Study On A New Model For Network Traffic Matrix Estimation

Internet technology is one of the fastest growing technologies in twenty-first Century. It has been widely used in our life, and made great contribution to social and economic development. However, with the development of Internet technology, many new network applications have emerged in recent years. Though they bring people much convenience and entertainment, they add great burden to the network providers. Meanwhile, different access ways provided by a large number of heterogeneous networks make it hard to manage Internet. Some important parameters used in network management can be obtained by network measurement. Therefore, efficient network measure is of great importance.The traffic matrix is one of the crucial inputs in many network planning and traffic engineering tasks. It reflects the traffic demand between Original-Destination (OD) pairs in the network. Network Tomography is often used in traffic matrix estimation, which estimates the path-level parameters from the link-level measurement data. But the nature of the Network Tomography determines that the traffic matrix estimation based on it is an ill-posed linear inverse problem. So some side information is needed such as prior models to solve such a problem. Lots of models and methods are proposed to estimate the network overall traffic matrix from link measurements. However because of the limits of the link measurements, the estimation on overall traffic matrix from link measurements based on these prior model assumptions do not perform well for large-scale networks.The theory of compressed sensing shows that with a small number of non-adaptive, randomized linear projection samples, any sufficiently compressible signal can be recovered. According to theory of compressed sensing, a probability model is proposed. It has been proved the probability model can reconstruct the traffic matrix from limited link measurements with small bias.The paper demonstrates the efficiency of the probability model and compares the model with the classical gravity model using real data of the Abilene network. It proves the probability model is applicable in real networks. When the probability model is applied in real networks, how to determine proper probability parameters which are suitable for a network is the critical problem. Since the gravity model can provide the probability parameters required in the probability model, we propose a model that combines the probability model and the gravity model, called the gravity-probability model. Finally it is proved the performance of TM estimation based on this model is better than that based on two sole models separately.