GPU加速高性能计算平台上容器性能评估  

Performance characterization of container-based virtualization for GPU-accelerated HPC

作　　者：胡鹤 赵毅[1] 庞飞 HU He;ZHAO Yi;PANG Fei(Computer Network Information Center,Chinese Academy of Sciences,Beijing 100190,China)

机构地区：[1]中科院计算机网络信息中心,北京100190

出　　处：《云南民族大学学报（自然科学版）》2021年第1期58-62,共5页Journal of Yunnan Minzu University:Natural Sciences Edition

基　　金：中国科学院战略性先导科技专项(XDC01000000).

摘　　要：容器是近年来新兴的虚拟化工具,可以实现资源和系统环境的隔离.容器能够帮助高性能计算应用程序将依赖打包进轻量级可移植的环境中,解决因软件配置无法在高性能计算资源上运行的问题.容器在虚拟化宿主机过程中具有性能开销,为了解GPU加速高性能计算平台上容器虚拟化技术的性能特征,使用标准基准测试工具对Docker容器进行了全面的性能评估,包括文件系统访问性能,并行通信性能及GPU计算性能.评估结果表明,在文件系统I/O开销及GPU计算开销方面,容器具备近乎原生宿主机的性能,容器的并行通信开销随着网络负载的增大而增大.在仅考虑性能的情况下,容器方案适用于通信负载不大的并行应用程序.The container is used as a new virtualization tool in recent years,which can isolate resources from system environments.Containers can help high-performance computing applications to package their dependencies into a lightweight and portable environment,and solve the problem of incapable of running on high-performance computing resources because of software configuration.Containers have performance overhead in the process of virtualizing the host.In order to understand the performance characteristics of container virtualization technology on GPU-accelerated high-performance computing platforms,this article uses standard benchmarking tools to conduct a comprehensive performance evaluation of the Docker container,including file system I/O performance,parallel communication performance and GPU computing performance.The evaluation results show that,in terms of file access overhead and GPU computing overhead,the container has nearly native performance,and the parallel communication overhead of the container increases with the increase of network load.When only considering performance,this solution is suitable for parallel applications with low communication load.

关 键 词：虚拟化 容器 基准测试 性能开销 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]