Sei sulla pagina 1di 4

TCP performance issues in LTE networks

Hyun-Seo Park Jae-Yong Lee, Byung-Chul Kim


Internet Future Technology Research Department Dept. of Information Communications Engineering
ETRI Chungnam National University (CNU)
Daejeon, KOREA Daejeon, KOREA
hspark@etri.re.kr jyl@cnu.ac.kr, byckim@cnu.ac.kr

Abstract We describe issues on the TCP performance in LTE network conditions. The test will measure the achieved average
networks. TCP throughput is determined by various factors such application-layer data rates using Transmission Control
as TCP window size, loss probability and RTT. We present Protocol (TCP) or User Datagram Protocol (UDP) in High
measurement results of TCP throughput in a real LTE testbed Speed Packet Access (HSPA) and LTE networks [3].
with various factors and TCP auto-tuning enabled. As expected,
the test results show that the AM bearer outperforms the UM Although TCP performance over various wireless networks
bearer due to less loss probability in the former. However, test has been heavily studied in the literature, different wireless
results show that delay spikes incurred by the RLC ARQ environment may introduce new challenges to TCP. TCP
increase TCP RTT and degrade TCP performance. We propose window auto-tuning is the default feature on various operating
some enhancements for improving TCP throughput in LTE systems such as Windows Vista, Windows 7, Windows 2008
networks. Server and Linux, and so on. However, TCP performance over
wireless networks with auto-tuning enabled is almost never
Keywords LTE, TCP, ARQ, HARQ, RLC, Reordering, Auto-
mentioned. Reference [4] is the only one that discusses TCP
Tuning.
performance with auto-tuning enabled. It shows the TCP
throughput measurement of TCP variants over commercial
I. INTRODUCTION WiMAX networks with auto-tuning enabled. The results
rd indicate that the WiMAX link does not adversely affect the
The 3 Generation Partnership Project (3GPP) specified the
intended operation of TCP variants, but that it is not well-suited
Long Term Evolution (LTE) radio interface to meet the
for the aggressive Cubic and window auto-tuning. Therefore it
increasing performance requirements of mobile broadband.
concludes that TCP Cubic and window auto-tuning are not
LTE provides a peak data rate of up to 300Mbps in the
recommended for connections traversing WiMAX links.
downlink and 75Mbps in the uplink with 20MHz bandwidth.
Currently, the 3GPP is specifying the LTE-Advanced radio In this paper, we describe issues on the TCP performance in
interface to meet the IMT-Advanced requirements. LTE- LTE networks. Section II briefly introduces the LTE radio
Advanced provides a peak data rate of up to 1Gbps in the interface focusing on the user plane protocol, including
downlink with peak spectral efficiency 30bps/Hz and 500Mbps Medium Access Control (MAC) Hybrid Automatic Repeat
in the uplink with peak spectral efficiency 15bps/Hz using up request (HARQ) and Radio Link Control (RLC) Automatic
to 100MHz bandwidth [1]. Repeat request (ARQ). Section III discusses TCP performance
factors such as TCP window size, loss probability, and Round
LTE is now on the market. In July 2011, an update from the
Trip Time (RTT). We also show the test results in the ETRI
Global Mobile Suppliers Association (GSA) confirms that 218
LTE-Advanced system with varying TCP performance factors
operators in 80 countries now make investments in LTE
in LTE links and TCP auto-tuning enabled. To the best of our
networks. The organization's latest update states that 166 firm
knowledge, this work is one of the first detailed measurements
commercial LTE network deployments are in progress or
of TCP performance on a real LTE-Advanced system and on
planned in 62 countries. This includes 24 networks which have
LTE-family networks with TCP auto-tuning enabled. Section
commercially launched, while a further 52 operators in 19
IV discusses a discovered problem in which delay spikes
additional countries are engaged in LTE technology trials, tests
incurred by MAC HARQ reordering and RLC ARQ increase
or studies.
TCP RTT and degrade TCP performance. Finally, we propose
In January 2011, ETRI demonstrated the worlds first LTE- some enhancements to alleviate the problem and improve TCP
Advanced system. The system provides a peak data rate of up throughput in LTE networks as a result. Some concluding
to 600Mbps and a peak effective data rate, i.e., application remarks, including future works, are given in section V.
layer throughput, of up to 440Mbps in the downlink and
220Mbps in the uplink with 40MHz bandwidth [2]. II. BACKGROUND INFORMATION

In December 2010, the 3GPP approved the new Release 11 A. LTE User Plane Protocol
study item, UE Application Layer Data Throughput In the LTE user plane, the PHY layer provides a bit pipe
Performance. The Global Certification Forum (GCF) with Adaptive Modulation and Coding (AMC) by protected
indicated that they want 3GPP to perform UE application-layer turbo-coding and a cyclic redundancy check (CRC), and
data throughput measurements under various simulated processes CRC ACK/NACK feedback. The MAC sub-layer

This work was supported by the IT R&D program of MKE/KEIT.


[KI001484, Research on service platform for the next generation mobile
communication].

978-1-4577-1268-5/11/$26.00 2011 IEEE 493 ICTC 2011


provides multiplexing/de-multiplexing, error correction III. TCP PERFORMANCE IN LTE NETWORKS
through HARQ, and scheduling. The RLC sub-layer provides We measured the TCP throughput in the ETRI LTE-
the segmentation/concatenation/reassembly of RLC SDUs, Advanced system through varying TCP performance factors
reordering of RLC PDUs, in-sequence delivery, and error such as TCP window size, loss probability, and RTT in LTE
correction through ARQ. The Packet Data Convergence links. In the test, the TCP client in the UE is on Windows OS
Protocol (PDCP) sub-layer provides header compression, and the TCP server is on Linux. The TCP on Linux uses Cubic
ciphering/integrity protection, and in-sequence delivery and TCP as the default congestion control with auto-tuning enabled.
retransmission of PDCP SDUs at handover [5]. Fig. 1 shows In our testbed, the RTT between the TCP client and TCP server
the LTE user plane protocol stack. is about 13ms. Also, a UE can use two carriers, and the
maximum bandwidth between a UE and eNB is 110Mbps per
carrier, so the maximum bandwidth per UE is 220Mbps in total.
The Robust Header Compression (ROHC) function at the
PDCP sub-layer is disabled in the test. We show the test results
and discuss the TCP performance factors one by one. We
repeated each test ten times and measured the average TCP
throughput and utilization. Table I shows the test parameters.

TABLE I. TEST PARAMETERS

FTP File size 700.74MB AVI file


Figure 1. LTE user plane protocol stack TCP segment size (MSS) 1380Bytes
TCP Receive Window 65535Bytes (Windows XP)
If the receiver detects a transmission error, the receiver
Size (rwnd) Auto-tuning (Windows 7)
sends CRC NACK feedback to the sender. When the MAC
sub-layer HARQ function of the sender receives CRC NACK UE IP connection IPv6
feedback from the receiver, it performs retransmissions of the Bandwidth 220Mbps
corrupted transport blocks, and thereby corrects the majority of PDCP/RLC/MAC Header Under 0.005% (Minimum)
all transmission errors. The HARQ protocol uses eight stop- overhead
and-wait HARQ processes, and HARQ RTT in the uplink is Max HARQ TX 4
8ms and that in the downlink is minimum 8ms [6]. Max ARQ ReTX 5
LTE supports two types of bearer, a Unacknowledged HARQ BLER Under 0.1%
Mode (UM) bearer and an Acknowledged Mode (AM) bearer. HARQ RTT 8ms
The UM bearer makes use of a fail recovery by the HARQ at PUSCH Always scheduled
the MAC sub-layer only, and thus can be used for radio bearers RF Bandwidth 40MHz (20MHz*2 Carriers)
that can tolerate a certain amount of loss. The AM bearer
RF Frequency 3.56 ~ 3.60GHz
provides more reliability than the UM bearer by the help of
ARQ retransmission at the RLC sub-layer. Delay bet. eNB and FTP Under 5ms
server
If the RLC sub-layer receiver detects a gap in the sequence
of the received PDUs, it starts a reordering timer assuming that A. TCP Window Size
the missing PDU still is being retransmitted in the HARQ The achievable TCP throughput is bounded at [Throughput
protocol. HARQ failures appear if a maximum number of <= TCP rwnd/RTT], where rwnd is the TCP receive window
HARQ transmission attempts are exceeded or HARQ feedback size and RTT is the round-trip-time of the path. To optimize
NACK-to-ACK errors occur. When the timer expires, usually the TCP throughput, the receive window is greater than [path
in a HARQ failure case, an RLC UM receiver delivers SDUs to bandwidth * RTT], which is known as the bandwidth-delay
PDCP with a certain amount of loss. However, an RLC AM product (BDP). Unfortunately, Windows XP TCP/IPv6 stack
receiver sends a status message comprising the sequence does not support window scaling, and thus the maximum
number of the missing PDUs to the sender. The ARQ function receive window size is only 65,535 bytes. As a result, the
of the RLC AM sender performs retransmissions based on the achievable TCP throughput is bounded at about 40Mbps.
received status message. The RLC ARQ protocol is a window-
based selective repeat ARQ [7], [8]. Next-generation TCP/IP stacks in Windows Vista and
Windows 7 support TCP receive window auto-tuning and no
Thus, the RLC UM is a high-persistent ARQ protocol longer use the TCPWindowSize registry value. TCP auto-
and RLC AM is a perfectly-persistent ARQ protocol except a tuning automatically determines the optimal receive window
Radio Link Failure (RLF) event. TCP applications can usually size on a per-connection basis. Thus, the user no longer needs
be mapped on the AM bearer due to greater reliability. to manually configure a TCP receive window for a specific
However, the ARQ persistence of the RLC AM ARQ increases connection or specific computers. It determines the optimal
latency and delay jitter to application data, and adversely receive window size by measuring the BDP and the application
delays the flow of TCP control information [9]. retrieve rate allowing up to a 16MB maximum receive window
size using the TCP window scaling option [10]. Through the
help of the auto-tuning feature, the measured TCP throughput

494
is around 200Mbps with the RLC PDU loss probability under RTT incurred by RLC ARQ retransmissions, as the achievable
0.001%. TCP throughput is inversely proportional to the RTT.
In conclusion, the UE should support the TCP window In the LTE, layer 2 only provides in-sequence delivery of
scaling option to obtain a better performance if the BDP of a SDUs to the upper layer. If an RLC AM receiver detects a gap
TCP connection is larger than 65,535 bytes. in the sequence of received PDUs, it starts a reordering timer
(t_Reordering) assuming that the missing PDU is still being
B. Loss Probability retransmitted in the HARQ protocol. When t_Reordering
The achievable TCP throughput is bounded at [Throughput expires, usually in a HARQ failure case, it sends a status
<= MSS/(RTT*sqrt(Ploss))], where the MSS is the maximum message comprising the sequence number of the missing PDUs
segment size, RTT is the round-trip-time of the path, and Ploss to the sender. The ARQ function of RLC AM sender performs
is the probability of packet loss. We measured the TCP retransmissions based on the received status message. After the
throughput of the UM and AM bearers by varying the RLC gap is filled by the ARQ retransmission, it delivers
PDU loss probability from 0.001% to 1%. In the UM bearer, reassembled RLC SDUs to the upper layer in sequence.
the probability of a packet loss is much the same as the RLC Therefore, the TCP RTT of a packet that is retransmitted by the
PDU loss probability. However, in the AM bearer, the RLC ARQ is proportional to the RLC ARQ RTT.
probability of a packet loss in the wireless last-hop between
eNB and UE is nearly 0% through the help of the RLC ARQ. The RLC ARQ RTT is influenced by many factors such as
t_Reordering, status prohibit timer (t_StatusProhibit), and UL
Fig. 2 shows the measured TCP throughput of the UM scheduling (e.g., scheduling request, buffer status reporting).
bearer vs. AM bearer. As expected, the AM bearer almost We measured the TCP throughput of the AM bearer with
always achieves better throughput than the UM bearer due to varying t_Reordering of 30ms, 15ms, and 0ms. Actually, 0ms
less loss probability. The UM bearer shows much the same t_Reordering is not suitable, but we used it anyway as a
throughput as the AM bearer with loss probability on the order reference of the minimum TCP RTT with RLC ARQ.
of 10-5. With the loss probability on the order of 10-5, the
utilization of the AM bearer is almost 90%, and that of the UM Figure 3 shows the measured TCP throughput of the AM
bearer is nearly 87%. The missing 10% in utilization is due to bearer with t_Reordering of 30ms, 15ms, and 0ms. As expected,
the time period of the TCP slow start phase in the TCP server the AM bearer with t_Reordering of 0ms almost always
and the time taken to determine the optimal TCP receive achieves better throughput than the larger t_Reordering case
window size by auto-tuning in the TCP client. with every loss probability. The 0ms t_Reordering case shows
4%, 9%, and 32% gain compared with 30ms t_Reordering with
Compared with that of AM bearer, the utilization of the a loss probability on the order of 10-4, 10-3, and 10-2,
UM bearer is 75% with a loss probability on the order of 10-4, respectively. The 15ms t_Reordering case shows 3%, 2%, and
25% with a loss probability on the order of 10-3, and at worst, 10% gain compared with 30ms t_Reordering with loss
only 12% with a loss probability on the order of 10-2. From an probability on the order of 10-4, 10-3, and 10-2, respectively.
LTE design, the HARQ feedback NACK-to-ACK error
probability is on the order of 10-4 and the RLC PDU loss In conclusion, as we decrease the RLC ARQ RTT, the
probability after the HARQ is on the order of 10-3. Thus, the throughput of TCP applications can be increased.
UM bearer is not appropriate for TCP applications.
In conclusion, TCP applications should be serviced on the
AM bearer to obtain better performance.

Figure 3. TCP throughput of AM bearer


with variable t_Reordering of 0ms, 15ms, 30ms

IV. ISSUES AND CONSIDERATIONS


As shown previously, the TCP throughput of the AM bearer
is degraded in proportion to the RLC PDU loss probability
Figure 2. TCP throughput of UM bearer vs. AM bearer
with variable loss probabilities of 10-5, 10-4, 10-3 and 10-2
despite nearly no application-layer PDU loss. This is due to
increased TCP RTT incurred by RLC ARQ retransmissions.
C. RTT We discuss two points for decreasing TCP RTT, that is, a way
to decrease RLC ARQ RTT and a way to decrease TCP RTT
In Fig. 2, the TCP throughput of the AM bearer is degraded only. Also, we discuss the behavior of the TCP sender with
in proportion to the RLC PDU loss probability despite nearly
no application-layer PDU loss. This is due to increased TCP

495
TCP auto-tuning enabled when the TCP ACK suffers delay the TCP performance. We propose some enhancements for
spikes. improving the TCP throughput in LTE networks. First, a local
NACK decreases the TCP RTT as well as the RLC ARQ RTT.
We can decrease the RLC ARQ RTT by introducing a local Second, an out-of-sequence delivery helps decrease the delay
NACK [11], [12], but the gain from decreasing the RLC ARQ spikes of TCP data and TCP ACK compression. Finally, an
RTT is not great. The local NACK has HARQ-assisted ARQ out-of-sequence delivery can help the TCP sender interpret
mechanisms. There are three mechanisms for a local NACK. the available bandwidth and RTT of the TCP connection
First, if the HARQ transmitter detects a HARQ failure, the exactly. For future work, we will study and test extensively an
transmitter ARQ entities are notified and retransmissions can out-of-sequence delivery in LTE networks with TCP auto-
be initiated. Second, if the HARQ receiver detects a NACK to tuning enabled.
ACK error, the transmitter ARQ entities are notified via
explicit signaling. Third, if the HARQ receiver detects a ACKNOWLEDGMENT
HARQ failure, the transmitter ARQ entities are notified. A
local NACK clearly can be a help in decreasing the RLC ARQ We thank all members who actively participated in tough
RTT, but the effect of a local NACK is similar to the use of but fruitful development and testing of the worlds first
t_Reordering between 30ms and 15ms. From the previous RTT operational LTE-Advanced System demonstrated outside lab
section, the gain from using 15ms t_Reordering is under 10%. conditions.
Also, simulation results from [11] show that the gain from a REFERENCES
local NACK is marginal.
[1] The 3rd Generation Partnership Project (3GPP), http://www.3gpp.org/,
We can decrease the TCP RTT by introducing an out-of- online.
sequence delivery into LTE layer 2. An in-sequence delivery [2] ETRI LTE-Advanced system, http://www.etri.re.kr/eng/,
of RLC and PDCP in AM mode increases the TCP RTT of all http://eng.kcc.go.kr/download.do?fileSeq=30503, online.
SDUs in every PDU from the SN of the retransmitted PDU. [3] 3GPP TR 37.901, UE Application Layer Data Throughput Performance
Therefore, this frequently incurs delay spikes in the TCP data Study Item Technical Report, V1.0.0, June 2011.
and TCP ACK compression. On the other hand, an out-of- [4] E. Halepovic, Q. Wu, C. Williamson and M. Ghaderi, TCP over
WiMAX: A measurement study, IEEE International Symposium on
sequence delivery can increase the TCP RTT of the SDUs in Modeling, Analysis and Simulation of Computers and
the retransmitted PDU only. However, an out-of-sequence Telecommunication Systems (MASCOTS), pp. 1-10, 2008.
delivery is not included in the LTE specification due to a lack [5] 3GPP TS 36.300, Evolved Universal Terrestrial Radio Access (E-
of study on TCP performance of a UE when TCP auto-tuning UTRA) and Evolved Universal Terrestrial Radio Access Network (E-
is enabled. Also, an out-of-sequence delivery incurs packet UTRAN); Overall description; Stage 2, V10.4.0, June 2011.
reordering in the upper layers more, but packet reordering is [6] 3GPP TS 36.321, Evolved Universal Terrestrial Radio Access (E-
not a rare event in current high-speed networks, and there are UTRA); Medium Access Control (MAC) protocol specification,
V10.2.0, June 2011.
some mechanisms to alleviate the reordering problem [13]. We
also discovered that the TCP sender stops sending data when a [7] 3GPP TS 36.322, Evolved Universal Terrestrial Radio Access (E-
UTRA); Radio Link Control (RLC) protocol specification, V10.0.0,
delay spike of TCP data is incurred by an RLC ARQ and the Dec. 2010.
delay of the TCP ACK is large. We consider this to be due to [8] A. Larmo, M. Lindstrm, M. Meyer, G. Pelletier, J. Torsner and H.
the fact that the TCP sender misinterprets an increased TCP Wiemann, The LTE link-layer design, IEEE Communications
RTT as an indication of decreased bandwidth with the same Magazine, vol. 47, no. 4, pp. 52-59, April 2009.
BDP or possible congestion. Therefore an out-of-sequence [9] G. Fairhurst and L. Wood, Advice to link designers on link Automatic
delivery can help relieve a misinterpretation of the TCP Repeat reQuest (ARQ), RFC 3366, Aug. 2002.
sender. [10] J. Davies, The Cable Guy: TCP receive window auto-tuning,
http://msdn.microsoft.com/en-us/library/2007.01.cableguy.aspx, online.
V. CONCLUSIONS [11] 3GPP R2-062772, Simulation results on HARQ assisted ARQ scheme,
Samsung, RAN2#55, Oct. 2006.
In this paper we described the TCP performance in LTE
[12] D. M. Kim, Y. K. Choi, S. G. Jin, K. H. Han and S. H. Choi, A
networks. We measured the TCP throughput in a real LTE MAC/PHY cross-layer design for efficient ARQ protocols, IEEE
testbed based on various factors and TCP auto-tuning enabled. Communications Letters, vol. 12, no. 12, pp. 909-911, Dec. 2008.
As expected, the test results show that the AM bearer [13] K. C. Leung, V. O.K. Li and D. Yang, An overview of packet
outperforms the UM bearer due to less loss probability in the reordering in TCP: problems, solutions, and challenges, IEEE
former. However, the test results show that delay spikes Transactions on Parallel and Distributed Systems, vol. 18, no. 4, pp.
incurred by the RLC ARQ increase the TCP RTT and degrade 522-535, April 2007.

496

Potrebbero piacerti anche