Parts | Days | Selection | Search | Updates | Downloads | Help
MM: Fachverband Metall- und Materialphysik
MM 15: Poster Session
MM 15.61: Poster
Monday, March 11, 2013, 18:00–20:00, Poster E
KKRnano: Improvements in efficiency and order-N scaling for use on massively parallel computers — •Elias Rabel, Rudolf Zeller, and Stefan Blügel — Peter Grünberg Institut and Institute for Advanced Simulation, Forschungszentrum Jülich and JARA, D-52425 Jülich
Petascale high-performance computers open completely new opportunities for the first-principles treatment of materials problems. The recently developed code KKRnano [1] implements the screened KKR Green’s function method for DFT calculations of unit cells with thousands of atoms. It targets modern supercomputing architectures, which requires being able to utilise more than 100000 processors.
The KKR method considers multiple scattering of partial waves up to a maximum angular momentum ℓmax. We investigate the structure of matrices with sparse block structure occuring in the treatment of very large unit cells. A distance dependent angular momentum cutoff is employed to reduce the effective matrix size with little compromise in accuracy. The special case of neglecting all partial waves at reasonably large distances leads to an O(N)-scaling method.
Usage of an iterative solver necessitates an appropriate preconditioning strategy. A block-circulant preconditioner has been proven to be highly effective in the case of uniform block sizes. Solution strategies for the case of the distance dependent ℓ-cutoff, where one has to deal with variable block sizes, are currently under investigation.
[1] A. Thiess et al., Phys. Rev. B 85, 235103 (2012)