A study on the performance of reproducible computations

Nicola Bombace; Michele Weiland

doi:10.1007/978-3-030-34356-9_33

A study on the performance of reproducible computations

Research output: Chapter in Book/Report/Conference proceeding › Chapter (peer-reviewed) › peer-review

Abstract

Parallel computations are intrinsically non-reproducible, due to a combined effect of non-deterministic parallel reductions and nonassociative floating point operations. Different strategies have been proposed in literature to alleviate this issue or eliminate it altogether, however at present there is no study on the performance impact of associative floating point operations on large scale applications. In this work, we implement associative operations using binned doubles in MiniFE, and perform various performance tests on Cirrus and Fulhame, two state-ofthe-art HPC systems.

Original language	English
Title of host publication	High Performance Computing
Subtitle of host publication	ISC High Performance 2019 International Workshops, Frankfurt/Main, Germany, June 16-20, 2019, Revised Selected Papers
Editors	Michele Weiland, Guido Juckeland, Sadaf Alam, Heike Jagode
Publisher	Springer
Pages	441-451
Number of pages	12
ISBN (Electronic)	978-3-030-34356-9
ISBN (Print)	978-3-030-34355-2
DOIs	https://doi.org/10.1007/978-3-030-34356-9_33
Publication status	Published - 3 Dec 2019
Event	ISC High Performance 2019 - Frankfurt, Germany Duration: 16 Jun 2019 → 20 Jun 2019 https://www.isc-hpc.com/research-papers-2019.html

Publication series

Name	Lecture Notes in Computer Science
Publisher	Springer
Volume	11887
ISSN (Print)	0302-9743
ISSN (Electronic)	1611-3349

Conference

Conference	ISC High Performance 2019
Abbreviated title	ISC HPC 2019
Country	Germany
City	Frankfurt
Period	16/06/19 → 20/06/19
Internet address	https://www.isc-hpc.com/research-papers-2019.html

Keywords

Reproducibility
Binned doubles
Performance

Access to Document

10.1007/978-3-030-34356-9_33Licence: All Rights Reserved

A_study_on_the_performance_of_reproducible_computations__ISC19_Accepted author manuscript, 246 KB

https://link.springer.com/chapter/10.1007/978-3-030-34356-9_33Licence: All Rights Reserved

1 Active

Strategic Partnership in Computational Science for Advanced Simulation and Modelling of Engineering Systems - ASiMoV
Parsons, M.
EPSRC
1/10/18 → 31/03/24
Project: Research

Michele Weiland
- College of Science and Engineering - Senior Research Fellow
- Edinburgh Parallel Computing Centre
- Computer Systems
Person: Academic: Research Active , Non-Academic: Research Active

Cite this

Bombace, N., & Weiland, M. (2019). A study on the performance of reproducible computations. In M. Weiland, G. Juckeland, S. Alam, & H. Jagode (Eds.), High Performance Computing: ISC High Performance 2019 International Workshops, Frankfurt/Main, Germany, June 16-20, 2019, Revised Selected Papers (pp. 441-451). (Lecture Notes in Computer Science; Vol. 11887). Springer. https://doi.org/10.1007/978-3-030-34356-9_33

Bombace, Nicola ; Weiland, Michele. / A study on the performance of reproducible computations. High Performance Computing: ISC High Performance 2019 International Workshops, Frankfurt/Main, Germany, June 16-20, 2019, Revised Selected Papers. editor / Michele Weiland ; Guido Juckeland ; Sadaf Alam ; Heike Jagode. Springer, 2019. pp. 441-451 (Lecture Notes in Computer Science).

@inbook{03b6c66689c54c9ea584780b832d5735,

title = "A study on the performance of reproducible computations",

abstract = "Parallel computations are intrinsically non-reproducible, due to a combined effect of non-deterministic parallel reductions and nonassociative floating point operations. Different strategies have been proposed in literature to alleviate this issue or eliminate it altogether, however at present there is no study on the performance impact of associative floating point operations on large scale applications. In this work, we implement associative operations using binned doubles in MiniFE, and perform various performance tests on Cirrus and Fulhame, two state-ofthe-art HPC systems.",

keywords = "Reproducibility, Binned doubles, Performance",

author = "Nicola Bombace and Michele Weiland",

year = "2019",

month = dec,

day = "3",

doi = "10.1007/978-3-030-34356-9_33",

language = "English",

isbn = "978-3-030-34355-2",

series = "Lecture Notes in Computer Science",

publisher = "Springer",

pages = "441--451",

editor = "Michele Weiland and Guido Juckeland and Sadaf Alam and Heike Jagode",

booktitle = "High Performance Computing",

note = "ISC High Performance 2019, ISC HPC 2019 ; Conference date: 16-06-2019 Through 20-06-2019",

url = "https://www.isc-hpc.com/research-papers-2019.html",

}

Bombace, N & Weiland, M 2019, A study on the performance of reproducible computations. in M Weiland, G Juckeland, S Alam & H Jagode (eds), High Performance Computing: ISC High Performance 2019 International Workshops, Frankfurt/Main, Germany, June 16-20, 2019, Revised Selected Papers. Lecture Notes in Computer Science, vol. 11887, Springer, pp. 441-451, ISC High Performance 2019, Frankfurt, Germany, 16/06/19. https://doi.org/10.1007/978-3-030-34356-9_33

A study on the performance of reproducible computations. / Bombace, Nicola; Weiland, Michele.

High Performance Computing: ISC High Performance 2019 International Workshops, Frankfurt/Main, Germany, June 16-20, 2019, Revised Selected Papers. ed. / Michele Weiland; Guido Juckeland; Sadaf Alam; Heike Jagode. Springer, 2019. p. 441-451 (Lecture Notes in Computer Science; Vol. 11887).

Research output: Chapter in Book/Report/Conference proceeding › Chapter (peer-reviewed) › peer-review

TY - CHAP

T1 - A study on the performance of reproducible computations

AU - Bombace, Nicola

AU - Weiland, Michele

PY - 2019/12/3

Y1 - 2019/12/3

N2 - Parallel computations are intrinsically non-reproducible, due to a combined effect of non-deterministic parallel reductions and nonassociative floating point operations. Different strategies have been proposed in literature to alleviate this issue or eliminate it altogether, however at present there is no study on the performance impact of associative floating point operations on large scale applications. In this work, we implement associative operations using binned doubles in MiniFE, and perform various performance tests on Cirrus and Fulhame, two state-ofthe-art HPC systems.

AB - Parallel computations are intrinsically non-reproducible, due to a combined effect of non-deterministic parallel reductions and nonassociative floating point operations. Different strategies have been proposed in literature to alleviate this issue or eliminate it altogether, however at present there is no study on the performance impact of associative floating point operations on large scale applications. In this work, we implement associative operations using binned doubles in MiniFE, and perform various performance tests on Cirrus and Fulhame, two state-ofthe-art HPC systems.

KW - Reproducibility

KW - Binned doubles

KW - Performance

U2 - 10.1007/978-3-030-34356-9_33

DO - 10.1007/978-3-030-34356-9_33

M3 - Chapter (peer-reviewed)

SN - 978-3-030-34355-2

T3 - Lecture Notes in Computer Science

SP - 441

EP - 451

BT - High Performance Computing

A2 - Weiland, Michele

A2 - Juckeland, Guido

A2 - Alam, Sadaf

A2 - Jagode, Heike

PB - Springer

T2 - ISC High Performance 2019

Y2 - 16 June 2019 through 20 June 2019

ER -

Bombace N, Weiland M. A study on the performance of reproducible computations. In Weiland M, Juckeland G, Alam S, Jagode H, editors, High Performance Computing: ISC High Performance 2019 International Workshops, Frankfurt/Main, Germany, June 16-20, 2019, Revised Selected Papers. Springer. 2019. p. 441-451. (Lecture Notes in Computer Science). https://doi.org/10.1007/978-3-030-34356-9_33

A study on the performance of reproducible computations

Abstract

Publication series

Conference

Keywords

Access to Document

Strategic Partnership in Computational Science for Advanced Simulation and Modelling of Engineering Systems - ASiMoV

Michele Weiland

Cite this