Self-Similar Network Traffic And Performance Of High-Speed Routing Structure

Measuring and modeling of network traffic are the fundament of network architecture design and performance analysis, the accuracy of measurement and the fitness of model are very important for an operational network to achieve the optimization of system performance.For a long time, network traffic has been only regarded as a Markov chain, and mainly modeled by Poisson process. These kinds of models have a short-range dependence structure and play an important role in network analysis, especially in traditional telecommunication network that was mainly designed to provide voice service. However it was found that besides short-range dependence, self-similarity (i.e. long-rang dependence) exists in actual measured network traffic. This character has a great influence on network performance and control scheme at large time scales, such as delay, jitter, congestion control, and so on, which is much different from Markov process. So self-similar network traffic and its performance evaluation attract much attention and are studied world widely.In this thesis, the self-similar network traffic and its performance evaluation are studied extensively. Firstly, we discuss the second-order parameter estimation and the prediction of self-similar network traffic. Then the impact of self-similar network traffic on queueing performance is discussed at different points of view with different methods and some useful results are obtained. The main research results are presented as follows.We know that self-similar process is mainly different from Markov process with its long-range dependence (i.e. second-order self-similarity) at multiple time scales, so it is very important to estimate its second-order parameter in practical traffic model and performance evaluation. With fractional Gauss process, we analysis the estimation of autocorrelation and obtain an evaluating formula, in which the relation of the accuracy of estimated autocorrelation, sampled data length and Hurst parameter is accurately described. Experiments confirm that our result is not only correct for fractional Gauss process, but also a perfect reference to other self-similar process.Meanwhile, a sharply variety in the accuracy of estimated autocorrelation we called "jumping burstiness" is discovered around H=0.75. It is a very interesting phenomenon that hadn't been reported before. This "jumping burstiness" shows that there is a non-linear relationship between Hurst parameter and traffic burstiness, which can be a good instruction for the utilization of second-order self-similar character. It also shows that all the traffic varieties cannot be captured only with Hurst parameter. Afterward, a traffic predicting algorithm based on FARIMA (Fractional Auto- Regression Integrated Moving Average) model is proposed, which estimates the self-similar parameters explicitly. This algorithm considers the character of long-range dependence of network traffic. Simulation result confirms the convergence of estimated parameter, which guarantees the predicting accuracy of this algorithm.The effect of self-similar network traffic on queueing system relates directly to the whole network performance. Packet delay and jitter in a FCFS (First Come First Service) queueing system with fractional Brown motion input are discussed. Then, the upper bound of buffer overflow probability is obtained with this model. An effective bandwidth allocation and admission control scheme is also proposed. Furthermore, the onset time scale that self-similar traffic begins to affect system performance is discussed. It is found that the onset time scale is not only related to short-range dependence, but also related to the system parameters, such as buffer, bandwidth and so on. This result remedies the defect of fractional Brown motion that cannot be operated at little time scale and it is a good reference to the time scale related problems.Based on the extensive analysis of the inputting IP packet size and inter-arrival time, a mixture Pareto traffic model is proposed from the point of v...