The Research On Node Repair Problem Of Distributed Storage System

Jim Gray, a winner of ACM Turing Award, put forward a new empirical law for worldwide data growth in his acceptance speech that in the future, data produced in every18month amount to the sum of that in the past. In recent years, explosive growth of the worldwide information has resulted from the rapid development of IT industry and widespread deployment of internet, thus many requirements from all kinds of applications have imposed to storage system. With some advantages like low price and high scalability, distributed Storage Systems are favorable among a large amount of data storage systems. However, low availability of storage nodes lead to the frequent node repair to guarantee the reliability of data. Therefore, it is urgent to deal with effective node repair and the related research is of great interest from the viewpoint of the practice.Based on network information flow graph, node repair in distribution storage system can be regarded as a data transmission mode in an abstract manner by applying network coding principles, and then a theoretic lower bound on bandwidth consumption for node repair can be derived. Most of the studies in the literature focus on single-failure recovery. However, there are many scenarios where multi-failures must be considered. The existing models foucs on multi-failure are all non-symmetric, this will increase the complexity of the actual system design. Aiming at the repair of multiple nodes, this thesis puts forward a model named Sym-metric Mutually Cooperative Recovery (SMCR) such that the new nodes need can exchange data equally. As for SMCR model, a theoretic lower bound on node repair bandwidth in dif-ferent redundancies is also established.Finally, a comparison between SMCR repair model and uncooperative node repair model is given. Through numerical analysis, it can be seen that SMCR model outperforms the unco-operative node repair model under low redundancy. In the case of high redundancy, it is not necessarily that SMCR model is superior to the uncooperative node repair model depending on different parameters. Further, this thesis gives deep analysis, putting forward some sufficient conditions for that SMCR model is superior to the uncooperative node repair model, which can be used as guidance for practical distribution storage system design.