Experiments in Automated Load Balancing

One of the promises of parallelized discrete-event simulation is that it might provide significant speedups over sequential simulation. In reality, high performance cannot be achieved unless the system is fine-tuned to balance computation, communication, and synchronization requirements. In this paper, we discuss our experiments in automated load balancing using the SPEEDES simulation framework. Specifically, we examine three mapping algorithms that use run-time measurements. Using simulation models of queuing networks and the National Airspace System, we investigate (i) the use of run-time data to guide mapping, (ii) the utility of considering communication costs in a mapping algorithm, (iii) the degree to which computational ``hot-spots'' ought to be broken up in the linearization, and (iv) the relative execution costs of the different algorithms. We compare the performance of the three algorithms using results from the Intel Paragon.