Research On Self-Similarity Of Network Traffic And Its Impaction On Queueing Performance

Large of monitoring result have shown the self-similar nature of network traffic. The characteristic of traffic has great effects on the analysis, design, control, and performance evaluation of computer networks. In this thesis, several mathematical definitions of self-similarity are given, some of their features are described and the methods of multi-fractal analysis are discussed. Research in recent years has presented that multi-fractal model is more precise than fractal model to discribe the characteristic of small time scale of real networks.In this thesis, the principle and implementation of seven estimate algorithms about Hurst coefficient are described. Three factors i.e. variance, periodic signal, and fractal structure, which affect the performance of the estimate algorithms, are discussed. By rearranging the FGN (Fractional Ganssian Noise ), the author constructs a new serial of correlation structure in a specific scale range serial. By changing the scale range continuously and estimating the new serial, the author concludes the estimation of each algorithm depends on the fractal structure in a specific scale range, but the structures out of the scale range have no effects on the estimation. The measurement of real network traffic shows that the differences of fractal structures result in the differences of the algorithms estimations.ON/OFF FGN RMD( Random Midpoint Displacement ), FARIMA are four usual self-similar traffic models. The thesis introduces the implementation of the models and analyses the precision of self-similarity traffic generated by the four models. Although the Hurst coefficient of sequence generated by ON/OFF model is close to the range of expected value, the Hurst value does not remain stable and changes according to the length of the sequence. Compared to other models, Durbin FGN model is more precise and stable.In network performance's analysis, queuing performance is a key point. The authors feeds different network traffic to the G/M/1 model to modulate and discussthe factors affecting delay and length of the G/M/1 queue. The research shows that the queuing performance is determined by packet interval time;self-similar sequence generates much worse queuing performance than short dependence;sequence even then the variation is more violent.Variance has an important affect on the queuing performance.Finally,the research shows that sequence generated by ON/OFF model is related to the corresponding Pareto distribution in ON and OFF period.