A new GPU implementation for lattice-Boltzmann simulations on sparse geometries

Abstract

We describe a high-performance implementation of the lattice Boltzmann method (LBM) for sparse 3D geometries on graphic processors (GPU). The main contribution of this work is a data layout that allows to minimise the number of redundant memory transactions during the propagation step of LBM. We show that by using a uniform mesh of small three-dimensional tiles and a careful data placement it is possible to utilise more than 70% of maximum theoretical GPU memory bandwidth for D3Q19 lattice and double precision numbers. The performance of our implementation is thoroughly examined and compared with other GPU implementations of LBM. The proposed method performs the best for sparse geometries with good spatial locality.

Esoteric Twist: An Efficient in-Place Streaming Algorithmus for the Lattice Boltzmann Method on Massively Parallel Hardware

“…Our results show comparable performance in terms of memory bandwidth (up to ∼68.2% peak bandwidth) to a block structured code with two distributions on recent Nvidia GPUs [36] and to full matrix codes. In terms of execution time, the sparse matrix EsoTwist method was found to obtain 99.4% of the performance of the AB-pattern on a recent GPU.…”

“…For the larger pipe oriented along the x-axis, we obtain 738.5 MNUPS (59.1% peak bandwidth); for the larger pipe oriented along the y-axis, we obtain 729.8 MNUPS (58.4% peak bandwidth) and, for the z direction, we obtain 727.3 MNUPS (58.2% peak bandwidth). Our performance figures compare to between 407 and 684 MNUPS reported for the GeForce GTX Titan by Tomczak and Szafran [36] using a lattice Boltzmann method with nineteen speeds (D3Q19 lattice) in double precision in a tiling layout (similar to block structuring) and two data sets (AB-pattern), which corresponds to between 48% and 72.6% bandwidth (the numbers are taken directly from [36] as our performance model does not apply directly to their code). The relatively wide range in the performance of [36] is due to differences in the data locality in the different test cases.…”

GPU Acceleration of the HemeLB Code for Lattice Boltzmann Simulations in Sparse Complex Geometries

“…A number of high-performance lattice Boltzmann codes have been developed by various groups, including Palabos [15], waLBerla [16], MUPHY [17], and HARVEY [18]. With the availability of application programming interfaces for general-purpose graphics processing units, there has been increasing interest in GPU implementations of the LBM [19], [20], [21], [22], [23], [14], [24]. These efforts address a variety of aspects including efficient data layouts [22], [14], indirect addressing solutions [19], [23], and multi-GPU implementations [25], [21], [17], [23], [26], [27].…”

A scalable parallel unstructured finite volume lattice Boltzmann method for three‐dimensional incompressible flow simulations

“…A number of works on the parallel LBM have been carried out on multi-CPUs 21,22 and multi-GPUs. [23][24][25][26][27][28] However, all of them were conducted on uniform Cartesian grids. In this paper, a parallel 3D FV-LBM on the unstructured grids is developed and analyzed on the Tianhe-2A supercomputer at the National Supercomputing Center in Guangzhou, China.…”