Enhanced metric caching for short TCP flows

Abstract

Internet-based applications that require low latency are becoming more common. Such applications typically generate traffic consisting of short, or bursty, TCP flows. As TCP, instead, is designed to optimize the throughput of long bulk flows there is an apparent mismatch. To overcome this, a lot of research has recently focused on optimizing TCP for short flows as well. In this paper, we identify a performance problem for short flows caused by the metric caching conducted by the TCP control block interdependence mechanisms. Using this metric caching, a single packet loss can potentially ruin the performance for all future flows to the same destination by making them start in congestion avoidance instead of slow-start. To solve this, we propose an enhanced selective caching mechanism for short flows. To illustrate the usefulness of our approach, we implement it in both Linux and FreeBSD and experimentally evaluate it in a real test-bed. The experiments show that the selective caching approach is able to reduce the average transmission time of short flows by up to 40%.

A measurement-based study of MultiPath TCP performance over wireless networks

“…When losses occur, the ssthresh value will be reset and cached for initialization of future TCP connections to the same destination. However, this is shown to be harmful for short flows [13] if an earlier connection to a particular destination encounters a sequence of losses. This is because ssthresh will be set to a small value and all the subsequent newly open flows to that destination will have the same small ssthresh.…”