Distributed Software RAID Architectures for Parallel I/O in Serverless Clusters*

In a serverless cluster of computers, all local disks can be integrated as a distributed software RAID (ds-RAID) with a single I/O space. This paper presents the architecture and performance of a new RAID-x for building ds-RAID. Through experimentation, we evaluate the RAID-x along with RAID-5, chained-declustering, and RAID-10 architectures, all embedded in a Linux cluster environment. All four ds-RAID architectures aim to scale in aggregate I/O bandwidth. The RAID-x is unique with its orthogonal striping and mirroring (OSM) architecture. The reliability comes from orthogonal mirroring, while the bandwidth is enhanced from distributed striping. To support single I/O space, we have developed cooperative disk drivers (CDD) at the Linux kernel to enable fast remote disk accesses without using a central file server such as the NFS. The performance of the RAID-x is experimentally proven superior in three areas: (1) significant improvement in I/O bandwidth especially in parallel write operations, (2) pipelined mirroring in the background with low overhead, and (3) enhanced scalability and reliability in cluster computing with a single I/O space. These claims are supported by Andrew and Bonnie benchmark results, obtained on the USC cluster of 16 Linux PCs. Reliable cluster middleware and Linux extensions are developed to enable not only single I/O space, but also shared virtual memory and global file hierarchy. Toward this end, we compare the RAID-x with four related parallel or distributed RAID projects: Digital Petal, Berkeley Tertiary Disk, Princeton TickerTAIP, and HP AutoRAID. Their relative strengths, shortcomings, and applications are discussed along with suggested further research.