Server-based Characterization and Inference of Internet Performance

Venkat Padmanabhan
Microsoft Research

We present a study wide-area Internet performance as inferred from passive observation of network traffic at a busy Web site. We first examine the correlation between topological distance (i.e., hop count) and network performance. We find that the correlation is weak. One reason is that all links are not ``equal''. The end-to-end performance may be dominated by a few lossy links. We also find that there is a significant spatial locality in the packet loss rate experienced by clients. This finding suggests that at least in some cases, a high packet loss rate experienced by clients is due to a shared lossy link.

These findings lead to our main focus: identifying lossy links based on end-to-end performance information. Unlike much of existing work on network tomography that is based on active probing, we investigate several techniques for identifying lossy links based on passive observation of end-to-end client-server traffic. We evaluate the accuracy of these techniques using extensive simulations and also Internet packet traces. We find that these techniques can identify most of the lossy links in the network with a manageable false positive rate.


Back to Large-Scale Communication Networks: Topology, Routing, Traffic, and Control