction in mo- states (idle, long-range reception, small-range transmission).
bile devices is peer-to-peer cooperation. In this paper, we present Results show promising opportunities to decrease the power
a testbed system implementation to evaluate the performance consumed by increasing the number of collaborative peers.
of an infrastructure controlled cooperative video streaming
architecture. A number of mobile devices in close proximity In this work, we present a system testbed implementation
are connected to the Internet through a wireless access point for an infrastructure controlled peer-to-peer (P2P) network
and are all requesting the same video stream from a dedicated architecture. We consider a cooperative P2P scenario in which
server. Instead of sending the stream independently to each a server streams real-time video in a distributive fashion
device, the server distributes the video packets among the devices, among the peers which exchange the received packets over
over a long-range wireless link using WLAN technology; the
devices then exchange the received packets among each other short-range connections. Two main themes highlight the ben-
over short-range wireless links using Bluetooth technology. The efits of the implemented cooperative video streaming testbed:
implemented testbed is used to experimentally demonstrate the cooperation between mobile devices allowing for lower power
power reduction gains in mobile devices and to derive an consumption thanks to the lower power demands of short-
analytical model that relates power consumption to various range communication links between the peers, and a reduction
design and system parameters.
of traffic on the network’s side relieving the operator from
unicasting the same video stream to each device. Based
I. I NTRODUCTION
on the implemented setup, we obtain accurate experimental
Cooperation among mobile devices is expected to play measurement results to verify the anticipated benefits of such
a central role in the evolution towards 4G [1]. Allowing a cooperative architecture. Moreover, we derive a new cus-
mobile devices to cooperate among each other via peer-to-peer tomized theoretical model that relates the power consumed in
connections can lead to throughput enhancement, coverage the mobile devices to various design and system parameters.
extension, and power reduction. Studies show that the most This paper is organized as follows. In Section II, we discuss
power-consuming component in a mobile device is normally the system design and compare it with a traditional system
its communication interface [2], [3]. The authors in [4] pointing out the differences. Section III presents a detailed
demonstrate throughput gains by establishing a peer-to-peer technical description of the testbed implementation. Section IV
network architecture between end users of a mobile internet describes the power measurement setup and presents measure-
community requesting services over a cellular network. In ment results. Section V presents an analysis of the obtained
[5], [6], different architectures are presented and evaluated measurement results in addition to an analytical-based study.
by allowing multihop communications among mobile nodes Finally, conclusions are drawn in Section VI.
in cellular networks in order to achieve coverage extension,
power reduction, throughput enhancement, or load balancing. II. S YSTEM M ODEL
In [7], a novel cooperation architecture is proposed to reduce Offering real-time video streaming services over wireless
power consumption among mobile devices during real-time systems can be critical for end-users as well as for network
streaming. In this architecture, a number of wireless devices operators and service providers in terms of QoS, throughput,
interested in the same multicast service are distributed within fairness, load, and power consumption requirements. A typical
the coverage of a single wireless access point, reachable scenario would consist of a group of mobile devices connected
through a long-range link. Each terminal is assumed to be able to the Internet via wireless links and requesting a video
to communicate with the other terminals over short-range links stream from a dedicated online server. In a traditional setup,
with higher bit rate capability compared to the long-range link. the server either separately streams the complete video to
Based on the proposed cooperative approach, every terminal each requesting device or multicasts the video once to all
receives a subset of the stream from the access point over the requesting devices. In both cases, the communication interface
long-range link and sends it to its peers over the short-range of each mobile device remains active for the whole reception
links. An analytical study is presented based on typical values duration, which depends on the length of the video stream.
for power consumption of mobile devices during different This results in high power consumption due to the required
Fig. 1. Cooperative video streaming scenario. Fig. 2. Testbed implementation setup.

RF and baseband processing during data reception. In terms measurements on the mobile device batteries and use a data
of network throughput, a drawback of sending individual acquisition system (DAQ) to analyze the results.
video streams over the long-range link to the different mobile One of the testbed implementation milestones was to de-
devices is the redundant transmitted traffic. While multicasting velop the necessary algorithms for the cooperative network
alleviates this drawback, it remains an expensive solution for architecture with the necessary intelligence at the server and
operators and might not be supported by some long-range the mobile devices to fulfill their expected roles. The server has
wireless technologies. It is worth noting that multicasting to distribute the data among the devices while respecting real-
results in the same incoming traffic at mobile devices and, time constraints, and the devices must cooperate to exchange
thus, does not have an impact on power consumption at the and process data streams. This behavior is illustrated by the
communication interface during video reception. packet sequence numbers in Figure 1.
In this work, we assume the establishment of a P2P network
between the mobile devices over short-range wireless links A. Frame Distribution at the Server
that are more power-efficient than the long-range wireless link. We used the J2SE platform to develop the server application.
In this scheme, each mobile device receives from the server The server awaits for a certain number of clients N to
over the long-range wireless link only a subset of the total connect to it on the TCP socket then starts transmitting the
data stream. If N devices are cooperating, then, each will be video frames in a round-robin fashion to the clients over
receiving one of N subsets from the server. This reduces the WLAN. The video format used is MJPEG where each MJPEG
power consumed at the long-range communication interface by frame is composed of a size header and a JPEG image.
a factor of N . Over the short-range wireless links, each device Using MJPEG is motivated by the codec design simplicity
receives the remaining data subsets from the other devices at the client side. The aim of this work is to evaluate the
in the P2P network. Being exchanged over a power-efficient power consumption of cooperative video streaming and not to
short-range wireless technology, the (N − 1) subsets require optimize the performance of the video codec. The frame size
lower reception power at the communication interface of the and the transmission rate are controlled at the server input
devices. However, an additional overhead in this case is that parameters that include the number of clients, the packet size
each mobile device needs to spend additional power to transmit (as in number of frames per packet), and the frame duration.
its received data subset to (N − 1) other devices. Therefore, This enables the opportunity to adjust the video play speed at
it is not directly clear whether the aggregate power consumed the client and to ensure that WLAN and Bluetooth bit rates
at the mobile devices would be less or more when compared are of the same order. From a performance perspective, this
to traditional video streaming architectures. would avoid overflowing of the buffer at the client side. Frames
buffered at the server are sent in UDP datagrams according to
III. T ESTBED I MPLEMENTATION
the specified frame rate in a round-robin fashion to the clients.
To experimentally address this question, we implement a A buffering algorithm is implemented at the clients in order
prototype cooperative video streaming testbed as illustrated in to reorder frames and reduce jitter.
Figure 2. A server streams to the devices real-time video in a
distributive fashion over a long-range wireless link, which we B. Distributed Processing at the Clients
experimentally implement using WLAN technology. The peers We implemented the client application on an HP iPAQ
will exchange the received packets over short-range wireless Pocket PC using Visual C# with the .Net Compact Frame-
links which we experimentally implement using Bluetooth work 2. To implement the Bluetooth functionalities, we used
technology. The aim is to verify whether using power-efficient the Franson Bluetools library.
short-range links would achieve the expected gains in terms of 1) Establishment of the P2P Links: Devices must establish
power reduction, and to anticipate the impact of short-range connections between them over Bluetooth in order to be able
throughput capability on performance. We conduct power to cooperate. We have chosen to implement the simplest
Bluetooth topology, a single piconet, in which one device
initiates inquiries (the master) and the others listen (slaves) [8].
The establishment of the P2P network is threefold, and aims
at restricting the P2P network membership to those devices
that are interested in cooperation. The main stages include
advertisement, discovery, and initiation. Each client starts
by advertising a Bluetooth serial streaming service with a
particular identifier. The cooperative peers are identified by
the services they advertise. Each client intending to request the
video discovers the existing cooperative peers in the vicinity.
The first client to actually initiate a Bluetooth connection
with each peer becomes the master in the piconet and all the Fig. 3. Power measurement setup implementation.
subsequent P2P traffic will be routed through it. It is assumed
that clients stay connected throughout the streaming duration.
2) Data Exchange over the P2P Network: After the setup of through the master, which implies higher levels of processing
the P2P network over Bluetooth links, streaming of the video and a higher power consumption at the master device. In this
can be initiated. Each client contacts the server through a TCP setup, we monitor the power consumption on a slave device.
socket and waits for incoming packets on a different UDP First, we are interested in comparing the power consumed
socket. When all the other peers have contacted the server, at the wireless interfaces of WLAN and Bluetooth, since it
the latter starts its round-robin streaming which triggers the is at the core of the perceived benefits of the system imple-
reception of the UDP datagrams at the mobile devices. Each mentation. Then, we observe the power consumed during the
packet received over WLAN from the server is then directly video streaming and we compare the implemented cooperative
sent to each connected peer over the Bluetooth links. Concur- architecture with the traditional architecture.
rently, packets over Bluetooth are also received in a separate V. R ESULTS AND A NALYSIS
thread at the clients. Each received packet is depacketized and
passed to a reordering algorithm that either buffers or directly We conduct several power measurements in various scenar-
displays the incoming frames according to certain quality of ios and display the captured instantaneous results. Then, we
service parameters to minimize delay and jitter. identify and separate the different components that contribute
3) Video Decoding at the Client: In order to enable MJPEG together to the overall aggregate average power consumption
video decoding at the clients, a special video player was in the device over a fixed time window (see Table I). The
implemented. Each frame after buffering is depacketized into processing power refers to the power consumed at the PDA
a header and a JPEG image. The header contains the image by the processing of the video frames prior to and during
size which is used to read and display the JPEG frame in a the display. Though peers are assigned threads to send and
picture box on the client’s GUI. We used some of the graphical receive frames over the P2P links, the power consumed by
capabilities of Windows Mobile, especially double-buffering, the multithreading remains minimal as the number of peers N
to prevent flickering of the images. Double-buffering keeps the increases, and the bulk of the processing activities is found
present image displayed until the next one is loaded to avoid to be due to frame processing. The power consumption due
blanks during the transition. to the reception and transmission activities is included as part
of the transmission and reception powers and not as part of
IV. P OWER M EASUREMENTS the processing power. Since the number of processed frames
Well known techniques are used to accurately measure the is the same for a given frame rate, regardless of the number
energy consumption of a mobile device. A typical one consists of cooperating entities in the network, the processing power is
of powering up the device with its own battery and calculating assumed to be independent of N . Another observation worth
the static power drain by measuring the voltage drop across
a known resistor on the line between the battery and the TABLE I
device [9]. The setup, shown in Figure 3, consists of one M EASUREMENT RESULTS .
mobile device (PDA) and an oscilloscope which we use to
State Average Power (W)
observe the voltage variations. The PDA is an HP iPAQ h6340,
powered by a 4 V battery via four pins. We use a DAQ Idle Mode 0.188
to collect the instantaneous voltage values across a resistor Processing 0.095
that connects the positive pin of the battery to that of the Receiving over WLAN 0.500
device. The iPAQ has a built-in 802.11b WLAN interface and Receiving over Bluetooth 0.015
a Bluetooth interface. Transmitting over Bluetooth 0.014
The testbed includes two types of devices: the Bluetooth Wireless Medium Average Bit Rate (kbps)
master and the Bluetooth slaves. The main difference from a Long Range (WLAN) 720
behavioral point of view is that traffic between slaves must go Short Range (Bluetooth) 221
mentioning is the dependency of the Bluetooth transmission 1
Traditional Scenario

power on distance and channel conditions. The HP iPAQ

Power [W]
h6340 implementation of Bluetooth does not comprise power 0.5

control which is an optional feature in the Bluetooth stan- 0


dard [8]. This resulted in a constant transmission power on 0 200 400
Time [ms]
600 800 1000

Cooperating Scenario − Two Peers


the Bluetooth interface. 1

Power [W]
A. Measurement-Based Study 0.5

It can be seen from the measurement results presented in 0


0 200 400 600 800 1000
Time [ms]
Table I that reception over WLAN is 33 times that required by Cooperating Scenario − Three Peers
1
either reception or transmission over Bluetooth. This is the key

Power [W]
observation that is expected to play a major role in making 0.5

the cooperative video streaming architecture power-efficient.


0
Moreover, measurements show that the WLAN interface is 0 200 400 600 800 1000
Time [ms]
constantly awakening to listen to periodic signals sent by
the wireless access point. These beacons notify the interface Fig. 5. Measurement results: Power consumption in traditional and cooper-
whenever data is expected over the wireless medium. The ative architectures.
peaks in Figure 4 correspond to the power required to power
up the interface at each beacon. The transitions from idle, to each period is causing a relatively small increase in the power
the Bluetooth Interface ON state, to the WLAN Interface ON floor. We shall note here that in a cooperative scenario with
state are obviously accompanied by an increase in the power N peers, each beacon in Figure 5 is synchronized with data
floor, which makes the mere activation of the WLAN interface exchange over Bluetooth between the device and one of its
expensive in terms of power consumption. (N − 1) peers.
In the measurement setup, we have programmed the server The resulting aggregate average powers over the measure-
to send data every 100 ms which is the beacon period at the ment time window are 0.394 W, 0.351 W, and 0.330 W
wireless access point. Therefore, in a traditional architecture, respectively for one, two, and three peers where we set the
where the complete data stream is sent to the device, data access point throughput to 1 Mbps. This proves that actual
is received and decoded at each beacon. However, in the power reduction gain in mobile devices can be achieved via
implemented cooperative architecture, data is received over the implemented cooperative video streaming architecture.
WLAN every N beacon periods. In Figure 5, we plot the The power reduction gain reaches around 16% with three
measured power consumption profile as a function of the cooperating devices.
number of cooperating peers (devices). The wide peaks cor-
B. Analytical-Based Study
respond to data reception, while the narrow ones correspond
to beacon reception. Results show that with three peers, there This study is conducted for three main reasons: To support
is a wide peak only every three periods, which reduces the the presented measurement results, to derive an expression
power consumed over the WLAN interface by around three for the additional power consumed at the master device
times compared to the traditional architecture (case of one due to traffic routing, and to obtain a generic model that
peer). Another important observation in Figure 5 is that the can be used to calculate power consumption for cooperative
power consumed by Bluetooth transmission and reception at video streaming as a function of different system and design
parameters. Based on the analysis of the measurement results
in different scenarios and on the identified power consumption
components presented in Table I, we derive the following
expression that relates aggregate power consumption to various
parameters:
µ ¶
F S PL,Rx (N − 1)
PAv = + (PS,Tx + PS,Rx )
N RL RS
+Pprocessing + Pidle (1)
where F S/(N RL ) is the reception time over the WLAN
interface for a given device, F SPL,Rx /(N RL ) is the energy
spent by a device during reception over WLAN, F S/(N RS ) is
the reception or transmission time over the Bluetooth interface
for a given device, (N − 1)F S(PS,Tx + PS,Rx )/(N RS ) is
the energy spent by a device due to transmitting the frame
N − 1 times to other peers and receiving N − 1 frames from
Fig. 4. Measurement results: Power transitions in idle mode. other peers over Bluetooth, Pprocessing is the processing power
TABLE II
T HEORETICAL M ODEL PARAMETERS .
Input Description
RL The long-range bit rate
RS The short-range bit rate
F The video frame rate in frames per second
S The average video frame size in bits
FS The number of bits sent in one second
N The number of cooperative peers
PL,Rx The average power consumed by the long-range
interface during reception
PS,Rx The average power consumed by the short-range
interface during reception Fig. 6. Analytical aggregate power consumption as a function of the number
PS,Tx The average power consumed by the short-range of cooperative peers N . LR stands for long range link (WLAN), SR stands
interface during transmission for short range link (Bluetooth), SR routing component applies only to the
case of the mobile device acting as Bluetooth piconet master.

VI. C ONCLUSIONS
which is nearly constant for various values of N as discussed
before, and Pidle is the power consumed during idle state (no We implement and evaluate a testbed for a cooperative video
communication) which is independent of N . streaming architecture in order to obtain power consumption
This model applies to all client devices in the network. measurement results and to derive a theoretical expression
For the master device in the Bluetooth piconet, an additional that relates power consumption to various design and system
component Prouting should be added to PAv , in order to parameters. Presented results prove that allowing peer-to-peer
account for the additional amount of power consumed at the cooperation among mobile devices notably reduces power con-
master due to routing which involves receiving frames from sumption for video streaming without compromising the video
a source client and forwarding them to a destination client. It quality. This is important due to the limited battery capabilities
can be calculated as: of mobile devices. In addition to the power reduction gains,
the implemented cooperative video streaming architecture can
FS result in throughput gains to network operators by reducing
Prouting = (N − 1)(N − 2) (PS,Tx + PS,Rx ) (2) the required traffic over long-range links and can result in
N RS
reduced service costs for end users who will be downloading
In addition to the different power components, the model less content directly from the service provider.
takes into account the number of cooperating devices, and the R EFERENCES
active duration of the wireless interface which is a function of
[1] S. Y. Hui and K. H. Yeung, Challenges in the Migration to 4G Mobile
the link bit rate and the data size. All parameters are defined Systems, IEEE Comm. Magazine, vol. 41, no. 12, pp. 5459, Dec. 2003.
in Table II. Using the measurement values from Table I, we [2] R. Kravets and P. Krishnan, Application-driven Power Management for
plot in Figure 6 the total average power consumption at the Mobile Communication, Wireless Networks Journal, vol. 6, pp. 263-277,
2000.
master mobile device as an aggregate of the different power [3] M. Katz and F. Fitzek, Cooperation in Wireless Networks: Principle and
consumption components versus the number of peers N . We Application. Cooperation in 4G Networks, Springer, pp. 463-496, 2006.
exclude the processing and idle power components since they [4] Y.-D. Lin and Y.-C. Hsu, On Using Peer-to-Peer Communication in
Cellular Wireless Data Networks, IEEE Trans. on Mobile Computing,
are assumed constant for various values of N . The same plot vol. 3, no. 1, Jan.-March 2004.
can be used to observe the aggregate power consumption at the [5] H. Wu, C. Qiao, S. De, and O. Tonguz, Integrated Cellular and Ad Hoc
slave mobile devices, by just excluding the upper part of the Relaying Systems: iCAR, IEEE J. Selected Areas in Comm., vol. 19, no.
10, pp. 2105-2115, Oct. 2001.
bars referring to SR routing. As N increases, we notice that [6] H. Luo, R. Ramjee, P. Sinha, L. Li, and S. Lu, UCAN: A Unified Cellular
the long-range reception power decreases at a faster pace than and Ad-Hoc Network Architecture, Proc. ACM MOBICOM 2003, Sept.
the increase in short-range powers. Results show that power 2003.
[7] F. Fitzek and P. Kyritsi and M. Katz, Cooperation in Wireless Networks:
reduction gains of up to 20% can be achieved by having five Principle and Application. Power Consumption and Spectrum Usage
cooperating peers. It is important to note that the analytical Paradigms in Cooperative Wireless Networks, Springer, pp. 365-386,
model results fit perfectly with the measured results for the 2006.
[8] Specification of the Bluetooth system, Bluetooth Core Specification v2.1
slave (cases with one, two, and three peers). This proves the + EDR.
accuracy and usability of the derived analytical expressions. [9] R. Lee and R. Nathuji, Power and Performance Analysis of PDA Archi-
As for the master, we notice that the routing power increases tectures, Advanced VLSI Computer Architecture, Dec. 2000.
[10] S. Gurun and P. Nagapurkar and B.y. Zhao, Energy Consumption and
as the number of cooperating entities increase in the network. Conservation in Mobile Peer-to-Peer Systems, Technical Report, UC
This centralization issue can be tackled by devising a scatternet Santa Barbara, 2006.
formation algorithm that ensures distributed routing within the [11] F. Fitzek, M. Rossi and M. Zorzi, Error Control Techniques for Efficient
Multicast Streaming in UMTS Networks, Proc. SCI 2003, July 2003.
P2P network.

Potrebbero piacerti anche