Paul Fischill

Sriramkrishnan Muralikrishnan, Paul Fischill, Andreas Adelmann et al.

We present and compare distributed parallelization strategies for the particle-in-Fourier (PIF) schemes used in kinetic plasma simulations. The different strategies are i) domain decomposition, where both the particles and Fourier modes are split between the MPI ranks ii) particle decomposition, where only the particles are split between the ranks and each rank carries all the modes, and, iii) space-time decomposition, in which time parallelization based on the parareal algorithm is added on top of the particle decomposition. We describe the different communication patterns involved in each of the strategies, the parameter regimes where they work best, and explain their advantages and disadvantages. We implement the strategies within the open-source, performance portable library IPPL and conduct scaling studies with 3D-3V Landau damping and Penning trap benchmark problems on Alps and JUWELS booster supercomputers. We analyze the dominant component timings in each of the strategies and identify areas for future optimizations.

Sonali Mayani, Paul Fischill, Sriramkrishnan Muralikrishnan et al.

We compare different Poisson solvers within the context of an electrostatic Vlasov-Poisson system. These schemes are implemented as part of the IPPL (Independent Parallel Particle Layer) library (Frey et al., 2024), which provides performance portable and dimension independent building blocks for scientific simulations requiring particle-mesh methods, with Eulerian (mesh-based) and Lagrangian (particle-based) approaches. The simulation used to compare the performance and portability of the schemes is Landau damping, part of a set of mini-applications implemented to benchmark and showcase the capabilities of the IPPL library (Muralikrishnan et al., 2024). We use grid-sizes of $512^3$ and $1024^3$ with 8 particles per cell, running with different algorithms in the solve phase of the Particle-in-Cell (PIC) loop: a Fast Fourier Transform (FFT) pseudo-spectral solver, a matrix-free finite difference Preconditioned Conjugate Gradient (PCG) solver, and a matrix-free Finite Element (FEM) solver. We also compare these PIC schemes to the novel Particle-in-Fourier (PIF) scheme, which performs interpolations using non-uniform FFTs thereby avoiding a grid in the real space. We obtain results on different computing architectures, such as AMD GPUs (LUMI at CSC), and Nvidia GPUs (Alps at CSCS and JUWELS Booster at Jülich Supercomputing Center), showcasing portability. In terms of absolute time the FFT solver is advantageous, but is limited in its applicability. All other field solvers in the PIC scheme are an order-of-magnitude more expensive in terms of time, but scale similarly to the FFT case in the electrostatic PIC context. The PIF scheme serves as a high fidelity alternative to standard PIC, and while it is costlier than the FFT-based PIC scheme, it shows excellent scalability on all the architectures.

Paul Fischill, Andreas Adelmann, Sriramkrishnan Muralikrishnan

The nonuniform fast Fourier transform (NUFFT) enables spectral methods for problems with irregularly spaced samples, with applications in medical imaging, molecular dynamics, and kinetic plasma simulations. Existing implementations are limited to shared-memory execution, restricting problem sizes to what fits on a single node. We present the first distributed, performance-portable NUFFT for heterogeneous supercomputers. Our Kokkos-based implementation runs without modification on NVIDIA and AMD GPUs. We develop multiple spreading and interpolation kernels optimized for different accuracy requirements and architectures. Our spreading kernels match or exceed the single-GPU throughput of the state-of-the-art CUDA-based NUFFT library cuFINUFFT at production particle densities, while our Kokkos-based implementation additionally supports AMD GPUs. Strong scaling experiments on Alps (NVIDIA GH200), JUWELS Booster (NVIDIA A100), and LUMI (AMD MI250X) demonstrate scaling up to 1024 GPUs. At scale, the distributed FFT is a significant part of the total runtime, making higher NUFFT accuracy less expensive. We apply the method to massively parallel Particle-in-Fourier simulations of Landau damping with up to $1024^3$ Fourier modes and 8.6 billion particles on Alps, JUWELS, and LUMI, demonstrating that distributed NUFFTs enable kinetic plasma simulations at resolutions previously inaccessible to spectral particle methods.