• 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

• 高性能计算 • 上一篇    下一篇

面向数据流结构的指令映射优化方法

李易1,常成娟1,卢圣健1,江道忠2,范东睿1,叶笑春1   

  1. (1.中国科学院计算技术研究所,北京 100094;2.北京石油化工学院,北京 102617)
  • 收稿日期:2018-08-20 修回日期:2018-10-16 出版日期:2019-01-25 发布日期:2019-01-25
  • 基金资助:

    国家重点研发计划(2018YFB1003500);国家自然科学基金重点项目(61332009)

An instruction mapping optimization method
based on dataflow architecture

LI Yi1,CHANG Chengjuan1,LU Shengjian1,JIANG Daozhong2,FAN Dongrui1,YE Xiaochun1
  

  1. (1.Institute of Computing Technology,Chinese Academy of Sciences,Beijing 100094;
    2.Beijing Institute of Petrochemical Technology,Beijing 102617,China)
  • Received:2018-08-20 Revised:2018-10-16 Online:2019-01-25 Published:2019-01-25

摘要:

在高性能计算领域,数据流是一类重要的计算结构,也在很多实际场景表现出很好的性能和适用性。在数据流计算模式中,程序是以数据流图来表示的,数据流计算中一个关键的问题是如何将数据流图映射到多个执行单元上。通过分析现有数据流结构的指令映射方法及其不足,提出了基于数据流结构的新型指令映射优化方法。
主要是根据多地址共享数据包的特性对指令映射方法进行优化,延迟多地址共享数据路由包的拆分,减少网络拥堵。
 

关键词: 高性能计算, 数据流计算, 指令映射, 多地址共享数据包

Abstract:

In the field of high-performance computing, dataflow computation is an important computing architecture and shows good performance and applicability in many practical scenarios. In the dataflow computation mode, the program is represented by a dataflow diagram. A key problem of dataflow computation is how to map dataflow diagrams to multiple execution units. By analyzing the instruction mapping methods of current dataflow architecture and their shortcomings, we propose a new instruction mapping optimization method based on dataflow architecture, and achieve code implementation and verification. We optimize the instruction mapping method according to the characteristics of multi-address shared data packets, delay the splitting of multiaddress shared data packets, and reduce network congestion.
Key words:

Key words: high performance computing, dataflow computation, instruction mapping, multi-address data shared packet