Cache-aware memory manager for Optimistic Simulations

Abstract

Parallel Discrete Event Simulation is a well known technique for executing complex general-purpose simulations where models are described as objects the interaction of which is expressed through the generation of impulsive events. In particular, Optimistic Simulation allows full exploitation of the available computational power, avoiding the need to compute safety properties for the events to be executed. Optimistic Simulation platforms internally rely on several data structures, which are meant to support operations aimed at ensuring correctness, inter-kernel communication and/or event scheduling. These housekeeping and management operations access them according to complex patterns, commonly suffering from misuse of memory caching architectures. In particular, operations like log/restore access data structures on a periodic basis, producing the replacement of in-cache buffers related to the actual working set of the application logic, producing a non-negligible performance drop. In this work we propose generally-applicable design principles for a new memory management subsystem targeted at Optimistic Simulation platforms which can face this issue by wisely allocating memory buffers depending on their actual future access patterns, in order to enhance event-execution memory locality. Additionally, an application-transparent implementation within ROOT-Sim, an open-source generalpurpose optimistic simulation platform, is presented along with experimental results testing our proposal.