An Empirical Study of Network Topology Inference

Abstract

Summary. Understanding network topology is important for evaluating the performance of network protocols, for detecting large-scale denial-of-service or malicious intrusion, improving the design of resource provisioning, or studying the scalability of multicast. Usually, to infer the topology of a large-scale network, researchers use multiple vantage points to conduct extensive traceroute-based measurements. However, an inferred topology is often incomplete or inaccurate because the traceroute technique on which we heavily rely has inherent limitations. Furthermore, the Internet is so complicated and dynamic that discovering an immediate snapshot of its topology to be a very challenging task. In our experiments, to identify a large ISP cloud, we spread vantage points inside the cloud and over the world, and collect topology information by probing a fixed list of IP addresses which consists of more than 25,000 routers and 36,000 links. Data analysis shows that sampling bias, if undetected, could significantly undermine the conclusions drawn from the inferred topologies.