Adaption and GPU based parallelization of the code TEMDDD for the 3D modelling of CSEM data.

Sommer, Malte and Hölz, Sebastian (2011) Adaption and GPU based parallelization of the code TEMDDD for the 3D modelling of CSEM data. [Poster] In: 71. Jahrestagung der Deutschen Geophysikalischen Gesellschaft gemeinsam mit der Arbeitsgemeinschaft Extraterrestrische Forschung. , 21.02.- 24.02.2011, Köln . 71. Jahrestagung der Deutschen Geophysikalischen Gesellschaft gemeinsam mit der Arbeitsgemeinschaft Extraterrestrische Forschung und dem Fachverband Extraterrestrische Physik der Deutschen Physikalischen Gesellschaft : 21.–24. Februar 2011 in Köln. ; p. 40 .

Full text not available from this repository. (Contact)

Abstract

For the interpretation of 3D marine CSEM data collected by the IFM-GEOMAR, we investigate the applicability of the 3D FD transient time domain code TEMDDD (K. Arnason). The code is adapted to massively parallel GPU (graphics processing unit) architecture to improve its performance, which for typical model sizes is in the order of 10-15 min on current CPUs. In the original code the spatial operators of the diffusion equation are discretized on a staggered grid (Yee, 1966), resulting in a large operator matrix. The large, sparse and banded volume matrix is solved by the Spectral Lanczos Decomposition Method. The tridiagonal, symmetric eigenvalue problem is solved via QR-decomposition. The dense surface matrix is reduced by the Householder Method to tridiagonal form, and consecutively solved as part of the volume matrix. More than 90% of the code’s runtime constitutes to the eigensolver and matrix reductions. The GPU based implementation of the Householder Method (30% time consumption in original code) shows - depending on matrix size - a speed up by a factor of 8x - 13x. The originally used QR-eigensolver was replaced by bisection on Gershgorin cycles to determine the eigenvalues and an inverse iteration to determine the corresponding eigenvectors. While the parallel cyclic reduction algorithm has turned out to be powerful on GPUs, the numerical stability of the dependent linear solver in the inverse iteration remains problematic. Solving this issue as well as the implementation of the SLDM algorithm are work in progress.

Document Type: Conference or Workshop Item (Poster)
Keywords: Meeresgeologie; Geodynamics
Research affiliation: OceanRep > GEOMAR > FB4 Dynamics of the Ocean Floor > FB4-GDY Marine Geodynamics
Open Access Journal?: Yes
Related URLs:
Date Deposited: 06 Jan 2012 12:17
Last Modified: 23 Feb 2012 06:03
URI: http://oceanrep.geomar.de/id/eprint/13302

Actions (login required)

View Item View Item