Transport protocol performance engineering

Transmission control protocols carry over 85% of the traffic in today's Internet. It is well known that currently deployed protocols (Reno, Reno with SACK and NewReno) do not scale well to paths with a high product of bandwidth times delay. Three recently proposed transport protocols that do not require router modification appear promising: BIC TCP, FAST TCP and TCP Madison. We evaluate these protocols on a wide range of Internet paths to get an in-depth understanding of which classes of paths they obtain bandwidth share on, as well as the specific causes of poor performance on other classes of paths. Internet path characteristics may change over time. Hence, a significant challenge to protocol performance evaluation is the absence of previous techniques and tools for rapid, accurate, and low-impact measurement of the path characteristics that affect transport protocol performance. TCP Madison also requires rapid and accurate host estimates of path capacity and minimum RTT at the start of a data transfer. Another significant challenge is that protocol comparisons require a platform that allows us to quickly switch between transport protocols without requiring time-consuming reboot. In this thesis, we develop new techniques and tools to quickly (within a few round-trips) and accurately estimate bottleneck link capacity, available bandwidth and bottleneck link buffer size. We also engineer the average round-trip time (RTTavg) used by TCP Madison throughout the data transfer, so that this measure is responsive to actual changes in cross-traffic while also being relatively insensitive to statistical fluctuations in bottleneck queue size during periods of stable cross-traffic. We develop a new user-level platform that achieves kernel level performance, enables seamless switching among transport protocols, and incorporates the new measurement tools. We use this new platform to perform an extensive protocol evaluation and analysis. In particular, for any given Internet path identified by available bandwidth, average and minimum RTT, packet loss rate, and tight link buffer size, we detail the throughput of each of the three recent protocols and TCP Reno. For paths where a protocol has significantly lower throughput than its bandwidth share, we provide the precise causes of the underperformance.