Skip to Main Content
A novel predictor for network traffic, the wavelet domain recursive least-squares (WDRLS) predictor is discussed. Empirical studies have shown that network traffic possesses diverse statistical properties and exhibits a complex correlation structure characterized by short-range dependence (SRD) and long-range dependence (LRD). A challenge in predicting network traffic is how to exploit such a complex correlation structure with both high accuracy and computational efficiency. In the proposed WDRLS predictor, we use the wavelet transform to tackle these issues. Specifically, we predict the wavelet coefficients and solve the prediction of network traffic through a reverse wavelet transform. Our approach is based on the discovery that, although the network traffic has both SRD and LRD correlation structures, the corresponding wavelet coefficients are all SRD. We further assume that the SRD wavelet coefficients can be well approximated by a linear correlation structure for prediction. A least-squares method is adopted to make predictions in the wavelet domain. An important feature about the WDRLS predictor is that it can make an on-line prediction of network traffic. This is made possible by implementing the least-squares method in a recursive way. The performance of WDRLS is investigated with real network traffic. Simulation results demonstrate the feasibility of using the linear correlation structure to predict SRD wavelet coefficients and show that WDRLS can achieve high prediction accuracy when working with real LAN or WAN traffic.