Server-based inference of Internet link lossiness

The problem of inferring the packet loss characteristics of Internet links using server-based measurements is investigated. Unlike much of existing work on network tomography that is based on active probing, we make inferences based on passive observation of end-to-end client-server traffic. Our work on passive network tomography focuses on identifying lossy links (i.e., the trouble spots in the network). We have developed three techniques for this purpose based on random sampling, linear optimization, and Bayesian inference using Gibbs sampling, respectively. We evaluate the accuracy of these techniques using both simulations and Internet packet traces. We find that these techniques can identify most of the lossy links in the network with a manageable false positive rate. For instance, simulation results indicate that the Gibbs sampling technique has over 80% coverage with a false positive rate under 5%. Furthermore, this technique provides a confidence indicator on its inference. We also perform inference based on Internet traces gathered at the busy microsoft.com Web site. However, validating these inferences is a challenging problem. We present a method for indirect validation that suggests that the false positive rate is manageable.