On predicting the performance characteristics of the ns-3 distributed simulator for scale-free internet models

Abstract

An inquiry is made into whether or not it is possible to learn a performance characterization which can predict the best way to configure the simulator executive when performing a simulation using the ns-3 network simulator distributed across many CPUs. In testing this inquiry, experiments are run in ns-3 using network topologies generated using BRITE and the total wallclock runtime of the simulation is measured. For each experiment, a set of features is formed which is composed of measurements made prior to the running of the simulation. For each experiment, a binary label is created indicating which simulator executive configuration produced better performance. A linear classifier is then designed which nearly separates the set of experiments by label, indicating that such a performance characterization may exist.