Research On Replication Strategy In Cloud Storage

With the development of information technology and the progress of informationization, the global data grows explosively. Confronting PB level storage requirements, traditional file system has a bottleneck on capacity and performance expansion. Cloud storage is widely acknowledged by IT industry because of its advantages, such as expandability, cost-effective, high fault-tolerance. As the basis of cloud storage, distributed file system achieves promotion of system availability and fault-tolerance by introducing data replication technology. However replica brings about a lot of problems such as bandwidth and storage consumption, unexpected data inconsistency, load balance.The replication problems in cloud storage distributed file system are network characteristics of dynamics and wide-area distribution; load balance in replication strategy; business characteristics of cloud storage. Aiming at solving the replication strategy issues in replica initial creating period and replica dynamic adjustment period, this thesis carries out researches as follows:1. Summarizes the related concepts—cloud computing and cloud storage, states the superiority and research point of data replication technology, analyzes the problems in replication strategy and introduces the research topics. Then investigates distributed file system, introduces system design assumptions, system architecture and replication strategies of several well-known distributed file systems.2. According to the problem of replication strategy in cloud storage, a distributed file system model MNFS (Master Node File System) is established in Chapter 3 firstly, and then two replication strategies are proposed in Chapter 4 and Chapter 5. MNFS model is based on centralized server pattern, which defines geographic location as node property, introduces virtual node to shield the performance differences between physical nodes.3. According to the reading intensive cloud application, Chapter 4 investigates the replication placement strategy in replica initial creating period. Then a replication placement strategy ULPS (Placement Strategy based on User Information& Load Balance) is proposed based on users'history information and node loads. The strategy takes consideration of load balance and places replications targeted to promote the system performance. Experiments show that ULPS strategy performances better in load balance and access rate than HDFS strategy. 4. According to replica dynamic adjustment period, a cost-driven replication strategy CDRS (Cost-Driven Replication Strategy) is proposed in Chapter 5. The strategy puts replica availability in the first place and takes considerations of load balance and the balance between consistency and availability by introducing market mechanisms. It copies, deletes and transfers replicas adaptively in order to minimize the replica cost and maximize the replica profit. Comparing with the static replication strategy, the experiment results show that CDRS strategy can improve load balance and the replica profit effectively.Aiming at the replication problems in cloud storage distributed file system, this thesis focuses on the replica initial creating period and replica dynamic adjustment period, builds up a distributed file system model MNFS and proposes two replication strategies—ULPS and CDRS. They give a new point of view and fine prospect for solving the problem of cloud storage replica management.