Active Storage Mechanism for Cluster-Wide RAID System
Active Storage Mechanism for Cluster-Wide RAID System
复制标题
DOI：
10.1109/dsdis.2015.101
复制
发表时间：
2015-12
期刊：
影响因子：
--
通讯作者：
Hiroki Ohtsuji;O. Tatebe
中科院分区：
文献类型：
--
作者：
Hiroki Ohtsuji;O. Tatebe
可靠、高性能的网络I/O机制是百亿亿级存储系统的关键组成部分。为了提高系统的可靠性，一些存储系统采用了复制和擦除编码。但是，这是一个额外的成本，并且需要更高的带宽。提出了一种采用主动存储机制实现零开销网络RAID-5的新方法。本文提出的方法主要针对全集群（基于网络的）RAID系统的写操作，并可推广到恢复过程。通常情况下，使用擦除编码的存储系统由于需要额外的编码计算和传输的数据量增加而降低了写吞吐量。我们的建议解决了这个问题，并使我们能够通过避免客户端计算成本和网络流量来实现零开销的集群范围的RAID系统。此外，我们提出的方法使系统能够分离条纹数据块（源）的流量和奇偶生成流量。通过为奇偶校验生成过程提供专用网络，可以大大提高集群范围内RAID的可伸缩性。此外，我们提出了一种有效的实现方法，利用InfiniBand远程直接内存访问（RDMA）机制，以尽量减少内存复制操作的数量。该实现使用零复制管道构建，使我们能够使用低开销的节点间数据处理机制。与原始方法相比，测量的吞吐量增加了32.6%。结果显示出与RAID-0相同的性能，这意味着我们提出的方法实现了零开销的集群范围的RAID。
Reliable and high-performance network I/O mechanism is a critical part of exascale storage systems. In order to improve system reliability, some storage systems use replication and erasure coding. However, this is an additional cost and requires higher bandwidth. This paper proposes a novel method to implement a zero-overhead network RAID-5 with an active storage mechanism. The proposed method focuses on the write operation of the cluster-wide (network based) RAID system, and it can be generalized to the recovery process. In general, write throughput of the storage system with erasure coding is degraded because of the additional coding computation and the increased amount of transferred data. Our proposal solves this problem and enables us to implement a zero-overhead cluster-wide RAID system by avoiding client-side computation cost and network traffic. Furthermore, our proposed method enables the system to separate the traffic of the striped data blocks (source) and the parity generation traffic. Scalability of cluster-wide RAID can be highly increased with a dedicated network for the parity generation process. In addition, we propose an efficient implementation method that utilizes an InfiniBand remote direct memory access (RDMA) mechanism to minimize the number of memory copy operations. This implementation is built with zero-copy pipelines, and enables us to use a low-overhead inter-node data processing mechanism. The measured throughput gain was 32.6% compared to that of the naive method. The results show the same performance as RAID-0, which means that our proposed method achieves the zero-overhead cluster-wide RAID.