| A point-to-point traffic matrix gives the volume of traffic between Origin-Destination (OD) pairs in an Internet Protocol (IP) network. This knowledge helps Internet Service Providers (ISPs) in performing various traffic engineering tasks such as routing protocols configuration, network planning and business planning. Measuring this traffic matrix directly is difficult and costly. However it is believed that it might be possible to obtain a reasonable estimate of this matrix from link loads. Hence, there is an on-going research effort to develop efficient and effective methods for inferring the traffic matrix from other readily available data such as link load measurements, routing and configuration data. There are different techniques existing in the literature for estimating traffic matrix from such easily measured data.;In this thesis, we selected three techniques which are known to perform very well most of the time, namely Tomogravity (TM), Entropy Maximization (EM) and Linear Programming (LP), and carried out a detailed comparative study among them. We found that for small networks Linear Programming performs better than Tomogravity and Entropy Maximization. On the other hand, for large networks, Tomogravity and Entropy Maximization perform equally well and much better than Linear Programming. Based on the comparative study, we proposed new directions for improving how to estimate traffic matrix using modified versions of the existing techniques. First, we improved the Linear Programming approach by incorporating additional network-specific information. We found that doing this improves the performance of Linear Programming further and also helps in estimating the traffic matrix of small networks precisely. Second, we developed a hybrid model that combines Tomogravity, Entropy Maximization and Linear Programming. This model best estimates the traffic matrix compared to the other three methods. Third, we proposed a hierarchical model that helps to estimate the traffic matrix of large networks at coarse level with better accuracy. This hierarchical model can also be used to assess the quality of traffic estimation for real life networks for which the true matrix is not known. |
