Sharing Event Data in Optimistically Scheduled Multicast Applications

Abstract

A major consideration when designing high performance simulation models is state size. Keeping the model state sizes small enhances performance by using less memory, thereby increasing cache utilization which leads to reduced model execution time. Assuming an otherwise efficient simulation executive and minimal model state, the only remaining area for reducing model size is within the events they create. The event population is typically the most memory intensive region within a simulation especially in the case of multi/broadcast like applications which tend to schedule many events within the atomic processing of a single event. To tackle the issue of excessive event memory consumption in multicast applications, this paper introduces the idea of shared event data. Here, the readonly data section is shared for a multicast event, which may then be delivered to several LPs. The critical aspect to processing events with shared data is that the simulation executive must ensure that the shared data is properly maintained not destroyed prematurely. In this paper we present an approach for sharing event data within optimistic simulation system and demonstrate performance on a multicast benchmark application. From our performance study, we report a 22% reduction in the data cache miss rate, a processor utilization in excess of 80% and a reduction in model memory consumption by a factor of 20.