In network performance tomography, characteristics of the network interior are inferred by correlating end-to-end measurements. In much previous work, the presence of correlations must be arranged at the packet level, e.g., using multicast probes or unicast emulations of them. This carries costs in deployment and limits coverage. However, it is difficult to determine performance characteristics without correlations. Some recent work has had success in reaching a lesser goal—identifying the lossiest network links— using only uncorrelated end-to-end measurements. In this paper we the required properties of network performance, and show that they are independent of the particular inference algorithm used. This observation allows us to design a quick and simple inference algorithm that identifies the worst performing link in a badly performing subnetwork, with high likelihood when bad links are uncommon. We give several examples of perforance models and that exhibit the required pro...
Nick G. Duffield