Crossover scaling effects in aggregated TCP traffic with congestion losses

Abstract

We critically examine the claims that TCP congestion control contributes to the observed self-similar traffic rate correlations. A simulation model is designed to analyze aggregated traffic of many TCP file transfers, with network topologies large enough so that each transfer has independent packet losses due to competition with other TCP traffic. To separate the effects of session-level variability from network-level variability we examine traffic consisting of small fixed-size files, and of heavy-tailed distribution of file sizes, with small variance of inter-session periods.We find that, with increasing packet loss rate, traffic rate scaling crosses over from the regime dominated by file size distribution to another scaling regime that is independent of file sizes. That loss-dominated scaling stretches over the timescales from RTT to the longest consecutive TCP time-outs (hundreds of seconds), and is not asymptotic. Analysis at the flow level exposes the mechanism of the crossover, from scaling dominated by variability of the flow ON-periods to that dominated by variability of the OFF-periods.However, it is unlikely that TCP timeouts contribute much to observed Internet traffic correlation structure, as they would matter only if widespread congestion losses exceeding 10% dominated the typical behavior of the Internet.