Item Details

Print View

ELFS: Object-Oriented Extensible File Systems

Grimshaw, Andrew; Loyot, Jr
Grimshaw, Andrew
Loyot, Jr
High performance scientific data analysis is plagued by chronically inadequate I/O performance. The situation is aggravated by ever improving processor performance. For high performance multicomputers, such as the Touchstone Delta that possess in excess of 500, 60 megaflops, processor I/O will be the bottleneck for many scientific applications. This report describes ELFS (an ExtensibLe File System). ELFS attacks the problems of 1) providing high bandwidth and low latency I/O to applications programs on high performance architectures, 2) reducing the cognitive burden faced by applications programmers when they attempt to optimize their I/O operations to fit existing file system models, and 3) seamlessly managing the proliferation of data formats and architectural differences. The ELFS solution consists of language and run-time system support that permits the specification of a hierarchy of file classes. 1. Introduction ELFS: Object-Oriented Extensible File Systems Contemporary high performance computer systems are becoming increasingly unbalanced. CPU speeds have increased dramatically over the last decade. At the same time I/O performance has improved only marginally. Thus, the performance of many scientific applications is bounded by the performance of their database 1 . They cannot get their data in and out of the machine fast enough. Evidence of this problem abounds. NRAO (the National Radio Astronomy Observatory), for example, has many database bound applications. One, a deconvolution algorithm, consumes 20 minutes of Cray CPU time, yet takes over 10 hours of wall clock time. The difference is due to database waits [20]. The advent of highly parallel architectures has made the problem even worse. For example, the Intel 128 node iPSC/860 has a peak performance of 7680 double precision mega-flops, and the recently released Delta has a peak rate almost four times that of the iPSC/860. Yet IO latency is still in the 10-20 millisecond range, and the aggregate bandwidth is only on the order of Note: Abstract extracted from PDF text
Date Received
University of Virginia, Department of Computer Science, 1991
Published Date
Libra Open Repository
In CopyrightIn Copyright
▾See more
▴See less


Access Online