Modeling of Packet Interactions in Dispersion-Based Network Probing Schemes

Measurements of end-to-end available bandwidth is getting increasingly important in the Internet. Many measurement methods are based on analysis of the dispersion of probe packets that have traversed the network according to some probing scheme. In this paper we present a theoretical framework to model interactions, at the discrete packet level, between probe packets and cross-traffic packets for dispersion-based probing schemes. We believe that an understanding of packet interactions is crucial to develop accurate and efficient measurement methods as well as to evaluate trade-offs between different probing schemes. Using the framework, we show that for statistical metrics such as the mean there are significant differences between using packet-pair and packet-train probing schemes when subject to cross traffic. Hence, these two probing schemes should not be used interchangabely. We show that the differences between packet-pair and packet-train probing schemes are due to correlations of dispersion values obtained from packet train probing. These correlations do not arise in packet-pair probing. The correlations are manifested as three characteristic patterns. The correlations and their corresponding patterns are identified and analyzed within our framework.