Evaluating the limits of network topology inference via virtualized network emulation

Abstract

The Internet measurement community is beset by a lack of ground truth, or knowledge of the real, underlying network in topology inference experiments. While better tools and methodologies can be developed, quantifying the effectiveness of these mapping utilities and explaining pathologies is difficult, if not impossible, without knowing the network topology being probed. In this thesis we present a tool that eliminates topological uncertainty in an emulated, virtualized environment. First, we automatically build topological ground truth according to various network generation models and create emulated Cisco router networks by leveraging and modifying existing emulation software. We then automate topological inference from one vantage point at a time for every vantage point in the network. Finally, we incorporate a mechanism to study common sources of network topology inference abnormalities by including the ability to induce link failures within the network. In addition, this thesis reexamines previous work in sampling Autonomous System-level Internet graphs to procure realistic models for emulation and simulation. We build upon this work by including additional data sets, and more recent Internet topologies to sample from, and observe divergent results from the authors of the original work. Lastly, we introduce a new technique for sampling Internet graphs that better retains particular graph metrics across multiple timeframes and data sets.