Open AccessPerformance benchmarks for a next generation numerical dynamo model
Author(s) -
Matsui Hiroaki,
Heien Eric,
Aubert Julien,
Aurnou Jonathan M.,
Avery Margaret,
Brown Ben,
Buffett Bruce A.,
Busse Friedrich,
Christensen Ulrich R.,
Davies Christopher J.,
Featherstone Nicholas,
Gastine Thomas,
Glatzmaier Gary A.,
Gubbins David,
Guermond JeanLuc,
Hayashi YoshiYuki,
Hollerbach Rainer,
Hwang Lorraine J.,
Jackson Andrew,
Jones Chris A.,
Jiang Weiyuan,
Kellogg Louise H.,
Kuang Weijia,
Landeau Maylis,
Marti Philippe,
Olson Peter,
Ribeiro Adolfo,
Sasaki Youhei,
Schaeffer Nathanaël,
Simitev Radostin D.,
Sheyko Andrey,
Silva Luis,
Stanley Sabine,
Takahashi Futoshi,
Takehiro Shinichi,
Wicht Johannes,
Willis Ashley P.
Publication year - 2016
Publication title -
geochemistry, geophysics, geosystems
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.928
H-Index - 136
ISSN - 1525-2027
DOI - 10.1002/2015gc006159
Subject(s) - dynamo , dynamo theory , computer science , scaling , computational science , earth's magnetic field , range (aeronautics) , observable , statistical physics , field (mathematics) , computational physics , algorithm , magnetic field , physics , aerospace engineering , mathematics , geometry , quantum mechanics , pure mathematics , engineering
Numerical simulations of the geodynamo have successfully represented many observable characteristics of the geomagnetic field, yielding insight into the fundamental processes that generate magnetic fields in the Earth's core. Because of limited spatial resolution, however, the diffusivities in numerical dynamo models are much larger than those in the Earth's core, and consequently, questions remain about how realistic these models are. The typical strategy used to address this issue has been to continue to increase the resolution of these quasi‐laminar models with increasing computational resources, thus pushing them toward more realistic parameter regimes. We assess which methods are most promising for the next generation of supercomputers, which will offer access to O (10 6 ) processor cores for large problems. Here we report performance and accuracy benchmarks from 15 dynamo codes that employ a range of numerical and parallelization methods. Computational performance is assessed on the basis of weak and strong scaling behavior up to 16,384 processor cores. Extrapolations of our weak‐scaling results indicate that dynamo codes that employ two‐dimensional or three‐dimensional domain decompositions can perform efficiently on up to ∼10 6 processor cores, paving the way for more realistic simulations in the next model generation.