Towards accelerating Smoothed Particle Hydrodynamics simulations for free-surface flows on multi-GPU clusters

TL;DR

This paper presents a multi-GPU SPH simulation framework for free-surface flows that significantly accelerates computations and handles larger particle counts than single-GPU systems, using spatial decomposition and MPI communication.

Contribution

It introduces a multi-GPU implementation of SPH with a novel domain decomposition and particle migration method, enabling faster and larger-scale simulations.

Findings

Achieves significant speedups over single-GPU implementations.

Supports simulations with up to 32 million particles.

Analyzes scalability and identifies computational bottlenecks.

Abstract

Starting from the single graphics processing unit (GPU) version of the Smoothed Particle Hydrodynamics (SPH) code DualSPHysics, a multi-GPU SPH program is developed for free-surface flows. The approach is based on a spatial decomposition technique, whereby different portions (sub-domains) of the physical system under study are assigned to different GPUs. Communication between devices is achieved with the use of Message Passing Interface (MPI) application programming interface (API) routines. The use of the sorting algorithm radix sort for inter-GPU particle migration and sub-domain halo building (which enables interaction between SPH particles of different subdomains) is described in detail. With the resulting scheme it is possible, on the one hand, to carry out simulations that could also be performed on a single GPU, but they can now be performed even faster than on one of these devices alone. On the other hand, accelerated simulations can be performed with up to 32 million particles on the current architecture, which is beyond the limitations of a single GPU due to memory constraints. A study of weak and strong scaling behaviour, speedups and efficiency of the resulting programis presented including an investigation to elucidate the computational bottlenecks. Last, possibilities for reduction of the effects of overhead on computational efficiency in future versions of our scheme are discussed.