基于KNL平台的LAMMPS优化

LAMMPS Optimization on KNL Platform

Intel新一代处理器KNL作为一种具有极强运算能力的多核处理器,拥有16GB高速片上内存(MCDRAM),物理核心数量高达72个,单CPU的双精度浮点峰值为3TFlops,为高并行负载应用提供强大的性能支持。各种主流的并行软件也纷纷使用KNL众核、高速内存技术,由于LAMMPS(large-scale atomic/molecular massively parallel simulator)在材料科学和计算化学中的广泛应用,因此在KNL节点上优化LAMMPS成为相关领域近些年的研究热点。本文以郑州超算中心的KNL集群为平台,采用MCDRAM和第三方扩展包两种方法对LAMMPS进行优化。MCDRAM可以加快CPU读取数据的速度,第三方扩展包从源码的角度对程序中的条件判断进行优化。试验结果表明,优化后的LAMMPS执行时间明显减少,加速比可达49x,是CPU平台加速比的5.5x。

Intel＇s new generation processors KNL（Knights Landing） bring in new memory technology, a high bandwidth on package memory and low capacity（up to 16 GB） called Multi-Channel DRAM. Additionally, With up to 72 cores, the KNL delivers over 3 TFlops of double precision peak, which provides powerful performance support for high parallel load applications. A variety of mainstream parallel software adopted Mangy-core and MCDRAM technology on KNL, LAMMPS（large-scale atomic/molecular massively parallel simulator） plays an important role in materials sciences and computational chemistry, so LAMMPS Optimization on KNL has become a research interest to many research scientists of related fields inrecent years. In this paper, we employ two methods to optimize the LAMMPS based on the KNL cluster of Zhengzhou Supercomputing Center. First, the MCDRAM is employed to make data form memory to CPU quickly, and second,we use third party packages to accelerate LAMMPS. Experimental results demonstrate that the optimized LAMMPS execution time is significantly reduced and speedup reaches 49 x while that is 5.4 x on the CPU platform.