Research And Implementation Of High Availability Technology For Object-based Storage

In the information ages with rapid development, data storage occupies a very important position. Under the promotion of internet, tendency of economic globalization requires the running mode to maintain 24x7 available instead of 8×5 available. To keep the normal running of an enterprise, data, most important capital of enterprise, must provide service anytime. It's a signification challenging task of how build a storage with high available and performance.In this thesis it researches high availability of object-based storage file system Lustre, including the following directions: the design and implementation of Lustre file level RAID across network and its inconsistent recovery, algorithms of storage object allocating and load balance, the technology of metadata server cluster and related recovery.First, we analyze the overall architecture and technologies for Lustre. To address the high price of current scheme to achieve high availability, we present and design a redundant storage scheme: Lustre file level RAID (LAID) across network, which can not only improve the aggregation I/O bandwidth but also can tolerance failure of node and disk, greatly improving the reliability and tolerance. Then To solve the inconsistency of LAID, we propose an algorithm for cooperating distributed inconsistent recovery with a little effect on performance and ability of online recovery. Finially, we do a series of I/O bandwidth benchmark test on various LAID modes in the Lustre file system. The result demonstrate that it can not only get very good aggregate bandwith but also improve the reliability and capability of fault tolerance of the system.On the basic of research to load balancing, as the available disk space, available network bandwidth and load change all the time, we propose algorithm of storage object allocation and load balance to uniform utilization of storage space and computational resource of all storage nodes. By object allocation schemes: round-robine, left space weight algorithm and create on write, it makes more uniform to use storage space, avoiding that space of single storage nod(?)runs out and potential load imbalance. And we also propose load balance algorithm based on feedback information of mirror LAID and LAID5 in special case. It targets that reduce the workload of server node and balance the IO load distribution, and choose a least synthetic workload to issue the IO according the feedback information on client from...