Internet Traffic Characterization Using Packet-Pair Probing

This paper presents an edge-based Internet traffic characterization approach. Our objective is to estimate the marginal distribution and the correlation structure of the packet arrival process at a queue, by sending probe packet pairs with a specific dispersion to sample the traffic. The aggregate work load process inferred from the output dispersions is a compound process of the packet arrival process and the packet size distribution. We show that the packet arrival marginal distribution and the packet size distribution can be decoupled by using the probability generating function; given one of the distributions, the other can then be estimated. We use the fact that the Internet packet size follows a known multi-modal distribution. Moreover, multiple series of packet pairs with different input dispersions can be used to estimate the packet arrival process at different time scales, and therefore to estimate the Hurst parameter, which characterizes the long-range dependence, by generating the variance-time plot. While the traffic characterization techniques are developed and validated in a single-queue context, we indicate how the techniques can be applied to a black box system for end-to-end quality of service provisioning along a multi-hop path.