Research On Data Access Optimizations For Erasure-Coded Storage Clusters

Recent years, with the rapid development of information technology, the amount ofdata was explosive growth, and distributed storage is widely used. The reliability andavailability of data has also been a great deal of attention. In this case, as one of theredundant mechanism, erasure codes are widely used in the data storage system in order toobtain high reliability and availability. However, erasure-coded storage cluster cost morein reading, writing, and reconstruction when failure. Therefore, designing new schemes onreading, writing and reconstruction to improve the overall performance of system has theimportant significance and broad application prospects.Aimed at reading, writing and reconstruction, the paper proposes three optimizedschemes under erasure-coded storage cluster respectively. And we call themMinimum-Load based Read optimization scheme, Partial-Updating Scheme andROW-based Online-Reconstruction scheme respectively.In the Minimum-Load based read optimization scheme, we first design loadbenchmarks according to the characteristics of erasure-codes system. According to workload of storage nodes, read requests to high loaded nodes will be transferred to other lowerloaded nodes. Then the requested desired data will be decoded based on the decodedalgorithm of erasure codes. So that, the system can be more load balanced to optimize theupdating time and user response time.In the scheme of PUS, it takes part of the updating work from updating manager tostorage nodes by taking full use of the computing capacity of them. Thereby, it not onlycan reduce the read, write, and transmission overhead caused by update, but caneffectively shorten the updating process, also the updating load of the updating manager isto be reduced by transferring. The results show that, PUS can optimize the updating timeby a factor of42%at least.In ROW-R scheme, based on the principle that minimizing the influence onreconstruction by user workload, all write requests and part read requests which orientedto the failed node are redirect oriented to the viable nodes, thus to separate user workflowand reconstruction workflow in a certain extent. Thus，it can speed up the reconstruction by taking full advantage of the high-performance characteristics of the disk in thecontinuous write to The results show that ROW-R can shorten the user response time for52%and speed up reconstruction process by a factor of6%.