Self-similar Traffic Forecast

The recent researches of network traffic have shown that many kinds of real network traffic have both self-similarity and long-range dependence. It is the long-range dependence that long-term prediction of network traffic is possible.In this essay, the characteristic of network traffic is narrated, and the definition of self-similarity and some useful theorems are given. Moreover, special aspects of self-similar traffic are summarized.For long-range dependent traffic, two prediction models are given and discussed The prediction results can be applied to reduce loss ratio in allocation of memories in network nodes.The first model is FARIMA ( Fractional AutoRegressive Integrated Moving Average). It is capable of capturing both the long-range and short-range behavior of network traffic. The research of FARIMA model is emphasized due to its linearity and easier access than nonlinear ones. Firstly, in this paper, the definition of FARIMA model is introduced, and the generating method of FARIMA process is given. Simulations have shown that FARIMA process is self-similar. Secondly, procedures are given to obtain the three parameters of FARIMA( p, d, q) in self-similar network traffic. Finally, the prediction method of network traffic is proposed. Simulations show that FARIMA( p, d, q) is effective for long-range dependent network traffic prediction. In order to minimize computing time, simplified prediction procedures are provided, which transform parameter seeking into ARMA process, and computing complexity and duration are decreased.The second model is BP of neural network. In this paper, the characteristic of BP is summarized. In long-range dependent case, a 5-layer BP is applied to predict the network traffic. Simulations show that, in terms of prediction, BP is more precise than FARIMA, but at the cost of computing complexity.The set-up of prediction model and its application in real network traffic are very important in the researches of bandwidth allocation, discarding methods, and loss ratio reducing. When prediction results are applied in memory allocation of network nodes, simulation results indicate that loss ratio can be reduced.