Raft存储集群中的日志分发机制优化  

Optimization of log distribution mechanisms in Raft storage clusters

作　　者：徐辉 高辉 Xu Hui;Gao Hui(School of Computer Science&Engineering,University of Electronic Science&Technology of China,Chengdu 611731,China;Kash Institute of Electronics&Information Industry,Kashgar Xinjiang 844000,China)

机构地区：[1]电子科技大学计算机科学与工程学院(网络空间安全学院),成都611731 [2]喀什地区电子信息产业技术研究院,新疆喀什844000

出　　处：《计算机应用研究》2024年第12期3755-3762,共8页Application Research of Computers

基　　金：四川省科技计划资助项目(2023YFG0021)。

摘　　要：在分布式存储系统中,Raft(replicated and fault tolerant)算法的强领导特性在节点数量增多时会带来巨大的日志分发开销,限制了系统性能和水平扩展能力。针对系统性能和扩展性瓶颈,提出了两种新的日志机制来优化一致性哈希集群分布式存储方案。第一种是基于动态优先级的日志分发机制,日志分发顺序由领导者与跟随者节点日志的同步程度决定,加快了日志项的提交速度;第二种是基于窗口流水线的日志分发机制,领导者节点指派日志同步程度较高的跟随者节点对同步程度较低的跟随者节点进行日志分发,缩短了系统中节点日志趋向一致的时间。相比于未优化方法,吞吐量和日志同步时间在多节点集群上有显著提升,证明了两种日志机制在改进系统性能上的有效性。In distributed storage systems,the Raft algorithm’s strong leadership characteristics introduce substantial log distribution overhead as the number of nodes increases,impacting performance and scalability.This paper proposed two innovative mechanisms to overcome these challenges.First algorithm was a dynamic priority based log distribution mechanism.It assigned log entries based on the synchronization levels between leader and follower nodes,thereby expediting log submission.Second algorithm was a window pipeline based log distribution delegation mechanism.Its leader node assigned more synchronized followers to disseminate logs to less synchronized ones,diminishing the time required to achieve system-wide log consistency.The experimental results demonstrate substantial enhancements in throughput and log synchronization time within multi-node clusters.This improvement shows the efficacy of these mechanisms in bolstering system perfor-mance.

关 键 词：分布式存储 一致性哈希 日志分发 动态优先级分配 窗口流水线 

分 类 号：TP311.13[自动化与计算机技术—计算机软件与理论]