Adaptive Time Warp Simulation of Timed Petri Nets

Time Warp (TW), although generally accepted as a potentially effective parallel and distributed simulation mechanism for timed Petri nets, can reveal deficiencies in certain model domains. Particularly, the unlimited optimism underlying TW can lead to excessive aggressiveness in memory consumption due to saving state histories, and waste of CPU cycles due to overoptimistically progressing simulations that eventually have to be “rolled back.” Furthermore, in TW simulations executing in distributed memory environments, the communication overhead induced by the rollback mechanism can cause pathological overall simulation performance. In this work, an adaptive optimism control mechanism for TW is developed to overcome these shortcomings. By monitoring and statistically analyzing the arrival processes of synchronization messages, TW simulation progress is probabilistically throttled based on the forecasted timestamp of forthcoming messages. Two classes of arrival process characterizations are studied, reflecting that a natural tradeoff exists among the computational and space complexity, and the respective prediction accuracy: While forecasts based on metrics of central tendency are computationally cheap but yield inadequate predictions for correlated arrivals (thus negatively affecting performance), time series based forecast methods give higher prediction accuracy, but at higher computational cost. The sensitivity of the adaptive optimism control with respect to forecast accuracy and computational overhead is analyzed for very large Petri net simulation models executed with the TW protocol on the Meiko CS-2 multiprocessor, and for a stress case scenario on the CM-5. Empirical evidence is delivered showing that: 1) probabilistic optimism control, regardless of the communication-computation speed ratio of the target execution platform, automatically finds the most appropriate synchronization policy in the spectrum between optimistic TW and conservative Chandy/Misra/Bryant schemes, 2) local control decisions yield an efficient exploitation of simulation model parallelism that is “local” to particular spatial regions, and 3) even if simulation progresses in “phases” of different performance behavior (nonstationary simulations), logical processes can dynamically readjust their synchronization policy, thus in a natural way evading the partitioning problem under imbalanced loads.