High-Performance GPU-Accelerated Diagonal-Concatenated RAID Schedulers in Cloud System

초록

This paper proposes a new GPU-accelerated diagonal-concatenated RAID in the cloud storage system. A master GPU environment is used to provide dedicated access to Grid GPU cores for each slave node enabling a higher degraded encoding and decoding performance for generating parity than the traditional vGPU technology and pass-through GPU does. The proposed GPU-accelerated diagonal-concatenated RAID reduces the allocation latency, increase the parallelism and resists against three types of node, disk, and sector failures. Experimental results show that the average encoding performance of the proposed RAID 555 is improved by 44% and 56% compared to Song et al.'s GPU-RAID 55 and Khasymski's PAID 51 with respect to various file sizes.