On Forecast And Simulation Of Large-Scale Network Traffic Based On S-Transform And Compressive Sensing

For the increasingly developed and expanded Internet, effective network management is an essential guaranty for maintaining networks' normal operations. As a key input parameter of numerous traffic engineering such as network management and network maintenance, traffic matrix has become the center of network research and traffic analysis. At the same time, it is becoming even more difficult to obtain traffic matrix because of networks' rapid development. Allowing for the huge consumption on network resources brought about by direct measurement, we actually seldom gains traffic matrix in such a way, indirect prediction has then become the major method of deriving traffic matrix, which has attracted extensive attention.At present, the biggest challenge of network traffic prediction is its highly ill-posed nature inherent in the prediction model. How to predict traffic matrix accurately depends on whether we can conquer the problem of ill-posed nature or not. Based on the ill-posed problem as well as various traffic features such as self-similarity, multifractality and space-time characteristic and so on, the thesis proposes three algorithms to predict network traffic based on S-transform and compressive sensing. As a powerful time-frequency analysis tool, S-transform can transform the signals from time domain to time-frequency domain and contain their whole frequency information. After performing S-transform on network traffic, we found out that, different from those complicated features in time domain, the transform results of low-frequency are larger than those of high-frequency, namely, network traffic display obvious sparsity in time-frequency domain. In light of that, the thesis bring forward three network prediction algorithms based on one-dimensional General S-transform and compressive sensing, two-dimensional General S-transform and compressive sensing, two dimensional orthogonal S-transform and compressive sensing respectively.The thesis firstly utilizes one-dimensional generalized S-transform to conduct time-frequency transform on network traffic, then exploits different prediction model to predict the high-frequency part and low-frequency part respectively. While since one-dimensional generalized S-transform can only independently analyze one single Origin-Destination (OD) flow, its ability to thoroughly describe the connection between different OD flows, namely the space characteristic, needs to be improved. Therefore, the thesis further proposes to use two-dimensional generalized S-transform to simultaneously analyze multiple OD flows, and then divide them into stationary and fluctuated parts to do the prediction respectively. Final, the thesis come up with the prediction algorithm based on two dimensional orthogonal S-transform and compressive sensing to overcome the problem of high computation and demanding memory requirements. Similar to the second algorithm, the algorithm also predicts the stationary and fluctuated parts of network traffic, respectively, which can not only get accurate traffic estimation, but also reduce the computation and redundancy.In a word, the three algorithms proposed by the thesis transform the problem of network traffic prediction from time domain to time-frequency domain and successfully conquer the highly ill-posed problem existed in time domain, which has developed a new approach for traffic analysis and prediction.