VLPL-S在Knights Landing上的优化与性能评估  

Optimization and Evaluation of VLPL-S on Knights Landing

作　　者：丁丹迪 文敏华 周姗 陈民[3] 林新华 

机构地区：[1]上海交通大学高性能计算中心,上海200240 [2]英特尔亚太研发中心,上海200240 [3]上海交通大学物理与天文系,上海200240 [4]东京工业大学

出　　处：《计算机科学与探索》2018年第2期241-251,共11页Journal of Frontiers of Computer Science and Technology

基　　金：国家重点研发计划Nos.2016YFB0201400;2016YFB0201800;日本学术振兴会RONPAKU项目~~

摘　　要：VLPL-S代码是基于Particle-in-Cell(PIC)算法开发的激光等离子体模拟程序,PIC算法是激光等离子模拟领域的常用主流算法之一。讲述了VLPL-S代码在Intel?新推出的Knights Landing平台上的早期移植及优化工作。通过采用在代码优化中常用的优化方法,例如访存优化、多线程优化、向量化,为VLPL-S代码实现了1.68倍的加速比。对于优化以后的VLPL-S代码,其在Knights Landing 7210P单节点上的性能是其在双路Xeon E5-2697v4节点上性能的1.53倍。还对比了不同优化方法在Knights Landing及Xeon平台上所获得的性能提升。结果表明,对于VLPL-S代码,以往CPU代码优化工作中常用的优化方法在新的Knights Landing平台中同样有效。VLPL-S code is developed based on the particle-in-cell(PIC) algorithm, which is the mainstream algorithm of laser plasma behavior research. This paper reports the early experience on porting and optimizing the VLPL-S PIC code on the Knights Landing. By applying general optimization methods such as memory access optimization,thread level parallelism and vectorization to the code, this paper achieves 1.68 times speedup compared with the original code. After optimization, 1.53 times speedup is achieved on Knights Landing 7210 P compared with that on a two-socket Xeon E5-2697 v4 node. This paper compares the performance improvement of different optimization methods on Knights Landing and the Xeon-based node. The results show that most of the commonly used optimization methods are effective for the VLPL-S code on Knights Landing.

关 键 词：激光等离子体模拟 PARTICLE-IN-CELL Knights LANDING 

分 类 号：TP39[自动化与计算机技术—计算机应用技术]