A Study on Balancing Parallelism, Data Locality, and Recomputation in Existing PDE Solvers

Catherine Olschanowsky; Michelle Mills Strout; Stephen Guzik; John Loffeld; Jeffrey Hittinger

doi:10.1109/SC.2014.70

A Study on Balancing Parallelism, Data Locality, and Recomputation in Existing PDE Solvers

Catherine Olschanowsky, Michelle Mills Strout, Stephen Guzik, John Loffeld, Jeffrey Hittinger

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

30 Scopus citations

Abstract

Structured-grid PDE solver frameworks parallelize over boxes, which are rectangular domains of cells or faces in a structured grid. In the Chombo framework, the box sizes are typically 16³ or 32³, but larger box sizes such as 128³ would result in less surface area and therefore less storage, copying, and/or ghost cells communication overhead. Unfortunately, current on node parallelization schemes perform poorly for these larger box sizes. In this paper, we investigate 30 different inter-loop optimization strategies and demonstrate the parallel scaling advantages of some of these variants on NUMA multicore nodes. Shifted, fused, and communication-avoiding variants for 128³ boxes result in close to ideal parallel scaling and come close to matching the performance of 16³ boxes on three different multicore systems for a benchmark that is a proxy for program idioms found in Computational Fluid Dynamic (CFD) codes.

Original language	English (US)
Article number	7013052
Pages (from-to)	793-804
Number of pages	12
Journal	International Conference for High Performance Computing, Networking, Storage and Analysis, SC
Volume	2015-January
Issue number	January
DOIs	https://doi.org/10.1109/SC.2014.70
State	Published - Jan 16 2014
Externally published	Yes
Event	International Conference for High Performance Computing, Networking, Storage and Analysis, SC 2014 - New Orleans, United States Duration: Nov 16 2014 → Nov 21 2014

ASJC Scopus subject areas

Computer Networks and Communications
Computer Science Applications
Hardware and Architecture
Software

Access to Document

10.1109/SC.2014.70

Cite this

A Study on Balancing Parallelism, Data Locality, and Recomputation in Existing PDE Solvers. / Olschanowsky, Catherine; Strout, Michelle Mills; Guzik, Stephen et al.
In: International Conference for High Performance Computing, Networking, Storage and Analysis, SC, Vol. 2015-January, No. January, 7013052, 16.01.2014, p. 793-804.

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

@article{ea1069bd1f994bb7a022dfcefd314558,

title = "A Study on Balancing Parallelism, Data Locality, and Recomputation in Existing PDE Solvers",

abstract = "Structured-grid PDE solver frameworks parallelize over boxes, which are rectangular domains of cells or faces in a structured grid. In the Chombo framework, the box sizes are typically 163 or 323, but larger box sizes such as 1283 would result in less surface area and therefore less storage, copying, and/or ghost cells communication overhead. Unfortunately, current on node parallelization schemes perform poorly for these larger box sizes. In this paper, we investigate 30 different inter-loop optimization strategies and demonstrate the parallel scaling advantages of some of these variants on NUMA multicore nodes. Shifted, fused, and communication-avoiding variants for 1283 boxes result in close to ideal parallel scaling and come close to matching the performance of 163 boxes on three different multicore systems for a benchmark that is a proxy for program idioms found in Computational Fluid Dynamic (CFD) codes.",

author = "Catherine Olschanowsky and Strout, {Michelle Mills} and Stephen Guzik and John Loffeld and Jeffrey Hittinger",

note = "Publisher Copyright: {\textcopyright} 2014 IEEE.; International Conference for High Performance Computing, Networking, Storage and Analysis, SC 2014 ; Conference date: 16-11-2014 Through 21-11-2014",

year = "2014",

month = jan,

day = "16",

doi = "10.1109/SC.2014.70",

language = "English (US)",

volume = "2015-January",

pages = "793--804",

journal = "International Conference for High Performance Computing, Networking, Storage and Analysis, SC",

issn = "2167-4329",

publisher = "IEEE Computer Society",

number = "January",

}

TY - JOUR

T1 - A Study on Balancing Parallelism, Data Locality, and Recomputation in Existing PDE Solvers

AU - Olschanowsky, Catherine

AU - Strout, Michelle Mills

AU - Guzik, Stephen

AU - Loffeld, John

AU - Hittinger, Jeffrey

PY - 2014/1/16

Y1 - 2014/1/16

N2 - Structured-grid PDE solver frameworks parallelize over boxes, which are rectangular domains of cells or faces in a structured grid. In the Chombo framework, the box sizes are typically 163 or 323, but larger box sizes such as 1283 would result in less surface area and therefore less storage, copying, and/or ghost cells communication overhead. Unfortunately, current on node parallelization schemes perform poorly for these larger box sizes. In this paper, we investigate 30 different inter-loop optimization strategies and demonstrate the parallel scaling advantages of some of these variants on NUMA multicore nodes. Shifted, fused, and communication-avoiding variants for 1283 boxes result in close to ideal parallel scaling and come close to matching the performance of 163 boxes on three different multicore systems for a benchmark that is a proxy for program idioms found in Computational Fluid Dynamic (CFD) codes.

AB - Structured-grid PDE solver frameworks parallelize over boxes, which are rectangular domains of cells or faces in a structured grid. In the Chombo framework, the box sizes are typically 163 or 323, but larger box sizes such as 1283 would result in less surface area and therefore less storage, copying, and/or ghost cells communication overhead. Unfortunately, current on node parallelization schemes perform poorly for these larger box sizes. In this paper, we investigate 30 different inter-loop optimization strategies and demonstrate the parallel scaling advantages of some of these variants on NUMA multicore nodes. Shifted, fused, and communication-avoiding variants for 1283 boxes result in close to ideal parallel scaling and come close to matching the performance of 163 boxes on three different multicore systems for a benchmark that is a proxy for program idioms found in Computational Fluid Dynamic (CFD) codes.

UR - http://www.scopus.com/inward/record.url?scp=84979764508&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=84979764508&partnerID=8YFLogxK

U2 - 10.1109/SC.2014.70

DO - 10.1109/SC.2014.70

M3 - Conference article

AN - SCOPUS:84979764508

SN - 2167-4329

VL - 2015-January

SP - 793

EP - 804

JO - International Conference for High Performance Computing, Networking, Storage and Analysis, SC

JF - International Conference for High Performance Computing, Networking, Storage and Analysis, SC

IS - January

M1 - 7013052

T2 - International Conference for High Performance Computing, Networking, Storage and Analysis, SC 2014

Y2 - 16 November 2014 through 21 November 2014

ER -

A Study on Balancing Parallelism, Data Locality, and Recomputation in Existing PDE Solvers

Abstract

ASJC Scopus subject areas

Access to Document

Other files and links

Fingerprint

Cite this