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

  • Authors:
  • Rosa Filgueira;Malcolm Atkinson;Alberto Nuñ/ez;Javier Fern$#225/ndez

  • Affiliations:
  • School of Informatics, University of Edinburgh, Edinburgh, U.K.;School of Informatics, University of Edinburgh, Edinburgh, U.K.;Dept. Sistemas Inform$#225/ticos y Computació/n, University Complutense de Madrid, Madrid, Spain;Dept. Arquitectura de Computadores, University Carlos III de Madrid, Legané/s, Spain

  • Venue:
  • Euro-Par'12 Proceedings of the 18th international conference on Parallel Processing
  • Year:
  • 2012

Quantified Score

Hi-index 0.00

Visualization

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.