On the sensitivity of network simulation to topology

Abstract

While network simulations for congestion control studies have often varied traffic loads and protocol parameters, they have typically investigated only a few topologies. The most common is by far the so-called "barbell" topology. We argue, first, that the "barbell" topology is not representative of the Internet. In particular we report that a measurable fraction of packets pass through multiple congestion points. Second, we argue that the distinction between the "barbell" topology and more complex topologies is relevant by presenting a scenario with multiple congestion points that exhibits behavior that seems unexpected based on intuition derived from the barbell topology (in particular, a TCP-only system that exhibits behavior technically considered "congestion collapse"). We make the larger argument that the typical methodology currently accepted for evaluating network protocols is flawed. Finally, we comment on some issues that arise in designing a simulation methodology that will be better suited to comparison of network protocol performance.