Experimental Evaluation of the TCP Simultaneous-Send Problem in 802.11 Wireless Local Area Networks

Authors: Sumathi Gopal Dipankar Raychaudhuri

Complete Citation

Gopal, S. and Raychaudhuri, D. 2005. Experimental evaluation of the TCP simultaneous-send problem in 802.11 wireless local area networks. In Proceeding of the 2005 ACM SIGCOMM Workshop on Experimental Approaches To Wireless Network Design and Analysis (Philadelphia, Pennsylvania, USA, August 22 - 22, 2005). E-WIND '05. ACM Press, New York, NY, 23-28. DOI= http://doi.acm.org/10.1145/1080148.1080153

Abstract

This paper is an experimental follow up to our earlier paper [1] that investigated the TCP simultaneous-send problem which arises in infrastructure mode 802.11 wireless local area networks. In particular it was observed that for file transfer traffic, 802.11 wireless nodes have a sustained supply of packets to send and hence experience a relatively high rate of MAC contention. We showed that for TCP, this resulted in competition among data and ACK packets for channel access which caused considerable deterioration in flow throughput. Simulations of TCP ACK skipping as an alleviation to the problem, showed improvements as high as 100% when MAC retries were disabled. There were gains in other scenarios too albeit more moderate. We evaluate the same TCP simultaneous-send problem with real world experiments on a wireless-cum-wired network testbed called ORBIT [2] at WINLAB, Rutgers University. ORBIT makes it feasible to conduct controlled and reproducible experiments in a wireless network scenario. The same network setup scenarios evaluated in simulations were considered here., particularly – scenarios with and without MAC retries, multiple TCP flows and multiple skipped ACKs. However not all scenarios could be reproduced in experiments for logistical reasons. In all, the experimental results confirm the original hypothesis on the detrimental effects of simultaneous-send and corroborate the advantages of ACK skipping, However the percentage gains in TCP throughput are far more moderate as compared to those observed in NS simulations. A reason could be differing TCP implementations, particularly with not all TCP optimizations implemented in NS. We share the experiences and challenges faced, particularly given that this work is among the first of its kind for testbed evaluation of transport protocols over wireless networks.

Annotations

The authors note a few important items:

• NS2 simulation results = Experimental Results
  • NS2 showed over 100% performance improvement when disabling MAC retries and skip ACKs
  • NS2 has default queue size of 50 packets, can cause queue overflow in TCP slow start, resulting in worse overall performance.
• Authors tried Two variations of the experiment
  • With MAC Layer Retries
    • Skipping 1 ACK always allowed for better performance.
    • Skipping 2 or 3 ACKs allowed for better performance most of the time.
  • Without MAC Layer Retries
    • Skipping 1 ACK allowed for better performance
    • Skipping 2 or 3 ACKs reduced performance
    • TCP ACK handles wireless link congestion better then MAC Layer retries.

-- DavidSalyers - 08 Aug 2007