基于Intel Xeon Phi的激光等离子体粒子模拟研究  

Research of LARED-P on Intel Xeon Phi

作　　者：姚文科 杜云飞[2] 吴强[1] 杨灿群[1] 

机构地区：[1]国防科学技术大学并行与分布处理国家重点实验室,湖南长沙410073 [2]国防科学技术大学计算机学院,湖南长沙410073

出　　处：《计算机工程与科学》2014年第5期809-813,共5页Computer Engineering & Science

基　　金：国家863计划资助项目(2012AA010903);国家自然科学基金资助项目(61170049)

摘　　要：激光等离子体粒子模拟广泛用于探索极端物质状态下的科学问题。将一种基于粒子云网格方法的三维等离子体粒子模拟程序LARED-P移植到Intel Xeon Phi协处理器上。在移植的过程中,综合运用了Native和Offload两种编程模式:首先运用Native模式对LARED-P程序中热点计算任务进行优化研究,通过采用SIMD扩展指令使该计算任务获得了4.61倍的加速;然后运用Offload模式将程序移植到CPU-Intel Xeon Phi异构系统上,并通过使用异步数据传输和双缓冲技术分别提升了程序性能9.8%和21.8%。Plasma simulations have been situations. The paper ports a particle-in-cell widely used to exploit scientific problems under extremely based plasma code, LARED-P, to Intel Xeon Phi co-proces- sor. In order to accomplish this, two modes are employed, i.e. the Native mode and the Offload mode. Firstly, the Native mode is employed to study on the hot computing tasks, which have been accelerated to 4.61 times faster by using SIMD extension instructions. Secondly, the Offload mode is employed to transplant the whole code onto a CPU-Intel Xeon Phi heterogeneous system, We also adopt optimiza- tions such as asynchronous data transferring and double buffer technique to improve the performance. And we obtain 9.8% and 21.8% improvement respectively.

关 键 词：LARED-P INTEL XEON PHI Native模式 Offload模式 512位SIMD扩展指令 异步数据传输 双缓冲 

分 类 号：TP311.56[自动化与计算机技术—计算机软件与理论]