IP Traffic Matrix Estimation Methods: Comparisons and Improvements

Determining point to point traffic matrix is essential for Internet service providers (ISPs) in carrying out traffic engineering tasks for network management and planning purposes. However, it is very difficult and costly to measure this traffic matrix directly. Hence, traffic matrices are inferred from link measurements through estimation, using different techniques. There are different techniques for this traffic matrix estimation and there is still a need for evaluating these existing techniques. Some of those techniques have been previously compared, but with new improved techniques recently developed there is a need to revisit the comparisons. In this paper, we have carried out studies to compare three very popular methods: the tomogravity, the entropy maximization and linear programming methods. We find that the tomogravity method best estimates the traffic matrix among the methods we tested. We then incorporate some enhancements which improve this method. Specifically we established that knowing some point to point traffic may improve the estimation but not necessarily, and this is counter-intuitive. We modify the existing entropy maximization method by adding more constraints and we find that our modified method outperforms the existing entropy maximization and tomogravity methods.