Presentation
IPPL: A Massively Parallel, Performance Portable C++ Library for Particle-Mesh Methods and Efficient Solvers
Presenter
DescriptionWe present the Independent Parallel Particle Layer (IPPL), a performance portable C++ library for particle-in-cell methods. IPPL makes use of Kokkos (a performance portability abstraction layer), HeFFTe (a library for large scale FFTs), and MPI (Message Passing Interface) to deliver a portable, massively parallel toolkit for particle-mesh methods. One of the advantages of such a framework is the ability to be a test-bed for new algorithms which seek to improve runtime and efficiency of large scale simulations, for example in the beam and plasma physics communities. Concretely, we have implemented an efficient and portable free-space solver for the Poisson equation based on the algorithm suggested by Vico et al. (2016). This fast solver has spectral convergence, as opposed to the second-order convergence of the state-of-the-art Hockney-Eastwood method. The ability to use coarser grids to achieve similar resolutions with the new solver allows for higher resolution simulations with a lower memory footprint, which is especially important for GPU usage. Finally, we show scaling studies on the Perlmutter machine at NERSC, on both CPUs and GPUs, with efficiencies staying above 50% in the strong scaling case.
TimeMonday, June 312:30 - 13:00 CEST
LocationHG F 26.3
Session Chair
Event Type
Minisymposium
Engineering
Physics
Computational Methods and Applied Mathematics
