Automated load balancing in SPEEDES

Parallel discrete event simulation offers the potential for significant speedup over sequential simulation. Unfortunately, high performance is often achieved only after rigorous fine tuning is used to obtain an efficient mapping of tasks to processors. In practice, good performance with minimal effort is often preferable to high performance with excessive effort. We discuss our research in adding automated load balancing to the SPEEDES simulation framework. Using simulation models of queuing networks and the National Airspace System, we demonstrate that using run time measurements, our automated load balancing scheme can achieve better performance than simple allocation methods that do not use run time measurements, particularly when large numbers of processors are used.