Asynchronous (Time-Warp) versus synchronous (Event-Horizon) simulation time advance in BSP

Abstract

This paper compares the very fundamental concepts behind two approaches to optimistic parallel discrete-event simulation on BSP computers [10]. We refer to (i) asynchronous simulation time advance as it is realised in the BSP implementation of Time Warp [3], and (ii) synchronous time advance as it is realised in the BSP implementation of Breathing Time Buckets [8]. Our results suggest that asynchronous time advance can potentially lead to more efficient and scalable simulations in BSP. IntroductionAs BSP is a (bulk) synchronous model of parallel computing [10], the obvious method of parallel discrete-event simulation [2, 7] on BSP computers seems to be synchronisation protocols based on synchronous time advance, such as Breathing Time Buckets [8]. In this paper we show that using synchronous time advance on a BSP computer might not be a good idea for two reasons. Firstly, synchronous simulation methods force the execution of periodical parallel minreductions during which no interesting advance in simulation time can be made. These operations take some number ofBSP supersteps to complete [10], and the simulation is stopped during their execution. The cumulative cost of these operations is relevant since the total number of min-reductions is equal to the total number of supersteps used to process the simulation events. Secondly and more importantly, synchronous methods impose barriers in simulation time which tend to reduce the rate of time advance per superstep of the simulation. This increases the total number of supersteps dedicated to parallel event-processing. Equivalently, time barriers tend to decrease the number of events that are processed in each superstep.In BSP, supersteps are delimited by the barrier synchronisation of all the processors. In turn this operation takes some amount of reM time to complete and it can become comparatively expensive in simulation models where the amount of computation involved in the processing of events is small (e.g., queuing networks). As many real life simulations can easily demand the execution of hundreds of thousands of event-processing supersteps, a significant reduction of the total number of supersteps can also cause a significant reduction of the total running time of the simulation. In addition, the cost of the remaining barrier

Automatic Parallel-Discrete Event Simulation

“…Those include the synthetic work-load PHold, Wind energy electricity generation systems, hard-particles models, Web crawlers, and a large toriodal queuing network. In all cases, specially in regular systems, we have observed that our prediction methodology is very effective in practice (this claim is supported by the results in [3,4]). …”