Research On Self-Similar Models Of Network Traffic And Parameter Estimations

The popularization and enhancement of new network applications and the demand of broad band service make network traffic increasing rapidly such as multimedia and VoIP etc, so that traditional traffic model character will no longer be adapted to current network traffic. The effective method for studying self-similar traffic is to build models that could describe network character more authentic, and be applied to simulation research. The self-similar traffic becomes a hotspot for researchers.In this paper, traditional traffic models, such as Markov model, Poisson model, are introduced in detail. Disadvantages of these models are analyzed.As one of the essential characteristics in the network traffic, self-similar phenomenon is analyzed in depth. Definition of self-similar is presented, and several properties of self-similar are discussed. Then, possible causes of self-similar are analyzed. Influence of self-similar on network performance is also discussed.Several self-similar network traffic models, such as ON/OFF model, FGN model, FARIMA model, Alpha stable model, are introduced in detail. Based on the simulations, the advantages and disadvantages of each model are listed. ON/OFF model is simple, but it is hard to describe complex situation. FGN model is self-similar, whereas it can't describe long-range dependence and short-range dependence in the same time. While FARIMA model can describe long-range dependence and short-range dependence in the same time, it's too complicated. Alpha stable model can show the burstiness and heavy-tailed distribution of network traffic. However, it has no close form.Major methods of Hurst parameter estimation of self-similar network traffic are introduced and compared. Variance-time plots method is studied in particular detail; both theoretical analysis and experiment suggest that estimation results depend largely on data blocks'range selection. Possible causes of this dependence are discussed, and an empirical formula for data blocks'range selection is proposed, which increases estimation precision.Finally, the main contributions in this dissertation are summarized and some suggestions and directions for the future work in this field are given.