Compressed Network Monitoring

This paper describes a procedure for estimating a full set of network path metrics, such as loss or delay, from a limited number of measurements. The approach exploits the strong spatial and temporal correlation observed in path-level metric data, which arises due to shared links and stationary components of the observed phenomena. We design diffusion wavelets based on the routing matrix to generate a basis in which the signals are compressible. This allows us to exploit powerful non-linear estimation algorithms that strive for sparse solutions. We demonstrate our results using measurements of end-to-end delay in the Abilene network. Our results show that we can recover network mean end-to-end delay with 95% accuracy while monitoring only 4% of the routes.