Heterogenous network traffic: Understanding and modeling

Heterogeneous network traffic possesses complex statistical properties. Those properties have been investigated from different aspects, resulting in different understandings. In this thesis, we address the problem of providing a unified understanding and modeling of the heterogeneous network traffic.; We use simulated/actual traffic traces and analytic approach to provide a unified view of the network traffic in the whole spectrum of link utilization and across multiple time scales. Using normalized network traffic, we show empirically and analytically how the marginal distributions and variance-mean relation of network traffic evolve with respect to time scales and link utilization. We demonstrate that in a open-loop bufferless scenario, the asymptotic behavior of temporal correlation tends to be unchanged for a wide range of link utilization. Based on the variances of in time and wavelet domain, we provide a unified view of network traffic.; In the second part of this thesis, we investigate the problem of modeling heterogeneous network traffic. Specifically, we study the independent wavelet models. First, we show analytically that the independent wavelet models are capable of matching any decay rate of auto-correlation functions asymptotically. Then, we show that the independent (Haar) wavelet models are inaccurate in capturing the small lags of ACF. We demonstrate that independent wavelet models, which employ orthogonal wavelets with higher vanishing moments, can improve the performance of ACF at small lags and still match the tail part of the ACF. Analytically, we show that the high-order vanishing moments provide smoother temporal correlation function of the wavelet function than that of the Haar wavelet, and thus improve the performance on modeling a smooth autocorrelation functions. Next, we prove that the independent wavelet models are mono-fractal. Therefore, they are applicable to traffic at large time scales only, where traffic is mono-fractal. Finally, we investigate the scalability issues of the IWM. We develop a scalable independent wavelet model, which uses fixed number of parameters to generate synthetic traffic with arbitrary length, link capacity and different sampling intervals.