Research On Self-similarity Of Network Traffic

Through the measurement and analysis of LAN and WAN, Leland and Paxson found that real traffic has statistical self-similarity. Traditional network traffic models neglect the important characteristic, when they describe the actual traffic. Compared with traditional models, self-similar models are more suitable for describing the real characteristics of traffic. Self-similarity is very important to system performance such as packet loss rate, network delay, throughput and so on, so research on self-similarity is very necessary.In this paper, mathematical definition and characteristics of self-similarity are given, and the estimation methods of Hurst parameter are introduced. Then several Self-similar traffic generation methods are summarized, and their advantages and defects are compared. In addition, the causes of self-similar traffic are discussed.Aggregation and decomposition of self-similar process are studied, certification of several related theorems is given, and the theorems are verified through examples. In the research on the impact of self-similarity to network traffic, emulation based on NS, the impact of self-similarity on packet loss rate is analyzed. The result shows that with the increasing of TCP load and the degree of self-similarity, the packet loss rate grows on. Moreover, in the emulation experiment, adding a loss module to study the impact of packet loss rate on self-similarity. The result of emulation experiment shows that with the enhancement of packet loss rate, the degree of self-similarity is increasing accordingly.Emulation experiments mentioned above show that self-similarity can cause degression of network performance, for example, the increment of packet loss rate. When it is severe, congestion will be caused. So a congestion alarm system is proposed in this paper, combined with dynamic bandwidth allocation. The system utilizes the correlation of self-similar traffic over long period of time, predicts the traffic over next period of time. When the buffer can't handle the predicted traffic load, the system will be startup, bandwidth is increased, so network performance is improved. The emulation experimental result shows that the system can predict forthcoming congestion. Moreover, compared with fixed bandwidth allocation, it can save bandwidth.