Determining initial states for time-parallel simulations

Abstract

Time-parallel simulations exploit parallelism by partitioning the time domain of a simulation model. Exploiting temporal parallelism requires predicting future states of a simulation model. A poor prediction of future states may cause extensive recomputation so that a time-parallel simulation requires more real time to execute than a corresponding sequential simulation. Recurrent states of a simulation model provide a way to predict future states. In this paper, we propose a time-parallel simulation method which uses a pre-simulation to identify recurrent states. For simulation models in which recurrent states do not exist or can not support sufficient time parallelism, an approximation technique is suggested to extend the class of simulation models to which time-parallel simulation can be applied. Several queueing network models are investigated with the proposed time-parallel simulation. Experimental result suggest that the proposed approach can exploit massive parallelism while yielding accurate results.