Improving Restart of Idle TCP Connections

TLDR: This paper proposes a third alternative: pacing some packets at a certain rate until the ACK clock can be restarted, and describes the motivation and implementation and simulation results which show that it achieves the elapsed-time performance comparable to NSSR and loss behavior of SSR.

Abstract: TCP congestion avoidance mechanisms are based on adjustments to the congestion-window size, triggered by the ACK clock. These mechanisms are not well matched to large but intermittent bursts of traffic, such as responses from a HTTP/1.1-based web server. Idle periods between bursts (web page replies) stop the ACK clock and hence disrupt even data flow. When restarting data flow after an idle period, current implementations either enforce slow start (SSR) or use the prior congestion window (NSSR). The former approach, while conservative, leads to low effective throughput in cases like P-HTTP. The latter case optimistically sends a large burst of back-to-back packets, risking router buffer overflow and subsequent packet loss. This paper proposes a third alternative: pacing some packets at a certain rate until the ACK clock can be restarted. We describe the motivation and implementation of this third alternative and present simulation results which show that it achieves the elapsed-time performance comparable to NSSR and loss behavior of SSR.