INFOCOM 2003Return to Main MenuThe Impact of Multihop Wireless Channel on TCPThroughput and LossZhenghua Fu, Petros Zerfos, Haiyun Luo, Songwu Lu, Lixia Zhang, Mario GerlaUCLA Computer Science Department,Los Angeles, CA 90095-1596Email:{zfu,pzerfos,hluo,slu,lixia,gerla}@cs.ucla.eduAbstract— This paper studies TCP performance over multihopwireless networks that use the IEEE 802.11 protocol as the accessmethod. Our analysis and simulations show that, given a specificnetwork topology and flow patterns, there exists a TCP windowsize W∗, at which TCP achieves best throughput via improvedspatial channel reuse. However, TCP does not operate aroundW∗, and typically grows its average window size much larger;this leads to decreased throughput and increased packet loss. TheTCP throughput reduction can be explained by its loss behavior.Our results show that network overload is mainly signified by wireless link contention in multihop wireless networks. As longas the buffer size at each node is reasonably large (say, largerthan 10 packets), buffer overflow-induced packet loss is rare andpacket drops due to link-layer contention dominate. Link-layerdrops offer the first sign for network overload. We further showthat multihop wireless links collectively exhibit graceful dropbehavior: as the offered load increases, the link contention dropprobability also increases, but saturates eventually. In general,the link drop probability is insufficient to stabilize the averageTCP window size around W∗. Consequently, TCP suffers fromreduced throughput due to reduced spatial reuse. We furtherpropose two techniques, link RED and adaptive pacing, throughwhich we are able to improve TCP throughput by 5% to 30%in various simulated topologies. Some simulation results are alsovalidated by real hardware experiments.I. INTRODUCTIONTCP is an adaptive transport protocol that controls itsoffered load (through adjusting its window size) according tothe available network bandwidth. It additively increases itscongestion window in the absence of congestion and throttlesdown its window when a sign of congestion is detected. In thewired Internet, congestion is identified by packet loss, whichresults from buffer overflow events at the bottleneck router.However, it is unclear how well such TCP mechanisms workin a multihop wireless network; this is the focus of this work.Multihop wireless networks have several characteristicsdifferent from wired networks. Firstly, in a typical wirelessnetwork that uses IEEE 802.11 MAC, packets may be droppeddue to either buffer overflow or link-layer contention causedby hidden terminals. Such losses directly affect TCP windowadaptation. Secondly, wireless channel is a scarce, shared re-source. Improving channel utilization through spatial channelreuse is highly desirable. Multiple nodes that do not interferewith each other should be encouraged to transmit concurrently.How well TCP utilizes the multihop wireless channel throughspatial reuse poses another important issue. A fundamentalproblem is that, TCP has to interact with ad hoc forwardingand IEEE 802.11 MAC in a multihop wireless network, whichexhibits features quite different from the wired or wirelesscellular networks. Earlier research on TCP performance overad hoc networks has been focused on the impact of mobility-induced factors, such as link breakage and routing failures [2],[3]. However, the interaction between TCP and the underlyingmultihop forwarding with the IEEE 802.11 MAC, is leftunaddressed.In this paper, we study the effect of multihop wireless linkon TCP throughput and loss behavior for several simple net-work configurations and flow patterns. Through both analysisand simulations, we reveal several interesting results. First,given a specific network topology and flow patterns, thereexists a TCP window size, say W∗, at which its throughputis highest through improved spatial channel reuse. Furtherincreasing the window size does not lead to further spatialchannel reuse, but results in increased link layer contention andperceived packet losses. Second, the standard TCP protocoldoes not operate around W∗, and typically grows its averagewindow much larger than W∗. Consequently, TCP experiencesthroughput decrease due to reduced spatial channel reuse. Weobserve 4% to 21% throughput reduction (from the highestthroughput) in our simulated scenarios.The suboptimal throughput of TCP can be explained by itsloss behavior over the multihop wireless channel. In a wirednetwork, all incoming packets are dropped if buffer overflowsat a bottleneck. It helps TCP to quickly reduce its windowsize to release congestion. Multihop wireless networks exhibitdifferent drop features. Unlike wired networks where bufferoverflow dominates packet losses, most packet drops experi-enced by TCP are due to link layer contention, incurred byhidden terminals. Buffer overflow induced packet loss is rare,and the contention-induced packet loss offers the first sign ofnetwork overload. Our analysis and simulations further showthat contention drops exhibit a load-sensitive feature: as theoffered TCP packets exceed W∗and increase further, link dropprobability becomes non-negligible and increases accordingly.After the offered TCP packets exceed another threshold¯W ,0-7803-7753-2/03/$17.00 (C) 2003 IEEE IEEE INFOCOM 2003AB DCGHIDATAACKDATAACKFEFig. 1. Spatial reuse and contention.An example of 8 hop chain. Optimal spatial reuse is achieved when nodes {A, D, G} and nodes {B, C, E} are scheduledfor transmission alternatively. Node D is the hidden terminal for transmission A→B.the link drop probability saturates and flattens out. It turns out,however, that the link-layer drop probability is not significantenough to stabilize the average TCP window size aroundW∗. It therefore leads to suboptimal TCP throughput. Somesimulation results are also validated with experiments.Our observations also shed light on how to improve TCPperformance over multihop wireless networks. In this paper,we propose two link layer techniques: a Link-RED algorithmto tune the wireless link’s drop probability, and an adaptivelink-layer pacing scheme to increase the spatial channel reuse.The goal is to let TCP operate in the contention avoidanceregion. These simple techniques lead to 5% to 30% throughputincrease compared with standard TCP.The rest of the paper is organized as follows. Backgroundinformation is provided in Section II. A thorough study of theTCP
View Full Document
Unlocking...