Performance Benchmark of a Parallel and Distributed Network Simulator

Simulation of large-scale networks requires enormous amounts of memory and processing time. One way of speeding up these simulations is to distribute the model over a number of connected workstations. However, this introduces inefficiencies caused by the need for synchronization and message passing between machines. In distributed network simulation, one of the factors affecting message passing overhead is the amount of cross-traffic between machines. We perform an independent benchmark of theParallel/Distributed Network Simulator (PDNS) based on experimental results processed at the Australian Centre for Advanced Computing and Communications (AC3) supercomputing cluster. We measure the effect of cross-traffic on wall-clock time needed to complete a simulation for a set of basic network topologies by comparing the result with the wall-clock time needed on a single processor. Our results show that although efficiency is reduced with large amounts of cross-traffic, speedup can still be achieved with PDNS. With these results, we developed a performance model that can be used as a guideline for designing future simulations.