Rigorous Statistical Analysis of Internet Loss Measurements

摘要

In this paper we present a rigorous technique for estimating confidence intervals of packet loss measurements. Our approach is motivated by simple observations that the loss process can be modelled as an alternating renewal process. We use this structure to build a Hidden Semi-Markov Model (HSMM) for the measurement process, and from this estimate both loss rates, and their confidence intervals. We use both simulations and a set of more than 18000 hours of real Internet measurements (between dedicated measurement hosts, PlanetLab hosts, web and DNS servers) to cross-validate our estimates, and show that they are significantly more accurate than any current alternative.