Particle-Mesh Ewald(PME)算法的GPU加速  

Implementation of Particle-Mesh Ewald(PME) on Graphics Processing Units

作　　者：徐骥[1,2] 葛蔚[1] 任瑛[1,2] 李静海[1] 

机构地区：[1]中国科学院过程工程研究所多相反应实验室,北京100190 [2]中国科学院研究生院,北京100039

出　　处：《计算物理》2010年第4期548-554,共7页Chinese Journal of Computational Physics

基　　金：国家自然科学基金(20221603);国家科技支撑计划(2008BAF33B01);中国科学院知识创新工程(KGCX2-YW-124)资助项目

摘　　要：讨论在NVIDIACUDA开发环境下,用GPU加速分子动力学模拟中静电作用的长程受力计算部分.采用Particle-Mesh Ewald(PME)方法,将其分解为参数确定、点电荷网格离散、离散网格的傅立叶变换、静电热能求解与静电力求解5个部分,并分别分析各部分的GPU实现.此方法已成功用于7个不同大小的生物分子体系的模拟计算,达到了7倍左右的加速.该程序可耦合到现有分子动力学模拟软件中,或作为进一步开发的GPU分子动力学程序的一部分,显著加速传统分子动力学程序.In this article,long range part of electrostatic interaction is accelerated using GPUs in NIVIDIA CUDA programming environment.Particle Mesh Ewald algorithm is adopted and split into 5 procedures：parameters designation,discretization of point charges into grids,Fourier transformation of grids,potential of electrostatic interaction,and force calculation of electrostatic interaction.The codes are tested by seven biomolecular systems with different sizes.Roughly 7-fold speedup over one core of mainstream CPUs is obtained.The codes can be integrated into molecular dynamic simulation software packages that already exst or used as part of GPU codes developed in the future to further speedup traditional MD simulations.

关 键 词：PME(Particle-Mesh Ewald)加速 GPU(Graphic Processing Unit图形处理单元) CUDA(Compute Unified Device Architecture计算统一设备架构) 

分 类 号：O241[理学—计算数学]