An Empirical Analysis of Parallelization Decisions Affecting Parallel Simulation Performance

This research investigated the effects of computer system architecture and simulation model decomposition strategies on the performance of parallel time-driven simulations. A Ballistic Missile Defense (BMD) t' imedriven simulation program was used as a test vehicle for investigating parallelization options for distributed and shared memory architectures. Eight implementations were developed for a 32-node Intel iPSC/l hypercube, one for an 8-node iPSC/2 hypercube, and one for a 16-node Encore Multimax. The iPSC/l implementations represented the range between purely functional decompositions and pure data decompositions, including both static data assignment and master-slave configureations where the data was dynamically assigned as nodes became idle. The knowledge gained from this investigation was captured as a set of simulation parallelization guidelines based on the relative performance of the different implementations and the lessons learned during the paralleliration process.