高性能计算系统RDMA Read机制研究  

Research on RDMA Read Mechanism in High Performance Computing System

作　　者：盛腾飞 卢宏生[1] 曹志强[1] 王梦嘉[1] 斯添浩 SHENG Tengfei;LU Hongsheng;CAO Zhiqiang;WANG Mengjia;SI Tianhao(Jiangnan Institute of Computing Technology,Wuxi,Jiangsu 214083,China)

机构地区：[1]江南计算技术研究所,江苏无锡214083

出　　处：《计算机工程》2018年第10期69-79,共11页Computer Engineering

基　　金：国家高技术研究发展计划"E级超级计算机新型体系结构及关键技术研究"(2015AA01A03);核高基专项"超级计算机处理器研发"(2013ZX0102-8001-001-001)

摘　　要：R&A和R2W两种远程直接存储读访问(RDMA Read)消息实现机制在设计复杂度、设计开销、通信性能方面均存在不足。为此,分别从3个方面对现有2种机制进行对比分析。在设计复杂度方面,针对2种机制设计基本的硬件实现流程并做定性分析,在设计开销、通信性能方面,采用基于性能分析模型的方法进行理论分析并设计软件验证平台实现实验验证。根据分析结果结合2种机制的方法和特点提出一种改进的RDMA Read实现机制。实验结果表明,相对R&A和R2W机制,该改进RDMA Read机制在设计复杂度、设计开销、通信性能方面均有较好且均衡的表现。R&A and R2W Remote Direct Memory Access Read(RDMA Read)message implementation mechanisms have some shortcomings in design complexity,design overhead and communication performance.To solve this problem,the three existing mechanisms are compared and analyzed from two aspects.In terms of design complexity,the basic hardware implementation flow is designed for the two mechanisms and qualitative analysis is made.In terms of design overhead and communication performance,performance analysis model based method is used to analyze the theory and software verification platform is designed to achieve experimental verification.An improved RDMA Read implementation mechanism is proposed according to the analysis results and the methods and characteristics of the two mechanisms.Experimental results show that compared with R&A and R2W,the improved RDMA Read mechanism performs well and equally in terms of design complexity,design overhead and communication performance.

关 键 词：远程直接存储读访问 消息机制 性能分析模型 设计复杂度 设计开销 通信性能 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]