Delay-based early congestion detection and adaptation in TCP: impact on web performance

Abstract

Concerns over the scalability of TCP's end-to-end approach to congestion control and its AIMD congestion adaptation have led to proposals for router-based congestion control, specifically, active queue management (AQM). In this paper we present an end-to-end alternative to AQM—a new congestion detection and reaction mechanism for TCP based on measurements of one-way transit times of TCP segments within a TCP connection. Our design, called Sync-TCP, places timestamps in TCP headers, measures variation in one-way transit times, and uses these measurements as a form of early congestion notification. We demonstrate empirically that: (1) Sync-TCP provides better throughput and HTTP response-time performance than TCP Reno, (2) Sync-TCP provides better early congestion detection and reaction than the Adaptive Random Early Detection with Explicit Congestion Notification AQM mechanism, (3) Sync-TCP's congestion detection and adaptation mechanisms are robust against clock drift, (4) Sync-TCP is an incrementally deployable protocol—Sync-TCP connections can co-exist with TCP Reno connections in a network, and (5) the performance of TCP Reno connections are improved with the addition of even a small percentage of Sync-TCP connections.