分布式存储系统中节点修复问题研究

发布时间：2018-01-27 02:30

本文关键词： 网络编码分布式存储节点修复网络信息流图数据容错　出处：《西南交通大学》2012年硕士论文　论文类型：学位论文

【摘要】：图灵奖获得者Jim Gray在其获奖演说时,对全球数据量的增长给出了一条新的经验定律：未来每18个月产生的数据量等于有史以来的数据量之和!近些年,IT行业的飞速发展与互联网的广泛应用,带来了全球信息资源的爆炸性增长,各种应用对存储系统提出了越来越高的要求。分布式存储系统,因其廉价性及高扩展性等优点,而倍受人们关注,理所应当地成为了海量数据存储的首要选择。然而,由于分布式存储系统中各个存储节点的可用性不高,为保证数据可靠性,系统会频繁的进行节点修复。因此,如何有效地进行节点修复就成为了亟待解决的问题,对其进行的深入的研究,具有很重要的现实意义。根据网络编码定理,分布式存储系统中的节点修复问题可以抽象成为一个基于网络信息流图的数据传输模型,从而可以利用图论中网络流的相关理论来分析节点修复时带宽消耗的理论下界。之前的大部分研究主要针对的是单节点修复的情形,然而,实际中多节点同时修复的情形非常常见。现有的针对多节点同时修复的模型都是非对称的,这会增加实际系统设计的复杂度。针对此问题,本文在现有模型的基础上,提出了一种对称的多节点协作修复模型(Symmetric Mutually Cooperative Recovery, SMCR),使得每个待修复的新节点之间可以进行等量的数据交换,并且计算出了模型在不同冗余度下节点修复带宽的理论下界。最后,本文将SMCR修复模型与不协作的节点修复模型进行比较。经过数值分析,发现在节点数据冗余度较低时,SMCR的修复带宽的理论下界好于不协作的修复模型。在数据冗余度比较高时,SMCR的修复带宽的理论下界与不协作的各有优劣,不同的参数会出现不同的结果。随后,本文对其进行了深入的分析,给出了一止匕SMCR修复模型好于不协作的节点修复模型的充分条件,为实际分布式存储系统的设计提供了的理论指导。
[Abstract]:In his speech, Turing Prize winner Jim Gray presented a new empirical law on the growth of global data volumes: the amount of data generated every 18 months over the next 18 months is equal to the sum of data in history! In recent years, the rapid development of IT industry and the wide application of the Internet have brought the explosive growth of global information resources. Because of its advantages such as low cost and high scalability, it has attracted much attention and become the first choice of mass data storage. However, the availability of each storage node in distributed storage system is not high. In order to ensure the reliability of the data, the system will frequently carry out node repair. Therefore, how to effectively repair the node has become a problem to be solved. Has very important realistic significance. According to the network coding theorem, the node repair problem in distributed storage system can be abstracted into a data transmission model based on network information flow graph. Therefore, we can use the theory of network flow in graph theory to analyze the theoretical lower bound of bandwidth consumption in node repair. Most of the previous studies focus on the case of single node repair, however. In practice, multi-node repair is very common. Existing models for multi-node simultaneous repair are asymmetric, which will increase the complexity of the actual system design. Based on the existing models, a symmetric Mutually Cooperative Recovery, a symmetric multi-node cooperative repair model, is proposed in this paper. The SMCRO makes it possible to exchange the same amount of data between each new node to be repaired and calculates the theoretical lower bound of the repair bandwidth of the model under different redundancy. Finally, this paper compares the SMCR repair model with the non-cooperative node repair model. Through numerical analysis, it is found that the node data redundancy is low. The theoretical lower bound of repair bandwidth of SMCR is better than that of non-cooperative repair model. When the data redundancy is high, the theoretical lower bound and non-cooperation of SMCR have their own advantages and disadvantages. Different parameters will produce different results. Then, this paper gives a thorough analysis of the SMCR repair model is better than the non-cooperative node repair model sufficient conditions. It provides theoretical guidance for the design of distributed storage system.
【学位授予单位】：西南交通大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP333

【共引文献】