A scaled version of the elastic time algorithm

摘要

Recently a new class of synchronization algorithms for parallel discrete event simulation has been proposed, namely the near perfect state information algorithms, which are based on a notion of error potential to control the optimism of event execution. An algorithm of this class, called elastic time algorithm (ETA), has been instantiated. In this algorithm, the error potential is computed using temporal information (next event timestamp, simulation clocks etc.) and is then translated into event execution delay based on a constant factor. In this paper we present a scaled version of ETA (SETA), in which the error potential is translated into event execution delay based on both a constant factor and an additional scaling factor determined dynamically as a function of the event granularity.

We have implemented versions of ETA and SETA for a cluster of PCs connected by a Myrinet switch and we have established in an empirical study that SETA outperforms ETA if there is difference inthe granularity of different event types.