Research On Performance Optimization For Erasure-Coded Storage Systems

With the scale of storage systems increasingly expands, data loss becomes commonplace. Erasure coding that tolerates failures by pre-storing a low degree of data redundancy, is intensively adopted in current storage systems. The equipment of erasure coding also brings some changes to I/O accesses and failure recovery in storage systems. We carry out this research to optimize the performance of erasure-coded storage systems from four aspects, i.e., code design, partial stripe write operation, single disk failure recovery and cooperative regeneration.? We propose an all-around MDS array code named HV Code(Horizontal-Vertical Code). HV Code utilizes horizontal parity chains to optimize degraded read operations and partial stripe writes that are in the same row. It also designs vertical parity chains to optimize partial stripe writes that are across rows. Parity elements are evenly placed in HV Code for load balancing. Experimental results show that compared with other representative RAID-6 codes, HV Code can provide better overall performance.? We propose PDP(Parity-Switched Data Placement) to optimize partial stripe writes for XOR-coded storage systems. PDP places continuous data elements to generate parity elements and explores the generation orders of parity elements, in order to decrease the parity updates for partial stripe writes. Experimental results show that PDP will introduce a little amount of storage overhead. In return, PDP can decrease up to 31.9% data updates and accelerate the write operation by up to 27.9%.? We propose an algorithm named SIOR(Seek-Efficient I/O Recovery) to optimize both repair traffic and disk seeks during single failure recovery. SIOR makes use of an existing algorithm to optimize the repair traffic in every stripe. Besides,SIOR also designs a greedy algorithm based on Tabu Search and proposes a filling algorithm to decrease disk seeks during the recovery. Experimental results show that SIOR significantly shortens the search time and reduces 31.8%?65.1%disk seeks.? We propose a new cooperative regeneration framework in heterogeneous storage networks and design its routing selection algorithm named HCR(Heterogeneous-aware Cooperative Regeneration). Based on node heterogeneity, the proposed framework first separates the regeneration process on a newcomer to decrease the network cost during data regeneration. HCR then designs greedy algorithms to select efficient routing when deploying existing cooperative regeneration schemes under the new framework. Experimental results show that compared with the conventional cooperative regeneration framework, HCR will need to transmit a bit more data, but it can decrease up to 75.4% transmission time by examining node heterogeneity.