基于RDMA的分布式键值存储系统性能优化  

Latency Optimization of Distributed Key-value Store Based on RDMA

作　　者：王哲 王玉玫[1] 吴亚非[1] 臧义华[1] WANG Zhe;WANG Yu-mei;WU Ya-fei;ZANG Yi-hua(North China Institute of Computing Technology,Beijing 100083,China)

机构地区：[1]华北计算技术研究所,北京100083

出　　处：《计算机与现代化》2023年第2期24-27,共4页Computer and Modernization

摘　　要：如今分布式键值存储系统在工业界得到了广泛的应用,如何提升其性能成为分布式系统和并行计算领域的研究热点。远程直接数据存取是高性能计算中常用的通信技术,利用其低延迟高带宽的特性,可以有效提升分布式键值存储系统的性能表现。MPI单边通信是一种支持RDMA的通信方法,本文使用MPI单边通信实现分布式键值操作,并根据其特点重新设计通信模型。为了解决RMA带来的数据一致性问题,本文改变键值对的存储结构并制定存取操作的通信策略。通过将单边通信与TCP/IP进行对比,对不同长度的数据测试通信延迟,验证了RDMA给系统带来的性能提升。Distributed key-value store is widely used in industry,and methods to improve the performance of data transfer in dis⁃tributed key-value store has drawn great interests in the study of distributed systems and parallel computing.RDMA(Remote Memory Direct Access)is a popular network technology in high-performance computing.Making use of its low latency and high bandwidth features can lead to enormous improvement of the performance of distributed key-value store.MPI offers routines to perform remote memory access(RMA),also known as one-sided communication,which can benefit from RDMA technology.In this paper,we propose leverageing MPI one-sided communication refactors key-value operations and redesigns the communica⁃tion model according to the features of one-sided communication.To solve the data inconsistency problem brought by RMA,we modify the storage structure of key-value pair and schedule the communication strategy.We evaluate the improvement by compar⁃ing the latency of RMA based key-value operations and TCP/IP based key-value operations when transferring key-value pairs of different lengths and get positive result.

关 键 词：分布式键值存储系统 远程内存直接访问 单边通信 

分 类 号：TP333[自动化与计算机技术—计算机系统结构]