Temporal clustering effects in the network traffic evaluated by queueing system performance
Viet Nguyen Duc, Araik Tamazian, Oleg A. Markelov, Mikhail I. Bogachev · 2016
We study the temporal clustering effects in the computer network traffic on the performance of a node with a given throughput limited by its outgoing channel capacity. The empirical data sets are exemplified by three different HTTP servers. We consider the inter-arrival times and the service times and evaluate the average system performance by the queuing system simulation. Our results indicate that the common M/M/1 model underestimates both the average queue size and the average sojourn time by one to two decades at 10-90% average system utilization. This leads to the 3-5 times underestimation of the required outgoing channel throughput to maintain the given average sojourn time. In contrast, by using an alternative model with Pareto-class distributions this underestimation can be reduced by more than one decade. The model is further improved once the long-range correlations in the inter-arrival times sequence are taken into account. Finally, we focus on the analysis of the long-range correlations in the total outgoing intraday traffic and discuss its potential implications on the occurrence of overloads and their reproducibility by the suggested model.