The Myth of Transfer Rate How and Why SCSI Is Better than IPI for NFS

Disk drives are often dismissed as mundane devices, but they are actually interesting, complicated, and misunderstood. As used in traditional Unix compute servers or general-purpose servers, disk storage subsystems have usually been optimized for excellent sequential-transfer performance. Perhaps counter-intuitively, however, NFS file servers exhibit marked random-access disk traffic. This report explores this apparent contradiction. It investigates the anatomy of a client-to-server NFS I/O operation in complete, microsecond-level detail. It debunks the myth of disk transfer rate, and shows instead that disk-drive concurrency is the most important factor in disk storage performance for most NFS network file servers. The paper presents a careful analysis of both SCSI and IPI disk drives. It begins with a concrete and detailed comparison of both performance-oriented and non-performance-oriented technical specifications of SCSI and IPI drive and interface types. It offers a thorough empirical evaluation of SCSI disk drive performance, varying parameters such as synchronous or asynchronous bus transfers, random and sequential access patterns, and multiplicity of drives per SCSI channel. It discusses (nonempirically) similar characteristics for IPI-2 drives. Measured disk traffic patterns are presented for several>5-GB NFS servers using SCSI disk arrays. As can be expected in a 40-200 user file server, request rates are random and independent--not sequential as often benchmarked in general-purpose servers. The report concludes with carefully benchmarked comparisons of file servers using SCSI-based disk arrays (like the Auspe