虚拟化高性能计算集群上MPI应用性能分析.pdfVIP

 虚拟化高性能计算集群上MPI 应用性能分析 陆华俊, 任开军, 宋君强, 刘少伟 ( 国防科学技术大学 计算机学院,长沙 410073) * MPI applications performance analysis in HPC virtual clusters LU Hua-Jun, REN Kai-Jun, SONG Jun-Qiang, LIU Shao-Wei (School of Computer, National University of Defense Technology, Changsha 410073, China) + Corresponding author: E-mail: luhuajun@ Abstract : Due to the advantages of customizing operating environment, resource utilization and fault tolerance, virtualization are considered to have the potential solving the problem that high-performance computing(HPC) facing to. However, the major problem of applying virtualization to HPC is the additional performance overhead. In order to verify the feasibility of virtualization using in HPC application, we build the virtualizat ion environment in HPC cluster, and carry out a series comparative examination of issues relevant to MPI-style applications. Our results show that, NPB LU and SP application runs longer time in virtual cluster. IS, MG, CG and BT in 16 nodes runs faster by 40%-50% per node than in 32(36)nodes, while FT runs faster by 27%. Excessive communication overhead in applications will reduce the efficiency of the virtual cluster. Key words: virtualization; high performance computing; Xen; scheduling policy 摘 要: HPC 上搭建虚拟集群可以较好地实现资源复用且可定制，同时保证各个用户需求并不会产生资源冲 突。被认为可以解决高性能计算面临的诸多问题的潜质，同时其主要面临的问题是额外的性能开销。为了验 证虚拟化在高性能计算中应用的可行性，本文在高性能计算集群中构建了虚拟化系统环境，并针对MPI 类应 用进行了系列对比测试。实验结果显示，NPB 测试中，LU 和SP 在虚拟集群上运行时间更长，IS 、MG 、CG、 BT 在 16 节点的集群单节点运行速度比32(36)节点大约快40%-50%，FT 在 16 节点的集群单节点运行速度比 32 节点大约快27% ，任务具有较大通信需求时，会降低虚拟集群的执行效率。 关键词: 虚拟化;高性能计算;Xen; 调度策略 中图法分类号: TP393 文献标识码: A  Supported by the National Natural Science Foundation of China under Grant No.( 国家自然科学基金) 作者简介: 陆华俊(1989 －),男,广东湛江人,硕士研究生,主要研究领域为虚拟化、云计算和大数据处理;任开军( 1975 －),男,副