Measurements on a high-performance switched Ethernet system are presented
that reveal new insights into the statistical nature of le server and web server
traÆc. Both le sizes and data requested from the web server are shown to match
well a truncated Cauchy distribution. This is a distribution with heavy tails similar
in nature to the commonly used Pareto distribution but with a much better t
over smaller le/request sizes. We observe self similar characteristics in the traÆc
at both servers and also at a CPU server elsewhere on the network. TraÆc from
this server is predominantly targeted at the le and web servers, suggesting that
self-similar properties at one point on a network are being propagated to other
points. A simple simulation model of an isolated server is presented with Poisson
arrivals and service (packet transmission) demands with the same Cauchy distribution
as we observed. The departure process is shown to follow a power law
and the corresponding power spectrum is shown to match extremely well that
of the observed traÆc. This supports the suggested link between le/request
size distribution and self-similarity. The resulting implication that self similarity
and heavy tails are primarily due to server-nodes, rather than being inherent
in o ered traÆc, leads to the possibility of using conventional queueing network
models of performance. This idea is further supported by an additional simulation
experiment and suitable models are proposed.