ACM SRC Poster: Optimizing all-to-all algorithm for percs network using simulation

Ehsan Totoni, Laxmikant V. Kale

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

Abstract

Communication algorithms play a crucial role in the performance of large-scale parallel systems. They are implemented in runtime systems and used in most parallel applications as a critical component. As vendors are willing to design new custom networks with significantly different performance properties for their new supercomputers, designing new efficient communication algorithms is an inevitable challenge. This task is desirable to be done before the machine comes online since inefficient use of the system before the new algorithm's availability is a huge waste of a possibly hundreds of millions of dollars resource. Here, we demonstrate the usability of our simulation framework, BigSim, in meeting this challenge. Using BigSim, we observe that the commonly used Pairwise-Exchange algorithm for all-to-all communication pattern is suboptimal for a supernode of the PERCS network (two-level directly connected similar to Dragony topology). We designed a new all-to-all algorithm for it and predict a five-fold performance improvement for large message sizes using this algorithm.

Original languageEnglish (US)
Title of host publicationSC'11 - Proceedings of the 2011 High Performance Computing Networking, Storage and Analysis Companion, Co-located with SC'11
Pages123-124
Number of pages2
DOIs
StatePublished - Dec 1 2011
Event2011 High Performance Computing Networking, Storage and Analysis, SC'11, Co-located with SC'11 - Seattle, WA, United States
Duration: Nov 12 2011Nov 18 2011

Publication series

NameSC'11 - Proceedings of the 2011 High Performance Computing Networking, Storage and Analysis Companion, Co-located with SC'11

Other

Other2011 High Performance Computing Networking, Storage and Analysis, SC'11, Co-located with SC'11
CountryUnited States
CitySeattle, WA
Period11/12/1111/18/11

Fingerprint

Communication
Supercomputers
Topology
Availability

Keywords

  • Design
  • General Terms Algorithms
  • Measurement
  • Performance

ASJC Scopus subject areas

  • Computer Networks and Communications

Cite this

Totoni, E., & Kale, L. V. (2011). ACM SRC Poster: Optimizing all-to-all algorithm for percs network using simulation. In SC'11 - Proceedings of the 2011 High Performance Computing Networking, Storage and Analysis Companion, Co-located with SC'11 (pp. 123-124). (SC'11 - Proceedings of the 2011 High Performance Computing Networking, Storage and Analysis Companion, Co-located with SC'11). https://doi.org/10.1145/2148600.2148664

ACM SRC Poster : Optimizing all-to-all algorithm for percs network using simulation. / Totoni, Ehsan; Kale, Laxmikant V.

SC'11 - Proceedings of the 2011 High Performance Computing Networking, Storage and Analysis Companion, Co-located with SC'11. 2011. p. 123-124 (SC'11 - Proceedings of the 2011 High Performance Computing Networking, Storage and Analysis Companion, Co-located with SC'11).

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

Totoni, E & Kale, LV 2011, ACM SRC Poster: Optimizing all-to-all algorithm for percs network using simulation. in SC'11 - Proceedings of the 2011 High Performance Computing Networking, Storage and Analysis Companion, Co-located with SC'11. SC'11 - Proceedings of the 2011 High Performance Computing Networking, Storage and Analysis Companion, Co-located with SC'11, pp. 123-124, 2011 High Performance Computing Networking, Storage and Analysis, SC'11, Co-located with SC'11, Seattle, WA, United States, 11/12/11. https://doi.org/10.1145/2148600.2148664
Totoni E, Kale LV. ACM SRC Poster: Optimizing all-to-all algorithm for percs network using simulation. In SC'11 - Proceedings of the 2011 High Performance Computing Networking, Storage and Analysis Companion, Co-located with SC'11. 2011. p. 123-124. (SC'11 - Proceedings of the 2011 High Performance Computing Networking, Storage and Analysis Companion, Co-located with SC'11). https://doi.org/10.1145/2148600.2148664
Totoni, Ehsan ; Kale, Laxmikant V. / ACM SRC Poster : Optimizing all-to-all algorithm for percs network using simulation. SC'11 - Proceedings of the 2011 High Performance Computing Networking, Storage and Analysis Companion, Co-located with SC'11. 2011. pp. 123-124 (SC'11 - Proceedings of the 2011 High Performance Computing Networking, Storage and Analysis Companion, Co-located with SC'11).
@inproceedings{b79f6a964fa247a3adac19844650c5a2,
title = "ACM SRC Poster: Optimizing all-to-all algorithm for percs network using simulation",
abstract = "Communication algorithms play a crucial role in the performance of large-scale parallel systems. They are implemented in runtime systems and used in most parallel applications as a critical component. As vendors are willing to design new custom networks with significantly different performance properties for their new supercomputers, designing new efficient communication algorithms is an inevitable challenge. This task is desirable to be done before the machine comes online since inefficient use of the system before the new algorithm's availability is a huge waste of a possibly hundreds of millions of dollars resource. Here, we demonstrate the usability of our simulation framework, BigSim, in meeting this challenge. Using BigSim, we observe that the commonly used Pairwise-Exchange algorithm for all-to-all communication pattern is suboptimal for a supernode of the PERCS network (two-level directly connected similar to Dragony topology). We designed a new all-to-all algorithm for it and predict a five-fold performance improvement for large message sizes using this algorithm.",
keywords = "Design, General Terms Algorithms, Measurement, Performance",
author = "Ehsan Totoni and Kale, {Laxmikant V.}",
year = "2011",
month = "12",
day = "1",
doi = "10.1145/2148600.2148664",
language = "English (US)",
isbn = "9781450310307",
series = "SC'11 - Proceedings of the 2011 High Performance Computing Networking, Storage and Analysis Companion, Co-located with SC'11",
pages = "123--124",
booktitle = "SC'11 - Proceedings of the 2011 High Performance Computing Networking, Storage and Analysis Companion, Co-located with SC'11",

}

TY - GEN

T1 - ACM SRC Poster

T2 - Optimizing all-to-all algorithm for percs network using simulation

AU - Totoni, Ehsan

AU - Kale, Laxmikant V.

PY - 2011/12/1

Y1 - 2011/12/1

N2 - Communication algorithms play a crucial role in the performance of large-scale parallel systems. They are implemented in runtime systems and used in most parallel applications as a critical component. As vendors are willing to design new custom networks with significantly different performance properties for their new supercomputers, designing new efficient communication algorithms is an inevitable challenge. This task is desirable to be done before the machine comes online since inefficient use of the system before the new algorithm's availability is a huge waste of a possibly hundreds of millions of dollars resource. Here, we demonstrate the usability of our simulation framework, BigSim, in meeting this challenge. Using BigSim, we observe that the commonly used Pairwise-Exchange algorithm for all-to-all communication pattern is suboptimal for a supernode of the PERCS network (two-level directly connected similar to Dragony topology). We designed a new all-to-all algorithm for it and predict a five-fold performance improvement for large message sizes using this algorithm.

AB - Communication algorithms play a crucial role in the performance of large-scale parallel systems. They are implemented in runtime systems and used in most parallel applications as a critical component. As vendors are willing to design new custom networks with significantly different performance properties for their new supercomputers, designing new efficient communication algorithms is an inevitable challenge. This task is desirable to be done before the machine comes online since inefficient use of the system before the new algorithm's availability is a huge waste of a possibly hundreds of millions of dollars resource. Here, we demonstrate the usability of our simulation framework, BigSim, in meeting this challenge. Using BigSim, we observe that the commonly used Pairwise-Exchange algorithm for all-to-all communication pattern is suboptimal for a supernode of the PERCS network (two-level directly connected similar to Dragony topology). We designed a new all-to-all algorithm for it and predict a five-fold performance improvement for large message sizes using this algorithm.

KW - Design

KW - General Terms Algorithms

KW - Measurement

KW - Performance

UR - http://www.scopus.com/inward/record.url?scp=84859017251&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=84859017251&partnerID=8YFLogxK

U2 - 10.1145/2148600.2148664

DO - 10.1145/2148600.2148664

M3 - Conference contribution

AN - SCOPUS:84859017251

SN - 9781450310307

T3 - SC'11 - Proceedings of the 2011 High Performance Computing Networking, Storage and Analysis Companion, Co-located with SC'11

SP - 123

EP - 124

BT - SC'11 - Proceedings of the 2011 High Performance Computing Networking, Storage and Analysis Companion, Co-located with SC'11

ER -