Inverting flow durations from sampled traffic

The flow duration distribution is an important metric to the network operator for traffic prediction and accounting but also arguably from the viewpoint of the user. The inversion problem of recovering the flow duration distribution from sampled traffic is addressed here under several sampling methods. A theoretical framework for the inversion problem is developed using a probabilistic flow model. In this framework, direct equations for the distributions of sampled flow quantities are derived based on the distributions of original flow characteristics. The inversion of these equations provides estimators for the flow duration distribution. Finally, the inversion techniques under several sampling schemes are evaluated on two Internet traces.