A superstatistical approach to the modeling of aggregate network traffic with long-term correlations

We study the long-term correlation effects on the performance of a network node with limited throughput. The empirical data are the traffic records of a busy cluster hosting the 1998 FIFA World Cup website. The traffic data is being split into individual daily fragments exhibiting in some cases stationary, but otherwise non-stationary traffic dynamics. Using queuing system simulations, we find that the conventional approximation for the G/G/1 queuing system given by the Kingman's formula underestimates the average sojourn time by one to two decades under high system utilization. In contrast, by using alternative model with long-term correlations in the local access patterns this underestimation can be reduced by more than one decade.