Why is MPI so slow? Analyzing the fundamental limits in implementing MPI-3.1

Ken Rafenetti, Abdelhalim Amer, Lena Oden, Charles Archer, Wesley Bland, Hajime Fujita, Yanfei Guo, Tomislav Janjusic, Dmitry Durnov, Michael Blocksome, Min Si, Sangmin Seo, Akhil Langer, Gengbin Zheng, Masamichi Takagi, Paul Cofman, Jithin Jose, Sayantan Sur, Alexander Sannikov, Sergey OblomovMichael Chuvelev, Masayuki Hatanaka, Xin Zhao, Paul Fischer, Thilina Rathnayake, Matt Otten, Misun Min, Pavan Balaji

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

Abstract

This paper provides an in-depth analysis of the software overheads in the MPI performance-critical path and exposes mandatory performance overheads that are unavoidable based on the MPI-3.1 speciication. We irst present a highly optimized implementation of the MPI-3.1 standard in which the communication stackDall the way from the application to the low-level network communication APIDtakes only a few tens of instructions. We carefully study these instructions and analyze the root cause of the overheads based on speciic requirements from the MPI standard that are unavoidable under the current MPI standard. We recommend potential changes to the MPI standard that can minimize these overheads. Our experimental results on a variety of network architectures and applications demonstrate signiicant beneits from our proposed changes.

Original languageEnglish (US)
Title of host publicationProceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, SC 2017
PublisherAssociation for Computing Machinery, Inc
ISBN (Electronic)9781450351140
DOIs
StatePublished - Nov 12 2017
EventInternational Conference for High Performance Computing, Networking, Storage and Analysis, SC 2017 - Denver, United States
Duration: Nov 12 2017Nov 17 2017

Publication series

NameProceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, SC 2017

Other

OtherInternational Conference for High Performance Computing, Networking, Storage and Analysis, SC 2017
CountryUnited States
CityDenver
Period11/12/1711/17/17

ASJC Scopus subject areas

  • Computer Networks and Communications
  • Software

Fingerprint Dive into the research topics of 'Why is MPI so slow? Analyzing the fundamental limits in implementing MPI-3.1'. Together they form a unique fingerprint.

Cite this