• 中国计算机学会会刊
  • 中国科技核心期刊
  • 中文核心期刊

计算机工程与科学

• 论文 • 上一篇    下一篇

异构高性能计算系统Linpack效率受限因素分析

贾迅,邬贵明,谢向辉   

  1. (数学工程与先进计算国家重点实验室,江苏 无锡 214125)
  • 收稿日期:2017-09-01 修回日期:2017-11-03 出版日期:2018-02-25 发布日期:2018-02-25
  • 基金资助:

    国家自然科学基金(91430214)

Analysis of the factors limiting Linpack efficiency of
heterogeneous high-performance computing systems
 

JIA Xun,WU Gui-ming,XIE Xiang-hui   

  1. (State Key Laboratory of Mathematical Engineering and Advanced Computing,Wuxi 214125,China)
  • Received:2017-09-01 Revised:2017-11-03 Online:2018-02-25 Published:2018-02-25

摘要:

能耗是目前高性能计算系统性能提升的一大挑战。主处理器连接加速器的异构计算技术可以有效提升系统能效,因而被广泛应用于当前高性能计算系统的设计。同等系统规模下,异构计算系统的Linpack效率普遍低于同构系统。针对这一问题,从结构设计的角度,基于真实计算系统的设计参数和性能数据,分析了大规模异构高性能计算系统Linpack效率受限的主要因素及其对结构设计的需求,并构建了针对异构计算系统的Linpack性能模型对分析结论进行了验证。研究成果对异构计算系统Linpack的性能优化以及未来高效异构架构的设计具有一定的指导意义。

关键词: 异构, 高性能, Linpack, 效率

Abstract:

Energy consumption is a major challenge for current high-performance computing (HPC) system design. The technique of heterogeneous computing with accelerators connected to host processor can improve the system energy efficiency, and has been widely applied in HPC. However, heterogeneous systems have lower Linpack efficiency in comparison to homogeneous systems with the equivalent scale. Aiming at this problem, from the perspective of structural design, based on the design parameters and performance data of real computing systems, this paper analyzes the main factors limiting the Linpack efficiency of heterogeneous HPC systems and their demands of structural design. Besides, the analysis is verified with a Linpack performance model tailored for heterogeneous systems. This study results are of guiding significance to Linpack performance optimization and efficient architecting on future heterogeneous systems.
 

Key words: heterogeneous, high performance, Linpack, efficiency