Research On Simulation And Performance Evaluation Of Self-Similar Traffic

The discovery of the self-similar characteristic of network traffic has great influence on network traffic modeling, performance evaluation and network control. Traditional Poisson-based models of network traffic are based on the hypothesis of Markov which has the nature of short-range dependence (SRD). Recent traffic analysis from various packet networks shows that network traffic processes exhibit ubiquitous properties of self-similarity and long range dependence (LRD), i.e. of correlation over a wide range of time scales. LRD exists on multiple time scales and has great influences on network performances such as delay, jitter, cell loss rate and throughput on the large time scale. The traditional Poisson-based models neglect the important characteristic of self-similariy. They can not capture the actual characteristic of network traffic accurately.In this thesis, the author studies the problems of network simulation and performance evaluation of self-similar traffic with depth. Firstly, several mathematical definitions of self-similarity are given. Some mathematical and physical features describing the self-similar processes are described. The methods of modeling and generation of the self-similar traffic are researched and discussed. The performance of these models is analyzed. The influence of network performance on self-similariy is studied through simulation.Based on the conclusion of the concepts of self-similarity, the implementation of the self-similar models, such as ON/OFF, Fractional Gaussian Noise (FGN), Fractional Brownian Motion (FBM), Fractional Auto-Regressive Integrated Moving Average (FARIMA), are introduced. The precision of self-similarity traffic series generated by these models is analyzed. The relults reveal that although the Hurst coefficient of traffic series generated by ON/OFF model is close to the expected value, its Hurst coefficient does not remain stable and changes according to the length of the series. Compared to other models, FGN model is more precise and stable. FARIMA model is consisting with both long and short range dependent structure and suit to the actual fractional structure. So it is much better than other modelsSecondly, relative models, parameters and indexes of network performance evaluation are analyzed. The author focuses on the queuing performance of FBM and FARIMA. The simulation model and the source of simulation are discussed. The performance of self-similar traffic is analyzed and summed up through extensive simulation.Queuing performance is a key index of network performances analysis. In this thesis, based on the OPNET simulation which is driven by the traces generated with self-similar traffic model described above, the performances and its influence factors of G/M/1 queuing model are analyzed. The author focuses on the study of influence factors on network performances, such as queuing delay, queue length and cell loss rate. The research results demonstrate that the queuing performance is determined by the distribution of packet interval time. The queuing performance of ON/OFF traffic model is related to Pareto distribution of ON or OFF period. Self-similar traffic results in much worse queuing performance than short range dependence traffic, and the variation of the performance of self-similar traffic is more violent. Variance of the self-similar traffic has a great affect on the queuing performance.