Performance analysis of the RAID 5 disk array

Abstract

While the processor and memory performance of computers continues to improve, I/O performance has remained relatively constant. Strategies to achieve better I/O performance than current disk systems have been investigated to address the growing I/O bottleneck. One effort is the RAID (Redundant Arrays of Inexpensive Disks) Level 5 disk array. RAID 5 offers increased parallelism of I/O requests through the disk array architecture and fault tolerance through rotated parity. Although RAID 5 offers the promise of improved I/O performance, such gains must be quantified. Accurate analytical modeling can be used to quantify and predict the I/O request response time for different workloads. While previous analytical models to calculate the I/O request response time have been constructed, they often rely upon simplifying assumptions or computational bounds, which are accurate for certain values of the possible workload parameters. This paper presents an analytical performance model to compute the mean steady state response time of a RAID 5 I/O request under a transaction-processing workload. By carefully considering individual disk accesses, the arrival process of disk requests, and correlation between disk requests, an accurate model for a wide range of the workload parameters is developed. >