darrell@terra.ucsc.edu (Darrell Long) (04/23/91)
The following UCSC technical report (UCSC-CRL-91-08) is available via anonymous FTP from midgard.ucsc.edu (128.114.14.6). The file is pub/tr/ucsc-crl-91-08.ps.Z Be sure to use binary more when doing the transfer. Exploiting Multiple I/O Streams to Provide High Data-Rates Luis-Felipe Cabrera Darrell D. E. Long IBM Almaden Research Center Computer & Information Sciences Computer Science Department University of California at Santa Cruz Internet: cabrera@ibm.com Internet: darrell@sequoia.ucsc.edu ABSTRACT We present an I/O architecture, called Swift, that addresses the problem of data-rate mismatches between the requirements of an application, the maximum data-rate of the storage dev- ices, and the data-rate of the interconnection medium. The goal of Swift is to support integrated continuous multimedia in general purpose distributed systems. In installations with a high-speed interconnection medium, Swift will provide high data-rate transfers by using multiple slower storage devices in parallel. The data-rates obtained with this approach scale well when using multiple storage devices and multiple interconnections. Swift has the flexibility to use any appropriate storage technology, including disk arrays. The ability to adapt to technologi- cal advances will allow Swift to provide for ever increasing I/O demands. To address the problem of partial failures, Swift stores data redundantly. Using the UNIX operating system, we have constructed a simplified prototype of the Swift architecture. Using a single Ethernet-based local-area network and three servers, the prototype provides data-rates that are almost three times as fast as access to the local SCSI disk in the case of writes. When compared with NFS, the Swift prototype pro- vides double the data-rate for reads and eight times the data-rate for writes. The data-rate of our prototype scales almost linearly in the number of servers and the number of network segments. Its performance is shown to be limited by the speed of the Ethernet-based local-area network. We also constructed a simulation model to show how the Swift architecture can exploit storage, communication, and processor advances, and to locate the components that will limit I/O performance. In a simulated gigabit/second token ring local-area network the data-rates are seen to scale proportionally to the size of the transfer unit and to the number of storage agents.