Scalability, locality, partitioning and synchronization PDES

Abstract

As parallel discrete event simulation becomes increasingly important to the solution of very large systems design problems, it becomes increasingly critical to establish whether PDES technology will scale up with increasing problem size and architecture. We address the problem in a general setting, and provide a resounding conclusion: maybe. To scale requires that the simulation model not grow in ways that defeat an ability to load balance, and that do not overwhelm any one processor with communication. It requires an architecture that scales as well. It requires a partitioner that balances workload and exploits locality of communication. The specific partition strategy we examined is very simple; our point is not to promote its specific use. Our point is that scalability is possible using it, and hence if a more refined partitioner can balance workload and maintain locality of communication, then simulations built using it will scale also. If these conditions apply, then we demonstrate by example a simple conservative synchronization protocol, QS, that scales. Using QS, we then examine the trade-off between load imbalance and synchronization overhead. We show how to efficiently manage that trade-off by probing the space of potential restricted partitions.