Browse Prior Art Database

Considerations for Massively Parallel UNIX Systems on the NYU Ultracomputer and IBM RP31

IP.com Disclosure Number: IPCOM000128238D
Original Publication Date: 1985-Dec-31
Included in the Prior Art Database: 2005-Sep-15
Document File: 20 page(s) / 73K

Publishing Venue

Software Patent Institute

Related People

Jan Edler: AUTHOR [+5]

Abstract

Novel challenges must be met when designing UNIX implementations for highly parallel shared-memory MIIVID architectures. Of primary importance is the need to avoid serial bottlenecks whenever possible, so that the potential speedup of such machines can be realized. Critical code sections far too short or infrequent to seriously impact performance on today's machines will be of concern on very large machines because the cost of each serial section rises linearly with the number of processors involved. In addition, the kernel interface .must provide for a structured and natural style of general-purpose parallel programming. We present the approaches taken to satisfy these requirements for machines such as the NYU Ultracomputer and the IBM 1RP3. We also describe our preliminary parallel implementation of UNIX, which is currently .running on an eight-processor prototype IJltracomputer.

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 5% of the total text.

Page 1 of 20

THIS DOCUMENT IS AN APPROXIMATE REPRESENTATION OF THE ORIGINAL.

Considerations for Massively Parallel UNIX Systems on the NYU Ultracomputer and IBM RP31

Jan Edler, Allan Gottlieb, Jim Lipkis

Courant Institute of Mathematical Sciences Ultracomputer Research Laboratory New York University 251 Mercer Street New York, NY 10012

LJltracomputer Note #91

December, 1985

ABSTRACT

Novel challenges must be met when designing UNIX implementations for highly parallel shared- memory MIIVID architectures. Of primary importance is the need to avoid serial bottlenecks whenever possible, so that the potential speedup of such machines can be realized. Critical code sections far too short or infrequent to seriously impact performance on today's machines will be of concern on very large machines because the cost of each serial section rises linearly with the number of processors involved. In addition, the kernel interface .must provide for a structured and natural style of general-purpose parallel programming. We present the approaches taken to satisfy these requirements for machines such as the NYU Ultracomputer and the IBM 1RP3. We also describe our preliminary parallel implementation of UNIX, which is currently .running on an eight-processor prototype IJltracomputer.

1. Introduction Continuing advances in microelectronics have inspired many to consider assembling large numbers of powerful processors into a single general-purpose machine capable of solving very large problems. Two such projects currently underway are the NYU IJltracomputer (Gottlieb et al. [83b]) and the IBM 1RP3 (Pfister et al. [85J, Pfister and Norton
[85], Norton and Pfister [85], Brantley et al. [85]), the former a shared-memory design and the latter supporting both shared and

. 'UNIX is a trademark of AT&T Bell Laboratories. 2This work was supported in part by the Applied Mathematical Sciences subprogram of the Office of Energy Research, . U. S. Department of Energy, under contract number DE-AC0276ER03077, and in part by the National Science Foundation, under grant number DCR-8413359.

h~U ~ private memory. At NYU we have been investigating the adaptation of the UNIX operating system to these architectures, . However, it remains to be demonstrated that such machines can be effectively utilized, and that UNIX is well-suited to the needs of such an environment. There are two aspects to this challenge. First, several thousand processors must be coordinated in such a way that their aggregate power is applied to useful computation. Serial code sections in which one processor works while the others wait become bottlenecks that drastically reduce the power obtained, even if the serial section is so small or infrequently executed as to be entirely acceptable on a machine with only modest parallelism. Indeed, the relative cost of a serial bottleneck rises linearly with the number of processors involved. Second, the machine must be programmable by humans. Effective use of high degrees of...