Measurement and modelling of self-similar traffic in computer networks

Abstract

Results are noted from the analysis of traffic measured over a departmental switched Ethernet. Self-similar characteristics are seen throughout the network, for example, at the compute servers, web server and intermediate routers. It is shown that data shipped by the web server (i.e. including both static files from a file server and dynamically-generated data) have a heavy-tailed distribution, which is matched extremely well by a Cauchy distribution. It is also shown that the fragmentation of the data (i.e. into Ethernet frames) leads to a departure process whose power spectrum is shown to follow a power law very similar to that of the observed traffic. Importantly, the power law appears to be largely independent of the input process; self-similar behaviour is observed even with Poisson arrivals. This supports the suggested link between file/request size distribution and self-similarity in network traffic. The resulting implication that self-similarity and heavy tails are primarily due to server nodes, rather than being inherent in offered traffic, leads to the possibility of using conventional queueing network models of performance.