计算机工程与应用 ›› 2021, Vol. 57 ›› Issue (18): 82-85.DOI: 10.3778/j.issn.1002-8331.2008-0449

• 理论与研发 • 上一篇    下一篇

IB网上CPU-GPU异构超算平台容器性能评估及优化

胡鹤,赵毅,王宪贺   

  1. 中国科学院 计算机网络信息中心,北京 100190
  • 出版日期:2021-09-15 发布日期:2021-09-13

Evaluation and Optimization of Container Performance on Heterogeneous Supercomputing Platforms of CPU-GPU on IB Network

HU He, ZHAO Yi, WANG Xianhe   

  1. Computer Network Information Center, Chinese Academy of Sciences, Beijing 100190, China
  • Online:2021-09-15 Published:2021-09-13

摘要:

为了实现资源和系统环境的隔离,近年来新兴了多种虚拟化工具,容器便是其中之一。在超算资源上运行的问题通常是由软件配置引起的。容器的一个作用就是将依赖打包进轻量级可移植的环境中,这样可以提高超算应用程序的部署效率。为了解基于IB网的CPU-GPU异构超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。该方法能够评估容器在虚拟化宿主机过程中产生的性能开销,包括文件系统访问性能、并行通信性能及GPU计算性能。结果表明,容器具备近乎原生宿主机的性能,文件系统I/O开销及GPU计算开销与原生宿主机差别不大。随着网络负载的增大,容器的并行通信开销也相应增大。根据评估结果,提出了一种能够发挥超算平台容器性能的方法,为使用者有针对性地进行系统配置、合理设计应用程序提供依据。

关键词: 容器, 超算, 性能, 基准测试, 优化

Abstract:

The container technology has been paid much attention as an virtualization tool in recent years, which can isolate resources and system environments. Containers help supercomputing applications to package dependencies into a lightweight and portable environment, and improve the effectiveness of application deployment. Containers have performance overhead in the process of virtualizing. In order to understand the performance characteristics of container virtualization on heterogeneous supercomputing platforms of CPU-GPU on IB network, standard benchmarks are used to conduct a comprehensive performance evaluation, including I/O, parallel communication and GPU computing. The evaluation results show that, as for I/O and GPU, the container has tiny overhead. The parallel communication overhead of the container increases with the increase of network load. An optimization method is proposed to give best performance of the container, which provides a basis for users to design applications and perform system configuration.

Key words: container, supercomputing, performance, benchmark, optimization