Abstract: "A computer system is most useful when it has well- balanced processor and I/O performance. Parallel architectures allow fast computers to be constructed from unsophisticated hardware. The usefulness of these machines is severely limited unless they are fitted with I/O subsystems that...

Ausführliche Beschreibung

Bibliographische Detailangaben

1. Verfasser:	Dibble, Peter C. (VerfasserIn)
Format:	Buch
Sprache:	English
Veröffentlicht:	Rochester, NY 1990
Schriftenreihe:	University <Rochester, NY> / Department of Computer Science: Technical report 334
Schlagworte:	Computer architecture Parallel processing (Electronic computers) Hochschulschrift
Zusammenfassung:	Abstract: "A computer system is most useful when it has well- balanced processor and I/O performance. Parallel architectures allow fast computers to be constructed from unsophisticated hardware. The usefulness of these machines is severely limited unless they are fitted with I/O subsystems that match their CPU performance. Most parallel computers have insufficient I/O performance, or use exotic hardware to force enough I/O bandwidth through a uniprocessor file system. This approach is only useful for small numbers of processors. Even a modestly parallel computer cannot be served by an ordinary file system Only a parallel file system can scale with the processor hardware to meet the I/O demands of a parallel computer. This dissertation introduces the concept of a parallel interleaved file system. This class of file system incorporates three concepts: parallelism, interleaving, and tools. Parallelism appears as a characteristic of the file system program and in the disk hardware. The parallel file system software and hardware allows the file system to scale with the other components of a multiprocessor computer. Interleaving is the rule the file system uses to distribute data among the processors Interleaved record distribution is the simplest and in many ways the best algorithm for allocating records to processors. Tools are application code that can enter the file system at a level that exposes the parallel structure of the files. In many cases tools decrease interprocessor communication by moving processing to the data instead of moving the data. The thesis of this dissertation is that a parallel interleaved file system will provide scalable high-performance I/O for a wide range of parallel architectures while supporting a comprehensive set of conventional file system facilities. We have confirmed our performance claims experimentally and theoretically
Beschreibung:	Zugl.: Rochester, NY, Univ. of Rochester, NY, Diss
Beschreibung:	VIII, 113 S. graph. Darst.