 |
| Titel |
Aggressive, accelerated subdomain smoothers for Stokes flow with highly heterogeneous viscosity structure |
| VerfasserIn |
Patrick Sanan, Dave May, Olaf Schenk, Karl Rupp |
| Konferenz |
EGU General Assembly 2016
|
| Medientyp |
Artikel
|
| Sprache |
en
|
| Digitales Dokument |
PDF |
| Erschienen |
In: GRA - Volume 18 (2016) |
| Datensatznummer |
250125514
|
| Publikation (Nr.) |
 EGU/EGU2016-5102.pdf |
|
|
|
|
|
| Zusammenfassung |
| Scalable solvers for mantle convection and lithospheric dynamics with
highly heterogeneous viscosity structure typically require the use of
a multigrid method. To leverage new hybrid CPU-accelerator
architectures on leadership compute clusters, multigrid hierarchies
which can reduce communication and use high available arithmetic
intensity are at a premium, motivating more aggressive coarsening
schemes and smoothers.
We present results of a comparative study of two competitive
GPU-enabled subdomain smoothers within an additive Schwarz method.
Chebyshev-Jacobi smoothing has been shown to be an effective smoother,
and its nature as a low-communication method built from basic linear
algebra routines allows its use on a wide range of devices with
current libraries. ILU smoothing is also of interest and is known to
provide robust smoothing in some cases, but has traditionally been
difficult to use in a fine-grained parallel environment. However, a
recently-introduced variant by Chow and Patel allows for incomplete
factorizations to be computed and applied in these environments, hence
allowing us to study them as well. We use and extend the pTatin3D,
PETSc, and ViennaCL libraries to integrate promising methods into a
realistic application framework. |
|
|
|
|
|
|