Efficient parallelization for 3d-3v sparse grid Particle-In-Cell: Shared memory architectures

Efficient parallelization for 3D-3V sparse grid Particle-In-Cell: Single GPU architectures

“…The present paper extends the works of [16,17] to the parallelization of 3D-3V sparse-PIC methods on GPU architectures. Although GPUs offer a limited amount of memory in comparison to CPUs, this is not an issue for sparse-PIC implementations thanks to the substantial gains upon the memory footprint.…”

“…for l ∈ are called the combination coefficients and ℎ l defines the interpolation onto the space ℎ l . In this paper, we consider an interpolation in the hierarchical basis of the space ℎ l , where a transformation from the nodal basis to a hierarchical basis is performed (see [17] for details). The rest of the scheme is similar to standard PIC methods and the sparse-PIC scheme is summarized in algorithm 1.…”

“…Recently, sparse-PIC method optimization and parallelization have been investigated in a shared memory architecture framework [17]. On the one hand, sparse-PIC methods have been proven to be particularly memory efficient, with respect to the size of the grids used to accumulate the density and compute the electric field, as well as the storage of the particles properties.…”

“…The efficient implementation introduced in this paper is compared to another one, which is merely the extension to GPU of sparse-PIC strategies introduced for shared memory CPU architecture in [17]. The novel implementation combines two level of parallelism for the charge density accumulation: a coarse-grain parallelism based on a particle population decomposition and a Single Instruction Multiple Data (SIMD) parallelism based on a component grid work sharing.…”

“…A brief overview of the hardware architecture and usual GPU programming is provided with the aim to outline the algorithmic issues to achieve performance. Then, two implementations of the sparse-PIC method are presented: an extension to GPU of the parallel version designed for shared memory architectures, which has been introduced in [17], and a version tailored specifically for GPUs. Finally in section 4, the efficiency of the implementations are assessed on three dimensional classical test cases: the non-linear Landau damping and the diocotron instability using two different hardware: a laptop GPU (Quadro T2000) and a Tesla V100.…”