Event scheduling based on combined rollback cost in time warp parallel simulation

An event scheduling policy for Time Warp parallel simulations based on combined rollback cost considerations is proposed. This policy takes into consideration a multitude of factors which affect the selection of the next event to be processed by any processor. These factors are: the event timestamp, the expected event execution time (granularity), the expected number of children events to be produced as a result of processing the candidate parent event and the particular event cancellation scheme used. Our scheduling policy is suitable for simulation models with event types having significantly different event granularity values and different numbers of children events. On any processor, it selects for scheduling from a set of candidate events the one having the smallest sum of granularity value and the expected cancellation cost for children events to be sent to other logical processes. This helps delay the execution of those candidate events which have high combined rollback costs as measured by the sum of granularity values and the expected cancellation cost of children events. Experimental results for a synthetic workload are reported. These results show that our scheduling policy can achieve a smaller simulation execution time and a fewer number of rollbacks than both LTF and grain-sensitive scheduling.