Why is Internet Traffic Self-Similar (or Is It?)

In this talk I present current evidence and explanations for self-similarity in network traffic, with an overrepresentation of my work in this area. This talk is accessible to a general audience, but a knowledge of statistics would help.

The slides are available in gzipped Postscript and PDF.

Abstract

In the early 1990's researchers at Bellcore showed that Ethernet traffic, contrary to expectations, showed statistical characteristics of a self-similar, long-range dependent process. This observation has been repeated in other environments including World Wide Web traffic in the Internet.

Several explanations have been proposed for this phenomenon. Most are based on the assumption that the distribution of file sizes (or some other network metric) is long-tailed. There is some evidence for this assumption, but there is also evidence that these distributions are lognormal, and reasons to expect them to be.

In this talk I will review the evidence for self-similarity and long-tailed distributions and present models that might explain the origins of lognormal and long-tailed distributions.