Abstract

This talk looks at the performance of TCP measurements “in the wild” as revealed by analysis of passive traces. It is shown that the old-fashioned model of TCP as being a protocol which “fills a pipe” (maximises throughput) controlled by loss is not the best way to understand how TCP really works in networks. The first section of the talk looks at CAIDA and MAWI traces and constructs a statistical model which relates TCP throughput with delay, session length and loss using statistical models. The second section of the talk digs in detail into the mechanisms which control TCP behaviour by analysis of unanonymised traces from MAWI (Japan). Traditionally TCP is thought of in terms of a sender pushing data to a receiver at a rate governed by a window size which is modulated by loss. Trace analysis shows that for the data set considered less than half of the sent data is controlled by this mechanism.

Bio: Richard Clegg is a Senior Research Fellow in Electronic Engineering at University College London. His PhD “The Statistics of Dynamic Networks” was awarded by the University of York in 2005. His current research interests include Software Defined Networking, traffic statistics, overlay networks and network topologies.