A. Nathan McNamara

EECS Department
University of California, Berkeley
Technical Report No. UCB/CSD-95-874
May 1995

Multiprocessor networks have been designed with several goals. In addition to performance, other goals have been versatility, fault-tolerance, economy, and simplicity. These goals are often in conflict; for example, a fault-tolerant network will be neither simple nor inexpensive.

The Connectionist Network Supercomputer (CNS) is a multiprocessor that is being designed to exploit the parallelism in connectionist applications. To achieve excellent performance at moderate cost, CNS is being optimized specifically for artificial neural network computation. Features that are not essential for such calculations are omitted.

Applying this principle of "cutting the fat" to the design of the CNS data network, a unidirectional ring topology performs best for small to medium sized systems, while larger systems are best suited to a unidirectional torus. Special hardware for fast multicasting improves broadcast execution time somewhat; however, the overall impact on benchmark applications is small. Further, hardware support for deadlock avoidance is unnecessary because the communication patterns required by the benchmark applications to not require it. Specifically, neither virtual channels nor separate request and reply networks are needed for CNS.

By eliminating unneeded mechanisms from the network, a substantial cost reduction is possible while increasing performance.

BibTeX citation:

@techreport{McNamara:CSD-95-874,
    Author = {McNamara, A. Nathan},
    Title = {Application Specific, Multiprocessor Network Design},
    Institution = {EECS Department, University of California, Berkeley},
    Year = {1995},
    Month = {May},
    URL = {http://www2.eecs.berkeley.edu/Pubs/TechRpts/1995/5632.html},
    Number = {UCB/CSD-95-874},
    Abstract = {Multiprocessor networks have been designed with several goals. In addition to performance, other goals have been versatility, fault-tolerance, economy, and simplicity. These goals are often in conflict; for example, a fault-tolerant network will be neither simple nor inexpensive.  <p>  The Connectionist Network Supercomputer (CNS) is a multiprocessor that is being designed to exploit the parallelism in connectionist applications. To achieve excellent performance at moderate cost, CNS is being optimized specifically for artificial neural network computation.  Features that are not essential for such calculations are omitted.  <p>  Applying this principle of "cutting the fat" to the design of the CNS data network, a unidirectional ring topology performs best for small to medium sized systems, while larger systems are best suited to a unidirectional torus. Special hardware for fast multicasting improves broadcast execution time somewhat; however, the overall impact on benchmark applications is small. Further, hardware support for deadlock avoidance is unnecessary because the communication patterns required by the benchmark applications to not require it. Specifically, neither virtual channels nor separate request and reply networks are needed for CNS.  <p>  By eliminating unneeded mechanisms from the network, a substantial cost reduction is possible while increasing performance.}
}

EndNote citation:

%0 Report
%A McNamara, A. Nathan
%T Application Specific, Multiprocessor Network Design
%I EECS Department, University of California, Berkeley
%D 1995
%@ UCB/CSD-95-874
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/1995/5632.html
%F McNamara:CSD-95-874