On the processor scheduling problem in time warp synchronization

Abstract

Time Warp is a synchronization mechanism for parallel/distributed simulation. It allows logical processes (LPs) to execute events without the guarantee of a causally consistent execution. Upon the detection of a causality violation, rollback procedures recover the state of the simulation to a correct value. When a rollback occurs there are two primary sources of performance loss: (1) CPU time must be spent for the execution of the rollback procedures and (2) waste of CPU time arises from the invalidation of event executions. In this paper we present a general framework for the problem of scheduling the next LP to be run on a processor in Time Warp simulations. The framework establishes a class of scheduling algorithms having the twofold aim to keep low the CPU time for the execution of the rollback procedures and also to guarantee low waste of time due to event executions invalidated by rollback. The combination of these two aims should actually lead to short completion time of Time Warp simulations.We collocate existing scheduling algorithms within the framework, pointing out how they miss previous aims, at least partially. Then we instantiate a Window-based Grain Sensitive (WGS) scheduling algorithm relying on the framework, which pursues the above twofold aim. We also identify the proper conditions, associated with the simulation model execution, under which any algorithm exploiting the framework structure is expected to benefit the performance of the Time Warp mechanism. Empirical evidence from an experimental study of WGS on classical benchmarks and on a mobile communication system simulation fully confirms the theoretical outcomes.