Parts | Days | Selection | Search | Updates | Downloads | Help
MM: Fachverband Metall- und Materialphysik
MM 23: Methods in Computational Materials Modelling (methodological aspects, numerics)
MM 23.11: Talk
Wednesday, April 3, 2019, 13:00–13:15, H44
Parallelization and Acceleration of the FLEUR Code: New Possibilities for All-electron Density Functional Theory — •Uliana Alekseeva, Daniel Wortmann, and Stefan Blügel — Institute for Advanced Simulation and Peter Grünberg Institut, Forschungszentrum Jülich and JARA, 52425 Jülich, Germany
The perpetual evolution of the high performance computing (HPC) hardware is a considerable challenge for scientific software, which has often been developed over decades. In particular, the need to adapt for new hardware can require substantial changes in the data layout, the parallelization strategy and the fundamental algorithms. The interplay of different functionalities amplifies the redesign challenge.
An example of such a scientific software package is the all-electron DFT code FLEUR developed at the Forschungszentrum Jülich (http://www.flapw.de). While its main field of application is the simulation of complex magnetic structures, the employed full-potential linearized augmented plane wave (FLAPW) scheme ensures a general applicability for all bulk and film setups. We will present our experiences with implementing the hybrid MPI+OpenMP parallelization, the challenges arising from finding an efficient algorithm and from optimizing the performance. In addition, we will report on porting of the code to NVIDIA GPUs. Finally, we demonstrate that the improved scalability of the code allows calculations for unit cells with over 1000 atoms.
Our work was supported by the MaX European Center of Excellence in HPC (H2020-EINFRA-2015-1 project: GA 676598).