Prediction and correction of traffic matrix in an IP backbone network

The prediction of traffic matrices (TM) is critical for many IP network management tasks. With the recent development of high-speed traffic measurement technologies, complete TM could be collected from operational IP networks. In this paper, we report our efforts in predicting the TM measured from a real IP backbone network in China. The new problem here is how to deal with the rich but noisy TM data and predict various traffics, ranging from the original-destination (OD) flow traffic, the node traffic to the total network traffic. After examining the traffic characteristics, we choose the node traffic as the principal data for prediction, and propose three prediction and correction methods: Independent Node Prediction (INP), Total Matrix Prediction with Key Element Correction (TMP-KEC) and Principle Component Prediction with Fluctuation Component Correction (PCP-FCC). TMP-KEC and PCP-FCC are designed with different purposes, i.e., for smaller prediction errors of the total network and the OD flows, respectively. The results show that, INP performs worst; TMP-KEC efficiently reduces the prediction errors of the large matrix elements; while, PCP-FCC achieves smaller average prediction errors for the elements as well as the complete matrix.