Abstract: The fixed data layout of traditional array system and the locality of workloads cause the partial disks of the array system to become hot disks, which affects the reliability and the overall concurrency performance of the array system. This paper proposes a new RAID architecture for SSD array systems, HA-RAID, which leverages hot/cold data separation and sliding window techniques. The main idea is that HA-RAID divides the disk array into hot disks and ordinary disks, which stores hot data on hot disks and cold data on ordinary disks, and it changes the role of each disk dynamically by moving a fixed-length sliding window. So, each disk has the opportunity to become a hot disk and stores hot data which achieves the purpose of storing hot data evenly on each disk. Experiments under real-world workloads on a RAID-0 array system composed of eight commercial SSDs show that HA-RAID can achieve an even distribution of hot data across all disks and reduce the percentage of hot disks appearing in the array to almost zero. This implies that HA-RAID achieves load balance and wear balance at the device level. In terms of performance, HA-RAID reduces the average response time by 12.01%~41.06% which achieves the I/O performance enhancement, compared with traditional RAID-0 array.

Key words: SSD array architecture design, cold and hot data identification, RAID-0, load balance, response time