RAIDX: Raid Extended for Heterogeneous Arrays

摘要

The computer hard drive market has diversified with the establishment of solid state disks (SSDs) as an alternative to magnetic hard disks (HDDs). Each hard drive technology has its advantages: the SSDs are faster than HDDs but the HDDs are cheaper. Our goal is to construct a parallel storage system with HDDs and SSDs such that the parallel system is as fast as the SSDs. Achieving this goal is challenging since the slow HDDs store more data and become bottlenecks, while the SSDs remain idle. RAIDX is a parallel storage system designed for disks of different speeds, capacities and technologies. The RAIDX hardware consists of an array of disks; the RAIDX software consists of data structures and algorithms that allow the disks to be viewed as a single storage unit that has capacity equal to the sum of the capacities of its disks, failure rate lower than the failure rate of its individual disks, and speeds close to that of its faster disks. RAIDX achieves its performance goals with the aid of its novel parallel data organization technique that allows storage data to be moved on the fly without impacting the upper level file system. We show that storage data accesses satisfy the locality of reference principle, whereby only a small fraction of storage data are accessed frequently. RAIDX has a monitoring program that identifies frequently accessed blocks and a migration program that moves frequently accessed blocks to faster disks. The faster disks are caches that store the solo copy of frequently accessed data. Experimental evaluation has shown that a HDD+SSD RAIDX array is as fast as an all-SSD array when the workload shows locality of reference.