Smoothed-Particle Hydrodynamics Models: Implementation Features on GPUs

TL;DR

This paper presents a GPU-accelerated implementation of Smoothed Particle Hydrodynamics for self-gravitating gas dynamics, demonstrating efficient parallelization on Nvidia Tesla GPUs for galactic collision simulations.

Contribution

It introduces a hybrid OpenMP-CUDA implementation of SPH with hierarchical grid sorting for neighbor list creation, optimized for multi-GPU systems.

Findings

High parallelization efficiency on Nvidia Tesla GPUs

Effective neighbor list creation using hierarchical grid sorting

Successful simulation of galactic gaseous halos collisions

Abstract

Parallel implementation features of self-gravitating gas dynamics modeling on multiple GPUs are considered applying the GPU-Direct technology. The parallel algorithm for solving of the self-gravitating gas dynamics problem based on hybrid OpenMP-CUDA parallel programming model has been described in detail. The gas-dynamic forces are calculated by the modified SPH-method (Smoothed Particle Hydrodynamics) while the N-body problem gravitational interaction is obtained by the direct method (so-called Particle-Particle algorithm). The key factor in the SPH-method performance is creation of the neighbor lists of the particles which contribute into the gas-dynamic forces calculation. Our implementation is based on hierarchical grid sorting method using a cascading algorithm for parallel computations of partial sums at CUDA block. The parallelization efficiency of the algorithm for various GPUs of the Nvidia Tesla line (K20, K40, K80) is studied in the framework of galactic' gaseous halos collisions models by the SPH-method.