Qualifying Examination, Neal Cardwell, November 1998
Recently researchers have proposed several analytic models of TCP performance. Several of these models accurately describe the steady-state behavior of long TCP connections transferring megabytes of data. Several address the performance of short TCP connections transferring a few kilobytes of data, but assume no packet loss. However, it is not obvious how applicable these models are, since a wealth of evidence suggests that in the most popular use of TCP today -- HTTP, the protocol of the World-Wide Web -- connections are short, often around 10KB, and often suffer high packet loss rates in the neighborhood of 5%.
This paper examines how well these existing models capture the behavior of short TCP connections under realistic loss rates. In addition, we describe two new models for TCP performance: a model for short TCP connections that experience no packet loss, and a generalization of this model for short connections that do experience packet loss. Comparing these models against simulated TCP flows, we find that both the new models and, somewhat surprisingly, the previously-proposed steady-state and lossless short flow models closely reflect TCP performance under a wide variety of simulated conditions. However, several experiments suggest that the new short flow models may provide a much closer fit for the measured performance of real TCP implementations in the Internet.
[ Paper: Postscript | Talk slides: Postscript ]