A Performance Evaluation of the Hemingway DSM System on a Network of SMPs

Abstract

Numerous designs for software distributed shared memory systems have been proposed. Most designs use uniprocessor workstations as the building blocks. In recent years there has been an increase in commodity multiprocessor workstations, with hardware-maintained internal memory coherence mechanisms. In this paper we investigate the performance of a software distributed shared memory system, Hemingway, which is built out of such multiprocessor workstations, utilizing off-the-shelf communication networks. The effectiveness of this system can be evaluated by studying performance as a function of both the total number of processors in the system and the degree of clustering (size of multiprocessor workstations). We evaluated the performance of Hemingway with systems of up to 8 processors, with different levels of clustering. We also compared the performance of our protocol with a similar, established protocol, the Munin protocol. Our results describe a system that scales well both with the number of processors and with clustering. Moreover, our studies indicate that the Hemingway protocol requires lower intra-workstation and inter- workstation network bandwidths than other protocols. Overall we have found that clustering is very effective in increasing performance in software DSM systems built with multiwriter, write-through memory consistency policies.

Open PDF

Document Details

Document Type
Technical Report
Publication Date
Jan 01, 1997
Accession Number
ADA461990

Entities

People

  • Anshu Aggarwal
  • Dirk Grumwald

Organizations

  • University of Colorado Boulder

Tags

DTIC Thesaurus Topics

  • Abstracts
  • Availability
  • Bandwidth
  • Classification
  • Clustering
  • Colorado
  • Commodities
  • Communication Networks
  • Computers
  • Consistency
  • Contracts
  • Information Operations
  • Instructions
  • Monitoring
  • Multiprocessors
  • Networks
  • Test And Evaluation

Fields of Study

  • Computer science
  • Engineering

Readers

  • Parallel and Distributed Computing.