PIConGPU: Unleashing the Full Computational Potential for the Many-Core Era


PIConGPU: Unleashing the Full Computational Potential for the Many-Core Era

Huebl, A.; Widera, R.; Pausch, R.; Worpitz, B.; Eckert, C.; Burau, H.; Garten, M.; Debus, A.; Kluge, T.; Bussmann, M.

Since its release as open source in 2013, PIConGPU is the fastest published 3D3V Particle-in-Cell (PIC) code in the world in terms of sustained peak performance with 7.2 Pflop/s scaling up to 18'432 GPUs on Titan (ORNL). Accelerator hardware is the key technology enabling an order-of-magnitude increase in computational power over conventional CPUs, but on the same time requires a general rethinking of particle-mesh and particle-particle algorithms in terms of multi-level parallelism.
We present the challenges that are common to all PIC codes in a heterogeneous computing environment and possible solutions in PIConGPU. Starting from a general description of mesh-based operations over communication and latency hiding down to efficient caching and register usage, a sustainable programming technique is explained that is both interchangeable in algorithms and performance portable.
The continuing trend of steady increase in theoretical peak performance for the world's leading machines diverges significantly from the bandwidths that are available for high-performance file systems, causing substantial change in established imulation and analysis chains. In-situ and staged processing are approaches to bridge that gap and will be presented on routines that are either memory limited, computationally highly expensive or communication bound.
Quantitatively, a dramatically lowered time-to-solution is the direct advantage of a many-core accelerated based PIC code. The former is indispensable for an equally significant, qualitative scientific improvement that allows to incorporate multi-physics models that are beyond the simple averaging over ensembles, e.g., kinetic collision and non-LTE ionization models. The potential impact for laser-ion acceleration on solid density targets will be illustrated in an example.

Keywords: HPC; PIConGPU; LPA; GPU; Simulation; PIC

  • Invited lecture (Conferences)
    ICNSP 2015 - International Conference on Numerical Simulation of Plasmas, 12.-14.08.2015, Golden (CO), USA

Permalink: https://www.hzdr.de/publications/Publ-22387
Publ.-Id: 22387