Dynamically switching between lazy and aggressive cancellation in a Time Warp parallel simulator

The performance of a Time Warp parallel discrete event simulator (PDES) depends on the efficiency of the cancellation strategy employed to undo the effects of the erroneous computation. Two known cancellation strategies exist, namely aggressive cancellation and lazy cancellation. Under aggressive cancellation (AC), when a straggler arrives at a logical process (LP), it rolls back to an appropriate previous state and immediately sends out anti messages for all the messages that were processed prematurely. In contrast, under lazy cancellation (LC), the anti messages are delayed until forward processing demonstrates that the originally sent output messages were incorrect. The performance under LC deteriorates if the probability of the regenerated output messages being different from the originally sent messages, is high. AC, on the other hand, performs badly if the same messages are generated before and after a rollback most of the time. In general, it has proven to be difficult to determine a priori the favorable cancellation strategy. Thus, we propose that the performance of a Time Warp simulator is best optimized by having the simulation dynamically select the cancellation strategy to be used based on the behavior of the application being simulated. A simple approach to achieve this is based on a parameter called the hit/miss ratio. The results obtained using this adaptive approach are compared with the results obtained using only aggressive cancellation or lazy cancellation. These results show that for our application, digital system simulation, the adaptive technique works better than either cancellation strategy