Formal metrics for large-scale parallel performance

Kenneth Moreland; Ron Oldfield

doi:10.1007/978-3-319-20119-1_34

Formal metrics for large-scale parallel performance

Kenneth Moreland, Ron Oldfield

Research output: Contribution to journal › Conference article › peer-review

15 Scopus citations

Abstract

Performance measurement of parallel algorithms is well studied and well understood. However, a flaw in traditional performance metrics is that they rely on comparisons to serial performance with the same input. This comparison is convenient for theoretical complexity analysis but impossible to perform in large-scale empirical studies with data sizes far too large to run on a single serial computer. Consequently, scaling studies currently rely on ad hoc methods that, although effective, have no grounded mathematical models. In this position paper we advocate using a rate-based model that has a concrete meaning relative to speedup and efficiency and that can be used to unify strong and weak scaling studies.

Original language	English
Pages (from-to)	488-496
Number of pages	9
Journal	Lecture Notes in Computer Science
Volume	9137 LNCS
DOIs	https://doi.org/10.1007/978-3-319-20119-1_34
State	Published - 2015
Externally published	Yes
Event	30th International Conference on High Performance Computing, ISC 2015 - Frankfurt, Germany Duration: Jul 12 2015 → Jul 16 2015

Funding

This material is based in part upon work supported by the U.S. Department of Energy, Office of Science, Office of Advanced Scientific Computing Research, Scientific Discovery through Advanced Computing (SciDAC) program under Award Number 12-015215.

Funders	Funder number
U.S. Department of Energy
Office of Science
Advanced Scientific Computing Research	12-015215

Access to Document

10.1007/978-3-319-20119-1_34

Cite this

@article{6574d51d85674631b6a7b0822a73e34b,

title = "Formal metrics for large-scale parallel performance",

abstract = "Performance measurement of parallel algorithms is well studied and well understood. However, a flaw in traditional performance metrics is that they rely on comparisons to serial performance with the same input. This comparison is convenient for theoretical complexity analysis but impossible to perform in large-scale empirical studies with data sizes far too large to run on a single serial computer. Consequently, scaling studies currently rely on ad hoc methods that, although effective, have no grounded mathematical models. In this position paper we advocate using a rate-based model that has a concrete meaning relative to speedup and efficiency and that can be used to unify strong and weak scaling studies.",

author = "Kenneth Moreland and Ron Oldfield",

note = "Publisher Copyright: {\textcopyright} Springer International Publishing Switzerland 2015.; 30th International Conference on High Performance Computing, ISC 2015 ; Conference date: 12-07-2015 Through 16-07-2015",

year = "2015",

doi = "10.1007/978-3-319-20119-1_34",

language = "English",

volume = "9137 LNCS",

pages = "488--496",

journal = "Lecture Notes in Computer Science",

issn = "0302-9743",

}

TY - JOUR

T1 - Formal metrics for large-scale parallel performance

AU - Moreland, Kenneth

AU - Oldfield, Ron

N1 - Publisher Copyright: © Springer International Publishing Switzerland 2015.

PY - 2015

Y1 - 2015

N2 - Performance measurement of parallel algorithms is well studied and well understood. However, a flaw in traditional performance metrics is that they rely on comparisons to serial performance with the same input. This comparison is convenient for theoretical complexity analysis but impossible to perform in large-scale empirical studies with data sizes far too large to run on a single serial computer. Consequently, scaling studies currently rely on ad hoc methods that, although effective, have no grounded mathematical models. In this position paper we advocate using a rate-based model that has a concrete meaning relative to speedup and efficiency and that can be used to unify strong and weak scaling studies.

AB - Performance measurement of parallel algorithms is well studied and well understood. However, a flaw in traditional performance metrics is that they rely on comparisons to serial performance with the same input. This comparison is convenient for theoretical complexity analysis but impossible to perform in large-scale empirical studies with data sizes far too large to run on a single serial computer. Consequently, scaling studies currently rely on ad hoc methods that, although effective, have no grounded mathematical models. In this position paper we advocate using a rate-based model that has a concrete meaning relative to speedup and efficiency and that can be used to unify strong and weak scaling studies.

UR - http://www.scopus.com/inward/record.url?scp=84984663150&partnerID=8YFLogxK

U2 - 10.1007/978-3-319-20119-1_34

DO - 10.1007/978-3-319-20119-1_34

M3 - Conference article

AN - SCOPUS:84984663150

SN - 0302-9743

VL - 9137 LNCS

SP - 488

EP - 496

JO - Lecture Notes in Computer Science

JF - Lecture Notes in Computer Science

T2 - 30th International Conference on High Performance Computing, ISC 2015

Y2 - 12 July 2015 through 16 July 2015

ER -

Formal metrics for large-scale parallel performance

Abstract

Funding

Access to Document

Other files and links

Fingerprint

Cite this