Accelerating the Evaluation of Parallel Program Performance Models Using Distributed Simulation

Petri nets with an explicit notion of time have proven to be a good modelling tool for the qualitative and quantitative study of asynchronous concurrent systems with parallelism, synchronization and resource sharing. The formalism could be successfully applied in performance modelling and evaluation of concurrent programs, parallel systems and mappings among the two. However, performance models of realistic systems are in general very costly to evaluate. In this work we develop a Time Warp based distributed discrete event simulation scheme to accelerate the evaluation of timed Petri net performance models. The particular model structure that typically results from performance modelling of parallel systems explicitly provides partitioning possibilities which are very cumbersome to identify in general Petri net models. With our implementations of a distributed simulation engine running on an Intel iPSC/860 multiprocessor we show that real speedup over sequential simulation model evaluation can be achieved depending on the inherent model parallelism.