Multiresolution decomposition and burstiness analysis of traffic traces

After a novel paper by W.E. Leland et al. published in 1994, self-similar properties of multimedia network traffic became a well known fact. More precisely, aggregate packet level network traffic exhibits fractal-like behavior over time scales on the order of a few hundreds of milliseconds and larger as long as the individual data streams that generate the aggregate traffic have a heavy tailed distribution. This finding has provided new insights into how self-similarity can have an impact on network performance and in what way the quality of service (QoS) parameters can be kept at an optimal level under self-similar traffic. The level of self-similarity is directly related to the degradation of the network performance and this level changes over time depending on the load on the network. In our previous study, we had investigated the methods to detect Hurst parameter (H), a measure of self-similarity, in terms of their reliability and complexity to be implemented on real time applications. Now, this paper is focused on the multiresolution decomposition technique. The optimum number of scales and the optimum number of the taps of a basis filter were addressed and the results have proved that different H is returned for different scale depth and the length of the filters deployed. Also, we, in this study, have demonstrated that H itself is not a good representative of the burstiness of traffic traces. It is shown that, in addition to H, also peak, mean, peak/mean ratio and the burst length should be considered to be able to characterize the subjected traffic