Strategies for Efficient Executions of Irregular Message-Driven Parallel Applications on GPU Systems

TL;DR

This paper presents strategies integrated into the G-Charm framework to improve the efficiency of message-driven irregular applications on GPU systems, achieving significant reductions in execution time.

Contribution

It introduces novel runtime strategies for irregular data access and hybrid execution, specifically designed for GPU systems, and demonstrates their effectiveness on real applications.

Findings

8-38% reduction in execution times

Effective handling of irregular data accesses

Improved performance over static strategies

Abstract

Message-driven executions with over-decomposition of tasks constitute an important model for parallel programming and have been demonstrated for irregular applications. Supporting efficient execution of such message-driven irregular applications on GPU systems presents a number of challenges related to irregular data accesses and computations. In this work, we have developed strategies including coalescing irregular data accesses and combining with data reuse, and adaptive methods for hybrid executions to minimize idling. We have integrated these runtime strategies into our {\em G-Charm} framework for efficient execution of message-driven parallel applications on hybrid GPU systems. We demonstrate our strategies for irregular applications with an N-Body simulations and a molecular dynamics application and show that our dynamic strategies result in 8-38\% reduction in execution times for these irregular applications over the corresponding static strategies that are amenable for regular applications.