Queueing performance comparison of traffic models for internet traffic

There are a fast-growing number of studies in the literature that deal with capturing the statistical properties of today's data traffic. While this is a good development and a lot of progress has been made in terms of understanding the properties of this traffic, little progress has been made in regard to the question of how traffic characteristics like long range dependence/self-similarity or heavy tailed distributions of interarrival times impact the engineering of data networks. A number of traffic models have been shown to capture certain aspects of the statistical properties of measured data traffic well, but it was not studied how well those models can be used to predict queueing performance for engineering purposes. We compare in this paper the performance prediction of a wide range of traffic models for a number of performance measures that are common in network engineering. In particular we investigate how accurately traffic models like Poisson, 2-state MMPP AR(I), Weibull, Pareto, and FBM predict queueing performance like loss probabilities and queue lengths (delays) when compared to results from measured Internet traffic. The numerical results lead to the conclusion that traffic models that are based on heavy-tailed distributions like the Pareto distribution are needed when dealing with small buffer sizes. For larger buffer sizes, the effect of long range dependence that is present in the data traffic influences the queueing behavior considerably and therefore traffic models that exhibit long range dependence like FBM are needed. It is interesting to see how the long range dependent traffic models fail to predict the queueing performance for small buffer sizes. This indicates that the choice of the traffic model depends not only on the type of traffic but also on the considered application/device in terms of buffering.