GPU acceleration of four-way coupled PP-DNS for compressible particle-laden wall turbulence

Abstract

This paper presents an efficient implementation of the four-way coupled point-particle direct numerical simulation (PP-DNS) for compressible particle-laden wall turbulence, utilizing the open-source finite-difference compressible Navier–Stokes solver, STREAmS. The proposed design integrates a GPU-based two-phase collision detection algorithm known as the spatial subdivision method, along with specialized storage and MPI communication strategies for Lagrangian particles on multi-GPU platforms. Specifically, a ‘page table’ like data structure is designed to store the particle information compactly and to enable highly parallelized packing and unpacking procedures for GPU-GPU data exchange. These advancements significantly reduce the computational cost of four-way coupled particle-laden flow simulations, enabling efficient simulations involving over $O\left(10^7\right)$ particles (an order of magnitude higher than that in the state-of-the-art simulations) on a single NVIDIA A100 GPU. To validate the proposed implementation, we perform simulations of compressible particle-laden wall-bounded turbulence using canonical configurations such as channel flows and zero-pressure-gradient boundary layers. The example results highlight the effects of inter-particle collisions and flow compressibility. Furthermore, we assess single-GPU performance and scalability by employing up to eight NVIDIA GPU devices. Even for four-way coupled simulations, the elapsed time per step scales approximately linearly with the number of particles (when the number of particles is large enough), and a parallel efficiency of 94.1% is achieved on 8 NVIDIA A100 GPUs.