Network Tomography with Router Interface Data

Keywords: traffic matrix, identifiability, network data

Abstract: We discuss methods to estimate source-by-destination traffic matrices using only link counts from routers on a network. These routers commonly collect link statistics and report them when polled by a network management system. Solutions to many network engineering problems, however, require information on the amount of traffic flowing between all pairs of edge nodes. Our methods take readily the available link count measurements and invert them to estimate full traffic matrices that vary with time.

We develop a model and estimation scheme using link count data from a local area network and then verify its performance using measurements of complete traffic matrices. The model has more parameters than the number of observed variables but is, nevertheless, identifiable. Our initial study uses a simple network with a single router and 4 edge nodes leading to 4x4 traffic matrices. A second example demonstrates a more complex 8x8 case.