The influence of long-range dependence on traffic prediction

Predicting traffic requirements is a fundamental objective of network management algorithms, whether explicitly stated or not. Considering the abundant evidence that traffic is long-range dependent (LRD) and the fact that the history of long-range dependent processes has significant impact on the present value of the process, it is natural to assume that predicting LRD traffic would be rather rewarding. Although there is indeed significant performance gain in using the correlation structure as the Hurst parameter increases, we show in this paper that this is primarily due to the utilization of the specific short-term correlations that occur within the structure of the traffic, rather than storing a long history of the traffic and utilizing the long-term correlations that exist between the future traffic level and this stored history. This conclusion is demonstrated using two specific management functions: rate control for elastic connections, and queue management using probabilistic packet discarding