Piggy-backed time-stepped simulation with 'super-stepping'

Abstract

We propose an optimization technique for reducing global synchronizations in traditional time-stepped simulations. Time-stepped simulations are known to be efficient when events are frequent or dense. However, when events are less frequent (when compared to the size of time-steps) the performance of time-stepped simulations drop noticeably. This paper aims at improving the performance of traditional time-stepped simulations during low frequency periods and maintaining its efficiency during high frequency periods. We focus on interactive simulations which have tight real-time interactive constraints. The proposed optimization is achieved by informing the host about future events. This information is 'piggybacked' on the ready messages sent by the participating Processing Elements (PE) to the host. We maintain simulation efficiency by switching between the proposed technique and the traditional technique depending on the observed event density. To achieve this switching we introduce a concept called 'super-stepping'. A probabilistic method is used to optimize 'super-step' size.