Approximate communication: Techniques for reducing communication bottlenecks in large-scale parallel systems

Filipe Betzel, Karen Khatamifard, Harini Suresh, David J. Lilja, John Sartori, Ulya Karpuzcu

Research output: Contribution to journalArticle

12 Scopus citations

Abstract

Approximate computing has gained research attention recently as a way to increase energy efficiency and/or performance by exploiting some applications' intrinsic error resiliency. However, little attention has been given to its potential for tackling the communication bottleneck that remains one of the looming challenges to be tackled for efficient parallelism. This article explores the potential benefits of approximate computing for communication reduction by surveying three promising techniques for approximate communication: compression, relaxed synchronization, and value prediction. The techniques are compared based on an evaluation framework composed of communication cost reduction, performance, energy reduction, applicability, overheads, and output degradation. Comparison results demonstrate that lossy link compression and approximate value prediction show great promise for reducing the communication bottleneck in bandwidth-constrained applications. Meanwhile, relaxed synchronization is found to provide large speedups for select error-Tolerant applications, but suffers from limited general applicability and unreliable output degradation guarantees. Finally, this article concludes with several suggestions for future research on approximate communication techniques.

Original languageEnglish (US)
Article number1
JournalACM Computing Surveys
Volume51
Issue number1
DOIs
StatePublished - Jan 2018

    Fingerprint

Keywords

  • Approximate communication
  • Approximate computing
  • Communication reduction
  • Scalability

Cite this