|
Abstract:
|
Many aspects of network management, such as network provisioning, and traffic anomaly detection, rely on knowing the traffic flow statistics between pairwise network nodes. However, rapid growth of networks in size and bandwidth makes obtaining such statistics very challenging. In this paper, we propose efficient streaming algorithms to generate compact on-line statistical digests of traffic, from which likelihood based estimates are derived. The new approaches are scale-free, i.e., the accuracy essentially does not depend on the unknown traffic parameters. We perform a thorough theoretical and simulation analysis of our algorithms and evaluate them using a traffic matrix from a top-tier Internet Service Provider. These analyses demonstrate that the new algorithms yield more accurate estimates than existing approaches, and are robust to heterogeneous and dynamic network conditions.
|
