Dynamic Cancellation: a Heuristic for Selecting Cancellation Strategies in Time Warp Simulators

The performance of Time Warp parallel discrete event simulators can be a ected by the cancellation strategy used to send anti-messages. Under aggressive cancellation, anti-message generation occurs immediately after a straggler message is detected. This results in quick termination of potentially incorrect computations (event processing) when out-of-order event processing occurs. In contrast, lazy cancellation delays the sending of anti-messages until forward processing from a straggler message con rms that the previous lookahead computation did indeed generate an incorrect message. Lazy cancellation errors on the side of assuming that out-of-order processing will frequently produce correct results; aggressive cancellation assumes that out-of-order processing will generally produce incorrect results. Unfortunately, previous studies have shown that neither approach is clearly superior to the other in all cases (even within the same application domain). It has proven to be extremely di cult to make an a priori determination of the favorable cancellation strategy and most existing Time Warp systems merely provide a switch for the user to select the cancellation strategy to be employed. This thesis explores the use of simulation time decision procedures to select cancellation strategies. The approach is termed Dynamic Cancellation and provides the capability for selecting cancellation strategies to the Logical Processes (LPs) in a Time Warp simulation. Thus, within a single parallel simulation both strategies may be employed by distinct LPs and even across the simulation lifetime of an LP. Empirical data using a non-linear control mechanism to select the cancellation strategy shows a nominal performance gain (3-5%) from dynamic cancellation. ii