An optimal repartitioning decision policy

The automated partitioning of simulations for parallel execution is a timely research problem. A simulation's run-time performance depends heavily on the nature of the inputs the simulation responds to. Consequently, a simulation's run-time behavior varies as a function of time. Since a simulation's run-time behavior is generally too complex to analytically predict, partitioning algorithms must be statistically based: they base their partitioning decisions on the simulation's observed behavior. Simulations which are partitioned statistically are vulnerable to radical changes in the run-time dynamics of the simulation. In this paper we discuss a dynamic repartitioning decision policy which detects change in a simulation's run-time behavior and reacts to this change. This decision policy optimally balances the costs and potential benefits of repartitioning a running simulation.