网络分簇BWRAID:更快的扩展、恢复和读写性能

发布时间：2018-07-21 19:51

【摘要】：存储区域网(storage area network,SAN)是重要的网络存储方法.使用商用硬件BWRAID在SAN上实现了分布式RAID.初始版本的BWRAID使用全对称结构,然而其存在3个问题:1)扩展时要读取数据重新计算校验,IO负载高、扩展时间长;2)将数据集中恢复到单个存储节点,没有分布的并发恢复;3)数据布局不合理,导致内部RAID4有大量同步更新.为解决上述问题,提出了"网络分簇BWRAID".新系统采用"分簇RAID"(declustering RAID)的非对称结构,分簇对象是相等大小的小虚拟盘而不是数据块;在扩展时,它在节点之间仅迁移虚拟卷,不需计算校验.由于一个恢复需要的节点数量小于节点总数,多个恢复就能并行.为优化IO使用新的数据布局,按内部RAID4条带组织用户的存储空间,并给出了搜索虚拟盘的算法,用于在系统分配、扩展、恢复时,搜索合适的虚拟盘.实验表明网络分簇BWRAID更好:在系统扩展时无需重新计算校验,加速扩展5~8倍;并行恢复成倍加速;新数据布局提高了IO性能.
[Abstract]:Storage area Network (storage area Network San) is an important network storage method. The distributed raid is implemented on San using commercial hardware BWRAID. The original version of BWRAID uses a fully symmetric structure, but it has three problems: 1) reading data, recalculating, checking IO, high load, long extension time, and restoring the data set to a single storage node. No distributed concurrent recovery 3) unreasonable data layout, resulting in a large number of internal RAID4 synchronous updates. In order to solve the above problems, a network clustering BWRAID is proposed. The new system adopts the asymmetric structure of "declustering raid". The clustering object is a small virtual disk of equal size rather than a data block. In the extension, it only migrates the virtual volume between nodes and does not need to calculate the verification. Since the number of nodes needed for one recovery is less than the total number of nodes, multiple recovery can be parallel. In order to optimize IO's new data layout, the storage space of users is organized according to the internal RAID4 strip, and the algorithm of searching virtual disk is given, which is used to search suitable virtual disk when the system distributes, expands and recovers. The experimental results show that the network clustering BWRAID is better: the expansion is accelerated by 5 times without the need to recalculate when the system is extended, the parallel recovery is doubled and the performance of IO is improved by the new data layout.
【作者单位】：中国科学院计算技术研究所;中科蓝鲸信息技术有限公司;中国科学院大学;
【基金】：国家“八六三”高技术研究发展计划基金项目(2013AA013205) 国家“九七三”重点基础研究发展计划基金项目(2011CB302304) 国家科技支撑计划基金项目(2011BAH04B04) 中国科学院战略性先导科技专项基金项目(XDA06010401);中国科学院重点部署基金项目(KGZD-EW-103-5(7))
【分类号】：TP333

【相似文献】