IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 13, NO. 1, FEBRUARY
2005.
We consider a modification of TCP congestion control in which the congestion window is adapted to explicit bottleneck rate feedback; we call this RATCP (Rate Adaptive TCP). Our goal in this paper is to study and compare the performance of RATCP and TCP in various network scenarios with a view to understanding the possibilities and limits of providing better feedback to TCP than just implicit feedback via packet loss. To understand the dynamics of rate feedback and window control, we develop and analyze a model for a long-lived RATCP (and TCP) session that gets a time-varying rate on a bottleneck link. We also conduct experiments on a Linux based test-bed to study issues such as faiess, random losses, and randomly arriving short file transfers.We find that the analysis matches well with the results from the test-bed. For large file transfers, under low background load, ideal fair rate feedback improves the performance of TCP by 15%–20%. For small randomly arriving file transfers, though RATCP performs only slightly better than TCP it reduces losses and variability of throughputs across sessions. RATCP distinguishes between congestion and corruption losses, and ensures faiess for sessions with different round trip times sharing the bottleneck link. We believe that rate feedback mechanisms can be implemented using distributed flow control and recently proposed REM in which case, ECN bit itself can be used to provide the rate feedback.
We consider a modification of TCP congestion control in which the congestion window is adapted to explicit bottleneck rate feedback; we call this RATCP (Rate Adaptive TCP). Our goal in this paper is to study and compare the performance of RATCP and TCP in various network scenarios with a view to understanding the possibilities and limits of providing better feedback to TCP than just implicit feedback via packet loss. To understand the dynamics of rate feedback and window control, we develop and analyze a model for a long-lived RATCP (and TCP) session that gets a time-varying rate on a bottleneck link. We also conduct experiments on a Linux based test-bed to study issues such as faiess, random losses, and randomly arriving short file transfers.We find that the analysis matches well with the results from the test-bed. For large file transfers, under low background load, ideal fair rate feedback improves the performance of TCP by 15%–20%. For small randomly arriving file transfers, though RATCP performs only slightly better than TCP it reduces losses and variability of throughputs across sessions. RATCP distinguishes between congestion and corruption losses, and ensures faiess for sessions with different round trip times sharing the bottleneck link. We believe that rate feedback mechanisms can be implemented using distributed flow control and recently proposed REM in which case, ECN bit itself can be used to provide the rate feedback.