This project was set up to investigate the statistical phenomenon known as Long-Range Dependence and its presence in internet traffic. The subject is of importance because traffic in which LRD is present is experiences significant extra delay and packet loss due to the extremely bursty nature of such traffic.
The specific aim of this project was to investigate the source of LRD in internet traffic -- that is to say, the mechanism by which it arises. Investigation of the (extensive) literature on the subject suggested four possible causes:
To expand on this, the first is the claim that some applications deliver traffic which inherrently contains LRD with no further interference from the network. This has particularly been noted in VBR video traffic. The second claim comes from the theoretical result (due to Taqqu) that an aggregation of heavy-tailed sources leads to a data-stream which is LRD. It has long been observed that quantities such as the length of files available for download on web-sites are LRD. The third claim arises from the widely-held belief that long-range dependence arises because of feedback in the TCP protocol. This has some experimental backing from results on Markov chain type models of TCP protocols and also from simulations of TCP window size parameters. The fourth claim arises from simple simulations of grid topology networks where poisson sources aggregate to LRD with simple assumptions about sources, sinks and routing.
While it is plausible (indeed likely) that all these mechanisms are true to some extent, it seems certain that some are more important than others. The work in this project suggested that the second cause was the most likely culprit and perhaps the fourth. For more information please consult the publications section.