Publikationsrepositorium - Helmholtz-Zentrum Dresden-Rossendorf

A scalable, performant, highly-parallel Particle-in-Cell code for fast simulations of large Laser-Plasma experiments. Investigating parameters for optimizing laser particle acceleration is a timeconsuming task since realistic simulations of laser plasma interactions using the particle-in-cell technique can require the computation of several hundred million particle trajectories on a grid of several ten million cells. The computational effort needed to investigate the dependence of the performance of new acceleration scenarios on only a few parameters thus normally requires the use of large-scale high-performance computing systems only available at central super computing centres. Thus, parameter scans are usually performed by reducing the system size, the particle density, the computation time and the dimensionality of the problem. Such a scan is then at best complimented by a small number of more realistic large-scale simulations with parameters closer to the experimental parameters. Recently, general purpose graphical processing units (GPGPUs) have entered the stage of high performance computing. This new hardware offers a computational power exceeding that of standard CPU-based computers by several orders of magnitude at much lower investment and maintenance costs. Making good use of this computational power is only possible if the algorithm can run on a massively parallel system consisting of a huge number of independently working processors. However, the memory on a single GPGPU and thus the system size that can be computed on it is limited We present PIConGPU [1], a particle-in-cell algorithm that can run efficiently on a cluster of GPGPU nodes. PIConGPU can run largescale, realistic simulations by mapping the physical system onto many GPGPUs. Thus, the time needed to calculate the evolution of the large system is comparable to the time it takes to compute the small sub-region that can fit on a single GPU and therefore can lead to turnaround times of only a few hours for a hundred thousand time steps and single time steps of under a nanosecond per macro-particle [2].
If computational stability and dispersion is treated appropriately, using GPGPUs to simulate for example laser wakefield acceleration of electrons can greatly enhance the study of large parameter spaces while at the same time using simulation parameters resembling those of the experimental system studied. We focus on real-world examples of using PIConGPU for the simulation of laser electron acceleration scenarios investigated with the DRACO laser system at the Forschungszentrum Dresden-Rossendorf and show how the fast response time of GPGPU-based simulations can open up the path for optimizing experimental parameters.