ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2019, Vol. 56 ›› Issue (4): 767-778.doi: 10.7544/issn1000-1239.2019.20170834

• 系统结构 • 上一篇    下一篇



  1. (武汉光电国家研究中心(华中科技大学) 武汉 430074) (信息存储系统教育部重点实验室(华中科技大学) 武汉 430074) (深圳华中科技大学研究院 广东深圳 518000) (
  • 出版日期: 2019-04-01
  • 基金资助: 

An Efficient Failure Reconstruction Based on In-Network Computing for Erasure-Coded Storage Systems

Tang Yingjie, Wang Fang, Xie Yanwen   

  1. (Wuhan National Laboratory for Optoelectronics (Huazhong University of Science and Technology), Wuhan 430074) (Key Laboratory of Information Storage System (Huazhong University of Science and Technology), Ministry of Education, Wuhan 430074) (Shenzhen Huazhong University of Science and Technology Research Institute, Shenzhen, Guangdong 518000)
  • Online: 2019-04-01

摘要: 目前分布式存储系统的规模越来越大,不论存储设备是磁盘还是固态盘,系统都始终面临着数据丢失的风险.传统分布式存储系统大多采用基于三副本的高可靠性技术,但为了追求较低的存储开销,大量系统正在转向基于纠删码的可靠性方法.但是在纠删码方案下,重建故障数据需要读取多个存储设备,这将导致大量的网络传输和存储I/O操作,增大系统恢复开销.为了能够在不损失其他性能的同时降低恢复开销,利用软件定义网络(software defined networking, SDN)技术,提出一种基于网络计算的高效故障重建方案——网络流水线(in-network pipeline, INP),其中SDN 控制器利用网络的全局拓扑信息构造重建树,系统依据重建树进行数据传输,并在交换机上完成部分计算,减少向后传输的网络流量,从而消除网络瓶颈,提升恢复性能.测试评估了不同网络带宽下INP的恢复效率.实验结果表明:与传统的纠删码系统相比,INP总是能大幅减少网络流量,并且在一定带宽条件下,能够接近正常读的时间开销.

关键词: 分布式存储系统, 纠删码, 软件定义网络, 恢复开销, 网络计算

Abstract: Nowadays, the scale of distributed storage systems is getting increasingly larger. No matter whether the storage devices are disks or solid-state drives, the system is always faced with the risk of data loss. Traditional storage systems maintain three copies of each data block to ensure high reliability. Today, a number of distributed storage systems are increasingly shifting to the use of erasure codes because they can offer higher reliability and lower storage overhead. The erasure codes, however, have an obvious shortcoming in the reconstruction of an unavailable block, because they need to read multiple disks, which results in a large amount of network traffic and disk operations and ultimately high recovery overhead. In this paper, INP (in-network pipeline), an effective failure reconstruction scheme based on in-network computing that utilizes SDN (software defined networking) technology is presented in order to reduce the overhead of recovery without sacrificing any other performance. We use the global topology information for network from SDN controller to establish the tree of reconstruction, and transmit data according to it. The switches do part of the calculation that can reduce the network traffic, therefore to eliminate the bottleneck of the network, and to enhance the recovery performance. We evaluate the recovery efficiency of INP in different network bandwidths. Compared with the common erasure code system, it greatly reduces the network traffic and in a certain bandwidth, the degraded read time is the same as that of normal reading.

Key words: distributed storage system, erasure code, software defined networking (SDN), recovery overhead, in-network computing