一种面向大规模并发的Gatherv优化方法  

A Gatherv optimization method for large scale concurrency

作　　者：孙浩男 王飞 魏迪 尹万旺 史俊达 SUN Hao-nan;WANG Fei;WEI Di;YIN Wan-wang;SHI Jun-da(National Research Center of Parallel Computer Engineering&Technology,Beijing 100080;Department of Computer Science and Technology,Tsinghua University,Beijing 100084,China)

机构地区：[1]国家并行计算机工程技术研究中心,北京100080 [2]清华大学计算机科学与技术系,北京100084

出　　处：《计算机工程与科学》2022年第9期1542-1549,共8页Computer Engineering & Science

基　　金：国家重点研发计划(2020YFB0204602)。

摘　　要：MPI不规则集合通信Gatherv为描述并行通信行为提供了极大的灵活性,但其不规则特性带来了较高的实现难度。现有方法存在通信热点突出、内存开销大和访存效率低等问题,难以满足当今大规模并行应用的性能需求。提出一种面向大规模并发的Gatherv优化方法,从优化等级、缓冲区管理等多个关键问题入手,将规则集合通信实现中常用的Binomial-Tree结构用于实现Gatherv,并提出消息链调度机制,进一步降低开销,提升优化效果。测试结果表明,该方法可以有效解决现有方法存在的性能问题,实现Gatherv集合通信性能在大规模并发条件下的高效可扩展。As an irregular MPI(Message Passing Interface)collective communication,Gatherv provides great flexibility for the description of parallel communication behavior,but its irregularity brings high implementation difficulties.Existing methods have some problems,such as outstanding communication hotspots,high memory overhead,low memory access efficiency,etc.,which are difficult to satisfy the performance requirements of today’s large-scale parallel applications.A Gatherv optimization method for large scale concurrency is proposed.Starting from the optimization level,buffer management and other key issues,the binomial tree model commonly used in the implementation of regular collective communication is applied to the implementation of Gatherv.Besides,a message chain scheduling is proposed to further reduce the overhead and improve the optimization effect.Test data shows that the proposed method can effectively solve the performance problems of the existing methods,and achieve efficient scalability of Gatherv performance under the condition of large-scale concurrency.

关 键 词：MPI 不规则集合通信 Gatherv Binomial-Tree 消息链调度 

分 类 号：TP302[自动化与计算机技术—计算机系统结构]