On parallel processing benchmarks

Abstract

This is a summary of preliminary work and experiences at NBS on benchmarks for parallel machines. Discussion covers the several roles that benchmarks play, ideal and realistic settings, and quick reviews of several types of benchmark sets. Several conclusions can be drawn, given the rudimentary nature of parallel processor characterization: (i) That the performance metrics be embedded in more comprehensive frameworks that can be appraised and modified as needed; (ii) That one universal framework is beyond reach, since distinct clusters of use are emerging with separate emphases; (iii) That large application benchmarks are most successf^ul when they run well on a machine, and thereby demonstrate compatibility of job and architecture; (iv) That the value of smaller metrics (fragments of code) is more diagnostic and preventive than predictive; small metric sets should encourage the parametric study of architectures and applications, and thus promote both economical hardware enhancement and suitable program design.

Cite this paper

@article{lyon1987parallel,
  title = {{On parallel processing benchmarks}},
  author = {Lyon, Gordon},
  year = {1987},
  doi = {10.6028/nbs.ir.87-3580}
}

Made from the index record. Check names, capitalisation and page numbers against the paper. Need another style? AI for Research has over 10,000.

Related papers