网络编码集群存储系统基于锁机制的热数据重建策略  

Lock-based Hot Data First Recovery Strategy for Network-coding-based Distributed Storage Systems

作　　者：赵昱帆 邓玉辉[1,2] ZHAO Yu-fan;DENG Yu-hui(Department of Computer Science,Jinan University,Guangzhou 510632,China;State Key Laboratory of Computer Architecture,Institute of Computing Technology,Chinese Academy of Sciences,Beijing 100190,China)

机构地区：[1]暨南大学信息学院计算机科学系,广州510632 [2]中国科学院计算技术研究所计算机体系结构国家重点实验室,北京100190

出　　处：《小型微型计算机系统》2018年第8期1707-1712,共6页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61572232)资助;广州市南沙区科技计划项目(2017CX006)资助;中国科学院计算机系统结构国家重点实验室开放基金项目(CARCH201705)资助

摘　　要：在数据规模与日俱增的背景下,网络编码成为了替代传统三副本机制的一种数据冗余机制.网络编码优势在于减少了存储成本,劣势在于节点失效时增加了系统重建带宽和响应延迟.基于热数据优先的网络编码分布式存储系统重建策略(HDFR)根据用户I/O负载优先重建热数据,使用户可以提前访问替换节点中已重建的失效热数据,这一定程度地使数据重建性能和系统响应性能得到了优化.该策略中,替换节点同时服务于重建数据流的同步写操作和用户访问数据流的读操作,这两种操作在替换节点中交替进行.相对于传统的重建策略(TR)中替换节点只服务于重建数据流的写操作,这种策略使替换节点中的I/O等待时间变长,使重建性能一定程度上削弱了.系统发生节点失效时,缩小重建窗口对于保障系统数据可靠性和可用性很有必要.本文基于上述热数据重建策略,提出了网络编码集群存储系统中基于锁机制的热数据重建策略(LHDFR),利用锁机制将写操作和读操作分割开来,使重建数据流的写操作具有更高优先级,而由此阻塞的读操作则由解码操作来代替.实验表明,相较于HDFR,LHDFR将重建性能优化了59.4%,用户访问性能优化了36.2%.With the explosive growth of data, network coding has become an alternative redundancy mechanism to three replica mecha- nism. The advantage of network coding is that the storage cost is reduced, and the disadvantage is that it increases the reconstruction bandwidth and response delay when nodes failed. Some propose a hot data first reconstruction method （ HDFR ） based on the character of I/O workload. The method optimizes the reconstruction and response performances to some extent. In HDFR, the replaced nodes serve both the writing operations of reconstruction stream and the reading operations of user accesses. Thus the read operations alter- nate with the write operations in the replaced nodes. Compare to the situation in traditional reconstruction { TR ） that replaced nodes only serve synchronous write operations of reconstruction stream, HDFR makes the ready queue of I/O grew longer, thus the reconstruction performance is weakened. When nodes failure happened, reducing the window of reconstruction has a significant impact on the system reliability and availability. This paper propose locke-based hot data first reconstruction strategy { LHDFR ） based on HDFR through bal- ancing the write operations and the read operations. Our method uses lock mechanism to separate the read and write operations in re- placed nodes. Write operations have priority over read operations to access the replaced nodes and the rest read operations are replaced by decoding. Our experiment shows that,compared with HDFR, LHDFR reduces the response time and the reconstruction time up to 59.4% and 36.2% ,respectively.

关 键 词：网络编码 分布式存储系统 AOE网络通信协议 热数据重建 数据流争用 

分 类 号：TP303[自动化与计算机技术—计算机系统结构]