High-performance disk I/O in a bus-based system

Processor and memory speeds continue to increase dramatically. The large gap in access times between memories and disk storage devices results in limiting the overall system performance. A major contribution to the slow response of storage devices are mechanical latencies arising from disk platter rotation and read/write head actuators, but other limitations are a result of system component interaction. Even though higher densities and rotational speeds reduce mechanical latencies, this reduction is not enough to satisfy the increasing appetite for higher bandwidth and lower latency. The concept of striping data across disk arrays is used to help solve this problem. Custom I/O subsystem design often proves uneconomical. The paper describes an I/O system implementation using off the shelf components and the issues involved in achieving the sustained 20 MByte/sec transfer rate required by the design