Berlin 2014 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
P: Fachverband Plasmaphysik
P 27: Theory of nonideal Plasmas
P 27.8: Vortrag
Freitag, 21. März 2014, 12:45–13:00, SPA HS201
Unleashing PFlop/s for Plasma Science with PIConGPU — •Axel Huebl1, René Widera1, Heiko Burau1, Felix Schmitt2, Richard Pausch1, Benjamin Schneider1, Guido Juckeland2, and Michael Bussmann1 — 1Helmholtz-Zentrum Dresden - Rossendorf, Dresden, Deutschland — 2Zentrum für Informationsdienste und Hochleistungsrechnen, TU Dresden, Deutschland
PIConGPU is a massively parallel open source particle-in-cell (PIC) code written for CUDA capable graphics cards (GPUs). PIC codes are used to model fully relativistic electro-magnetic interactions between particles and discrete mesh points (fields) in a self-consistent manner. That allows for ab-intio simulations ranging from astrophysical scenarios like emerging plasma instabilities to modern laser-driven particle accelerators like the laser-wakefield accelerator (LWFA).
In this talk, we will present the strategy to port a non-trivial algorithm to a modern, massively parallel hardware architecture like GPUs. We give a short introduction in our free multi-GPU framework libPMacc for general mesh and particle based simulations. Based on that, scalings up to 18,000 GPUs on the Titan supercomputer (Oak Ridge National Lab) will be discussed.
Applications will include results of in-situ far field radiation diagnostics for the relativistic Kelvin-Helmholtz-Instability (KHI) by evolving billions of electrons and calculating their individual Liénard-Wiechert potential for arbitrary directions and frequencies on the fly.