TCP with variance control for multihop IEEE 802.11 wireless networks

• Authors: Jiwei Chen;Mario Gerla;Yeng Zhong Lee;Medy Sanadidi
• Venue: MILCOM'06 Proceedings of the 2006 IEEE conference on Military communications
• Year: 2006

In this paper we argue that a cause of poor performance of TCP NewReno in wireless ad hoc networks is the use of packet drop as feedback for window adjustment. While packets can be dropped because of wireless congestion/interference, it is true however that these losses are typically recovered by link level retransmissions before they affect TCP. Eventually, the link level gives up after a maximum number of retransmissions and TCP detects a loss. But, at this time, congestion in the wireless medium has already built up. Thus, this feedback "comes too late". To overcome the problem, we propose a more reactive feedback mechanism, driven by the variance in achieved rate rather than by packet drop. This choice is motivated by the observation that MAC retransmissions lead to fluctuations in achieved rate and increased rate variance. Most important, the change in variance can be detected by the TCP source well before packet loss. The result is TCP-VAR (TCP with Variance Control), a novel congestion control algorithm for TCP. TCP-VAR implements rate-based scheduling where the source adapts the outgoing packet rate based on the feedback on achieved rate variance. The latter is measured from the rate of returning ACKs. Simulation studies confirm that TCP-VAR reacts more promptly to wireless congestion than TCP NewReno, with substantial improvements in efficiency while maintaining fairness. An important implementation advantage of TCPVAR is the fact that it requires "source only" modifications of TCP, without any dependence on low layer support.