Practical estimation of Internet traffic demands

Traffic engineering and network management tasks enable the operation of communication networks at optimal levels of performance and efficiency. Performance is usually captured by metrics perceived by users such as message delays and losses. Efficiency usually concerns how well the available network capacity is utilized. The ultimate objective of achieving both high performance and high efficiency is very challenging for operational networks for a wide variety of reasons. Consequently, network operators generally opt for straightforward but needed strategies to achieve the performance goals at the expense of sub-optimal levels of efficiency.; Accurate knowledge of the characteristics of end-to-end traffic demands in a communication network is a key factor in enabling network operators to succeed in achieving the aforementioned objective. Such traffic information can be succinctly represented in matrix form. A Traffic Matrix (TM) represents the volume of traffic that flows between all source-destination pairs in a network. In a TM X, rows and columns represent nodes in the network, and element x_ij represents the volume of traffic exchanged from node i to j. In this talk I will describe mechanisms we have investigated and proposed to enable network operators to fully populate TMs as accurately and efficiently as possible. In order to achieve this goal, the following steps were followed. First, a taxonomy of TMs is developed with the objective of providing a common language for the description and analysis of TMs. Second, the feasibility of fully measuring complete TMs in the context of Tier-1 backbone networks is analyzed. Third, a thorough comparative analysis of the main statistical techniques for estimating network traffic demands previously proposed when the thesis started is performed. Fourth, a mechanism is proposed for the generation of informed starting points to be provided as input for statistical techniques. Fifth, an approach to TM estimation is proposed which combines our mechanism for generating starting points with a fast variant of an Expectation-Maximization based statistical technique. Finally, a methodology is developed for estimating TMs in the context of a commercial Tier-1 backbone network. The developed methodology is embedded in a network analysis software toolkit.