基于异构平台的BH算法高效并行实现

High efficient parallel implementation of BH algorithm on heterogeneous platforms

针对多核CPU和众核加速器或协处理器异构平台的架构特征进行了研究,以MPI和Open MP混合编程模型实现了N体问题BH算法的并行,采用了正交递归二分法(ORB)使进程之间负载均衡,并对程序进行了并行优化和MIC加速。优化和加速后的程序性能提升到原版本的3.4倍以上,其中MIC加速后性能提升到加速前的1.7倍;程序具有较好的扩展性,计算粒子规模达到上亿时,可扩展到32个节点共4 480核心(640个CPU核心和3 840个MIC核心)。

Studying the architecture＇ s characteristics of the multi-core CPU and accelerators or coprocessors heterogeneous platforms, this paper was about the parallel implementation of N-body BH algorithm with hybrid MPI and OpenMP programming model. It used orthogonal recursive bisection （ORB） to balance load between processors, then carefully optimized the code on multi-core CPU and accelerated it on MIC. Testing result shows, after optimizing and accelerating, the code＇ s performance rea- ches above 3.4x speedup than original version and gets a 1.7x speedup than only running on muhi-core CPU. The code also has a good scalability with a 100 million particles running on a 32 nodes cluster, which has 4 480 cores （640 CPU cores and 3 840 MIC cores）.