DPG Phi
Verhandlungen
Verhandlungen
DPG

Berlin 2024 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

O: Fachverband Oberflächenphysik

O 38: Poster: Electronic Structure Theory

O 38.2: Poster

Dienstag, 19. März 2024, 18:00–20:00, Poster D

Accelerating plane-wave-based ab initio molecular dynamics by optimization of Fast-Fourier transforms for modern HPC architectures — •Christian Ritterhoff, Tobias Klöffel, Sagarmoy Mandal, and Bernd Meyer — Interdisciplinary Center for Molecular Materials and Computer Chemistry Center, FAU Erlangen-Nürnberg, Germany

The most important advantage of plane-wave basis sets is that wave functions can be transformed efficiently from reciprocal to real space and back by using the Fast-Fourier transform (FFT) algorithm. This allows to evaluate the kinetic and potential energy in reciprocal and real space, respectively, where both operators are diagonal. This reduces the computational cost for applying the Hamilton operator from N2 to N logN. However, the scalability of current FFT libraries is rather limited on today’s HPC systems, which offer large numbers of compute nodes, each of them with many cores. Here we present our optimization of the FFT routines in the Car-Parrinello molecular dynamics code CPMD (www.cpmd.org). Data distribution and communication patterns have been revised to make optimal use of combined MPI and OpenMP parallelization. Scalability is further increased by combining FFTs into batches and by introducing overlapping computation and communication. The improved performance of the new FFT routines will be demonstrated by a series of benchmark simulations with our optimized version of the CPMD code [1].


[1] T. Klöffel, G. Mathias, B. Meyer, Comput. Phys. Commun. 260 (2021) 107745

Keywords: fast Fourier transform (FFT); plane-wave DFT codes; Car-Parrinello molecular dynamics; high performance computing

100% | Mobil-Ansicht | English Version | Kontakt/Impressum/Datenschutz
DPG-Physik > DPG-Verhandlungen > 2024 > Berlin