Browse Prior Art Database

Towards a Shared Memory Hypercube

IP.com Disclosure Number: IPCOM000148181D
Original Publication Date: 1988-Nov-28
Included in the Prior Art Database: 2007-Mar-29

Publishing Venue

Software Patent Institute

Related People

Lindsay, Donald C.: AUTHOR [+2]

Abstract

Donald C. Llndsay 28 November 1988 CMU-CS-88-190 Copyright 63 1988 Donald C. Lindsay Department of Computer Science Carnegie Mellon University Pittsburgh, PA 1521 3-3890 lindsay@k.gp.cs.cmu.edo Towards a Shared Memory Hypercube Donald C. Lindsay Department of Computer Science Carnegie Mellon University Pittsburgh, PA 1521 3-3890 lindsay@k.gp.cs.cmu.edu Abstract Early generation hypercube computers have shown great promise, but only message-based programs have been successful in exploiting the potential of these machines. The article shows that shared memory programming techniques can be efficient on a mnventional hypercube with appropriate communications support. Spe- cific hardware features are proposed, based on double-ended circuit transactions. Introduction Although hypercube computers were construct- ed in the 1970's [6l, they have only recently become attractive. They are now seen as a way to assemble our increasingly powerful board-lev- el (or chip-level) microsystems into products having high aggregate bandwidths. This avoids the mainframe approaches to high computation- al bandwidth, high memory bandwidth, and high I10 bandwidth, all of which involve diminishing returns. There is a bright promise that hyper- cubes can be scaled to large sizes without diffi- culty. The high speed and sophistication of recent microprocessors shows that a hypercube may be composed of quite powerful elements, instead of being "an army of ants". Hypercubes are already supporting large disk farms, and it seems possible to eliminate disk cabling by dis- persing today's increasingly tiny disk drives throughout a machine. A large number of efforts have shown that hypercubes can be pro- grammed, and even time-shared, and that many

This text was extracted from a PDF file.
At least one non-text object (such as an image or picture) has been suppressed.
This is the abbreviated version, containing approximately 10% of the total text.

Page 1 of 12

Towards a Shared Memory Hypercube

Donald C. Llndsay 28 November 1988

CMU-CS-88-190

Copyright 63 1988

    Donald C. Lindsay Department of Computer Science Carnegie Mellon University Pittsburgh, PA 1521 3-3890 lindsay@k.gp.cs.cmu.edo

[This page contains 1 picture or other non-text object]

Page 2 of 12

[This page contains 1 picture or other non-text object]

Page 3 of 12

Towards a Shared Memory Hypercube

    Donald C. Lindsay Department of Computer Science Carnegie Mellon University Pittsburgh, PA 1521 3-3890 lindsay@k.gp.cs.cmu.edu

Abstract

Early generation hypercube computers have shown great promise, but only message-based programs have been successful in exploiting the potential of these machines. The article shows that shared memory programming techniques can be efficient on a mnventional hypercube with appropriate communications support. Spe- cific hardware features are proposed, based on double-ended circuit transactions.

Introduction

Although hypercube computers were construct- ed in the 1970's [6l, they have only recently become attractive. They are now seen as a way to assemble our increasingly powerful board-lev- el (or chip-level) microsystems into products having high aggregate bandwidths. This avoids the mainframe approaches to high computation- al bandwidth, high memory bandwidth, and high I10 bandwidth, all of which involve diminishing returns. $ There is a bright promise that hyper- cubes can be scaled to large sizes without diffi- culty. The high speed and sophistication of recent microprocessors shows that a hypercube may be composed of quite powerful elements, instead of being "an army of ants". Hypercubes are already supporting large disk farms, and it seems possible to eliminate disk cabling by dis- persing today's increasingly tiny disk drives throughout a machine. A large number of * efforts have shown that hypercubes can be pro- grammed, and even time-shared, and that many

$ Some examples: almost any cooling method; thin film interconnection; the extensive cross- bars of a multiported, highly interleaved memory.

applications can exploit the power of these machines [8,9,15].

If there is a problem, it is that the early genera- tion machines were unforgiving. For example, performance was quite dependent on the pro- grammer arranging that messages mostly went to directly adjacent nodes. In order to achieve this high locality, tasks and data had to be mapped onto the machine with great care. Although there is now a record of experience and success in this area, it is clear that this was limiting.

In a similar vein, the use of message based pro- gramming paradigms has often resulted in high latencies, which can easily impact performance if the nodes are not multiprograrnmed. It has become important to measure how synchronous an application can be, since highly synchronous codes can carefully overlap all...