Scalable lattice Boltzmann solvers for CUDA GPU clusters

•An MPI-CUDA implementation of the lattice Boltzmann method is described.•We propose a method to handle efficiently 3D partitions of the simulation domain.•We study the performance of our implementation on a cluster using up to 24GPUs.•Peak performance as well as weak and strong scalability are sati...

Full description

Saved in:
Bibliographic Details
Published in:Parallel computing 2013-06, Vol.39 (6-7), p.259-270
Main Authors: Obrecht, Christian, Kuznik, Frédéric, Tourancheau, Bernard, Roux, Jean-Jacques
Format: Article
Language:English
Subjects:
Clusters
Computation
Computational fluid dynamics
Computer Science
CUDA
GPU clusters
Kernels
Lattice Boltzmann method
Lattices
Networking and Internet Architecture
Performance measurement
Solvers
Three dimensional
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:•An MPI-CUDA implementation of the lattice Boltzmann method is described.•We propose a method to handle efficiently 3D partitions of the simulation domain.•We study the performance of our implementation on a cluster using up to 24GPUs.•Peak performance as well as weak and strong scalability are satisfactory. The lattice Boltzmann method (LBM) is an innovative and promising approach in computational fluid dynamics. From an algorithmic standpoint it reduces to a regular data parallel procedure and is therefore well-suited to high performance computations. Numerous works report efficient implementations of the LBM for the GPU, but very few mention multi-GPU versions and even fewer GPU cluster implementations. Yet, to be of practical interest, GPU LBM solvers need to be able to perform large scale simulations. In the present contribution, we describe an efficient LBM implementation for CUDA GPU clusters. Our solver consists of a set of MPI communication routines and a CUDA kernel specifically designed to handle three-dimensional partitioning of the computation domain. Performance measurement were carried out on a small cluster. We show that the results are satisfying, both in terms of data throughput and parallelisation efficiency.
ISSN:0167-8191
1872-7336
DOI:10.1016/j.parco.2013.04.001