PlanetSeer: Internet Path Failure Monitoring and Characterization in Wide-Area Services

Detecting network path anomalies generally requires examining large volumes of traffic data to find misbehavior. We observe that wide-area services, such as peer-to-peer systems and content distribution networks, exhibit large traffic volumes, spread over large numbers of geographically-dispersed endpoints. This makes them ideal candidates for observing wide-area network behavior. Specifically, we can combine passive monitoring of wide-area traffic to detect anomalous network behavior, with active probes from multiple nodes to quantify and characterize the scope of these anomalies.

This approach provides several advantages over other techniques: (1) we
obtain more complete and finer-grained views of failures since the
wide-area nodes already provide geographically diverse vantage points; (2)
we incur limited additional measurement cost since most active probing is
initiated when passive monitoring detects oddities; and (3) we detect
failures at a much higher rate than other researchers have reported since
the services provide large volumes of traffic to sample. In this talk, we
describe how to exploit this combination of wide-area traffic, passive
monitoring, and active probing, to both understand path anomalies and to
provide optimization opportunities for the host service.

Ming Zhang is a Ph.D. candidate at Princeton University. He received his
B.S. from Nanjing University in China in 1999, and his M.A. from Princeton
University in 2001, both in Computer Science. While at Princeton, he
worked on topics including Internet failure measurement and robust
multi-path transport layer protocols. Before that, he worked on
reordering-robust congestion control during his internship at
International Computer Science Institute (ICSI), and flow control and
queries in sensor networks at Intel Research Pittsburgh. His research
interests include overlay networks, network troubleshooting, network
routing, transport layer protocols, wireless and sensor networks.