Research On Self-similarity Of Network Traffic Based On Classified Packets Of Link Layer In Metro Backbone Network

With the increasement of network bandwith and transfer rate, the development of various network applications is speeding up, which leads to the deepen study on self-similarity of network traffic. The research on self-similarity of network traffic has important significance for queuing performance, network measurement, routing policy, network protocol and network performance analysis.However, the current study lacks for specific measurement standards, which leads to the inaccurate calculation of Hurst exponent and the uncertainty of self-similarity of network traffic. On the other hand, the practical application of self-similarity becomes more difficult because measurable indicator of self-similarity is too single and research methods are too cumbersome. Therefore, a novel method of self-similarity of network traffic is proposed, which based on link layer classified packets in metro backbone network.This method can accomplish more fine-grained inspection of self-similarity of network traffic based on classified packets.In this thesis, the best sequence length of network traffic is introduced to resolve the uncertainty of self-similarity of network traffic and a single Hurst exponent is broken down into many Hurst exponents which indicate the self-similarity of different types of classified packets. Moreover, to change the abstract study situation of self-similarity with a single Hurst exponent, two specific analysis methods of network traffic characteristics are proposed based on the self-similarity of classified packets, which can be applied quickly and easily into differentiated services of network traffic, anomaly of network traffic and analysis of network protocol. The main contributions of this thesis are as follows:(1)The strong self-similarity of network traffic of link layer in the metro backbone is confirmed in the larger time scales. (Hurst exponent is greater than 0.7)(2)The best sequence length of network traffic is calculated, which improves the accuracy of Hurst exponent and resolves the uncertainty of self-similarity of network traffic. Furthermore, an real-time algorithm for calculating the Hurst exponent is proposed.(3)The self-similarity of classified packets is the same as the self-similarity of network traffic, which is proved by theory and experiment. Two specific methods of dynamic trend analysis and variance analysis are combined with the self-similarity of classified packets to analyze the characteristics of network traffic. Eventually, an example of the application of network protocol analysis is accomplished by using above methods.