Scalability Study of Molecular Dynamics Simulation: Multi-core VS Many-core

Abstract

Abstract—Molecular dynamics (MD) simulation has broad applications, and increasing computing power is needed to satisfy the large spatiotemporal scales of the real world simulation. The advent of multi-core and many-core paradigm brings unprece-dented computing power, however, it remains a great challenge to harvest the computing power due to MD’s irregular memory-access pattern. To address the challenge, this paper presents a joint application/architecture study to enhance scalability of MD on multi-core and many-core architecture. First, a hierar-chical scheme is designed to explore the multi-level parallelism mapping application to hardware. Then to further harvest the many-core computing power, three incremental optimization strategies—-a novel data-layout to improve data locality, an on-chip locality-aware parallel algorithm to enhance data reuse, and a pipelining algorithm to hide latency to shared memory—- are