Research On Rotated Based Scaling Scheme For RAID6Storage System

In recent years, with the rise and development of technologies such as big data and cloud computing, user data and new applications is growing at an explosive rate, which proposes a higher demand for the reliability of data storage systems, the I/O performance and storage capacity. In order to cope with the performance bottlenecks of single disk storage system that may be encountered, RAID technology has been widely applied to various practical storage systems.But as the users’ demand of storage capacity increases,a storage system may inevitably encounter the system scaling. In order to meet the requirements of24hours uninterrupted service, today’s system requires online and realtime scaling. The traditional approaches to storage system scaling are based on the RoundRobin scheduling scaling scheme to meet the optimal system access performance after scaling. But they migrate almost all of the data during the process of scaling, which will degrade the access performance of the system and the users’ experience. To reduce the amount of data to be migrated during the scaling, some new approaches are put forward to minimize data migration.This thesis focuses on the scaling of RAID6storage system. The existing scaling schemes for RAID6systems are to minimize the amount of data migration and assumes that parity blocks are deployed on two dedicated disks. However, in practical systems, parity blocks are usually deployed among all disks with a rotation of different stripes to balance the load of parity blocks. Because the existing scaling schemes only migrate data blocks but not parity blocks, they lead to unbalanced load of the parity blocks in the system after the scaling.Based on the assumption that parity blocks are deployed rotatedlly in practical storage system, this thesis proposes a new scaling scheme called RSR based on the most commonly used two erasure codes RDP code and EVENODD code. RSR meets the requirement of the minimum amount of data/parity blocks being migrated. It also makes the data blocks and parity blocks balancedly deployed among all disks in the system after the scaling by splitting and then splicing multiple stripes.Besides, RSR also uses Piggyback technology to reduce the cost of updating the parity blocks.In order to verify the performance of RSR, we conducted detailed simulations on DiskSim, an accurate and efficient disk simulator extensively used in research involving storage systems. We implemented our scheme RSR and three existing schemes RS6, RR4, RR5, simulated on DiskSim under RDP and EVENODD code respectively using the same scaling parameters, and compared the scaling time and average access time after scaling. For RDP code simulation experiment, we scale from6disks to18disks in four consecutive scaling processes, adding2,4,2,4disks each time. The simulation results show that under RDP coding scheme, the scaling time of RSR decreased by53.45%to76.75%compared to RR4and RR5, and by-0.74%to6.89%compared to RS6. For the write-intensive trace Financial, the average access time after scaling of RSR increased by0.76%to15.92% compared to RR5that achieves the best average access performance currently, while RR4and RS6can not obtain efficient access time due to the unbalanced parity blocks. As for EVENODD coding scheme, RSR showed similar performance in the simulations. Through experimental comparison with existing approaches, RR4, RR5and RS6, we confirm that RSR has a lower cost of scaling, and has the optimal access performance in the system after the scaling.