An adaptive, scalable, and portable technique for speeding up MPI-based applications

Rosa Filgueira*, Malcolm Atkinson, Alberto Nuñez, Javier Fernández

*Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingConference contribution

5 Citations (Scopus)

Abstract

This paper presents a portable optimization for MPI communications, called PRAcTICaL-MPI (Portable Adaptive Compression Library- MPI). PRAcTICaL-MPI reduces the data volume exchanged among processes by using lossless compression and offers two main advantages. Firstly, it is independent of the MPI implementation and the application used. Secondly, it allows for turning the compression on and off and selecting the most appropriate compression algorithm at run-time, depending on the characteristics of each message and on network performance. We have validated PRAcTICaL-MPI in different MPI implementations and HPC clusters. The evaluation shows that compressing MPI messages with the best algorithm and only when it is worthwhile, we obtain a great reduction in the overall execution time for many of the scenarios considered.

Original languageEnglish
Title of host publicationParallel Processing - 18th International Conference, Euro-Par 2012, Proceedings
Pages729-740
Number of pages12
DOIs
Publication statusPublished - 2012
Event18th International Conference on Parallel Processing, Euro-Par 2012 - Rhodes Island, Greece
Duration: 27 Aug 201231 Aug 2012

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume7484 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference18th International Conference on Parallel Processing, Euro-Par 2012
Country/TerritoryGreece
CityRhodes Island
Period27/08/1231/08/12

Keywords

  • Adaptive systems
  • Compression algorithms
  • High-Performance Computing
  • MPI Library
  • Parallel techniques
  • Portable optimizations

Fingerprint

Dive into the research topics of 'An adaptive, scalable, and portable technique for speeding up MPI-based applications'. Together they form a unique fingerprint.

Cite this