SMuK 2023 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
P: Fachverband Plasmaphysik
P 12: Poster II
P 12.32: Poster
Mittwoch, 22. März 2023, 17:30–19:00, HSZ EG
Enabling GENE for Exascale Computing via Modern Data Science — •Luciana Tanzarella1, Tilman Dannert2, Tobias Görler1, and Frank Jenko1 — 1Max Planck Institute for Plasma Physics, Garching — 2Max Planck Computing and Data Facility, Garching
Theoretical plasma turbulence studies are typically based on numerical solutions of integro-differential equations in 5 or 6 dimensions over a very large parameter space. So the required computational power is huge. Higher speed, lower communication and energy costs are all benefits of lower precision arithmetic, but the outputs must be accurately assessed. Particular focus is put on the world-leading gyrokinetic plasma turbulence code GENE, based on the Eulerian approach. GENE allows for either single or double precision computations.The national DaREXA-F project’s specific objective is to develop methods to reduce the amount of data for transfer operations and leverage the power of reduced precision arithmetics on modern architectures. The first steps in this direction entail the creation and application of lower precision methods in selected operations performed by GENE. The precision must be scaled using existing libraries in addition to assessing how it impacts calculations as not every hardware supports arbitrary precision. To ascertain how much the discretization order influences the outputs on grids, this must be done on each of GENE’s several sections. The operation will be done on the stencil part first. In order to accurately evaluate the impact of implementing this part with reduced precision, a respective error model has to be developed.