Delay-based congestion avoidance for TCP

Abstract

The set of TCP congestion control algorithms associated with TCP/Reno (e.g., slow-start and congestion avoidance) have been crucial to ensuring the stability of the Internet. Algorithms such as TCP/NewReno (which has been deployed) and TCP/Vegas (which has not been deployed) represent incrementally deployable enhancements to TCP as they have been shown to improve a TCP connection's throughput without degrading performance to competing flows. Our research focuses on delay-based congestion avoidance algorithms (DCA), like TCP/Vegas, which attempt to utilize the congestion information contained in packet round-trip time (RTT) samples. Through measurement and simulation, we show evidence suggesting that a single deployment of DCA (i.e., a TCP connection enhanced with a DCA algorithm) is not a viable enhancement to TCP over high-speed paths. We define several performance metrics that quantify the level of correlation between packet loss and RTT. Based on our measurement analysis we find that although there is useful congestion information contained within RTT samples, the level of correlation between an increase in RTT and packet loss is not strong enough to allow a TCP/Sender to reliably improve throughput. While DCA is able to reduce the packet loss rate experienced by a connection, in its attempts to avoid packet loss, the algorithm will react unnecessarily to RTT variation that is not associated with packet loss. The result is degraded throughput as compared to a similar flow that does not support DCA.