A time warp synchronized distributed simulation organizes a discrete event simulation into concurrently executing Logical Processes (LPs) that use a rollback recover mechanism whenever a causality violation is detected. The LPs operate independently from one another and when performing a rollback, they must ensure the proper management of any prematurely generated output events. In general, there are two main techniques to manage the prematurely generated output events, namely: aggressive cancellation and lazy cancellation. When using aggressive cancellation, the LP will immediately send anti-messages to cancel the events which were prematurely generated. In contrast, lazy cancellation will not send anti-messages until re-computation shows that the original, prematurely sent output event is incorrect.
While lazy cancellation was originally proposed as an optimization to time warp, numerous studies failed to show that lazy cancellation would consistently deliver superior performance to aggressive cancellation. In fact, these studies failed to show that either strategy was consistently superior to the other. For some simulation models, lazy cancellation was superior and for others aggressive cancellation was superior. As a result of these studies, a third technique called dynamic cancellation was developed. Under dynamic cancellation, each LP selects either lazy or aggressive cancellation based on its past performance. LPs whose premature output events tend to be correct use lazy cancellation; all others use aggressive cancellation. Studies with dynamic cancellation have shown it to perform on par with the better performing static method. However all of the above studies have been on either shared memory or distributed memory platforms. The performances of cancellation strategies on multi-core Beowulf clusters with both shared memory and distributed memory construction have not been well studied. This thesis explores the comparative performance of various cancellation strategies implemented for Threaded Warped.