Loaded and Unlocked - Accelerating OpenSource Plasma Physics

Loaded and Unlocked - Accelerating OpenSource Plasma Physics

Bussmann, M.; Burau, H.; Cowan, T. E.; Debus, A.; Garten, M.; Helm, A.; Huebl, A.; Juckeland, G.; Kluge, T.; Pausch, R.; Schmitt, F.; Schneider, B.; Schramm, U.; Steiniger, K.; Widera, R.

PIConGPU is a fully relativistic 3D3V particle-in-cell (PIC) code for studying laser-plasma interactions. Todays graphics processing units (GPUs) are massively-parallel accelerators for scientific computing, pushing the limits for a new era of in situ plasma simulations.

During the last decades PIC codes became the workhorses for theoretical studies in laser-particle acceleration. Besides the ongoing demand for faster and more realistic simulations, interaction with the simulation and handling of peta-bytes of generated data per simulation are two of the challenging topics in high performance computing.

Moreover, we present a general explicit scheme to load charged relativistic particle beams in the PIC cycle. Taking care of the particle shape in combination with a discretization-dependent potential solver for one-time initialization allows for ab-initio conservation of Gauss' law.

PIConGPU is among the 2013 ACM Gordon Bell finalists, reaching the highest performance of a PIC simulation ever reported with 8 PFlop/s mixed precision. Utilizing the world's #2 supercomputer Titan (Oak Ridge National Lab), we present performance benchmarks up to 18,000 GPUs, scaling to a total of 50 million multi-processors.

PIConGPU is developed as open source by the Junior Group for Computational Radiation Physics at the Institute for Radiation Physics at Helmholtz-Zentrum Dresden - Rossendorf (HZDR) in close collaboration with the Center for Information Services and High Performance Computing (ZIH) of the Technical University Dresden (TUD).

Keywords: PIConGPU; LPA; OpenSource; GPU; HPC; Laser-Plasma

  • Invited lecture (Conferences)
    Séminaire LOA, 30.01.2014, Paris, France

Permalink: https://www.hzdr.de/publications/Publ-21326
Publ.-Id: 21326