high performance computing on graphics processing units: hgpu.org

Parallel programming is increasingly used to improve the performance of solving numerical methods used for scientific purposes. Numerical methods in the field of fluid dynamics require the calculation of a large number of operations per second. One of the methods that is easily parallelized and often used is the Lattice Boltzmann method (LBM). Today, it is possible to perform simulations of numerical methods not only on high performance computers (HPC) but also on commodity computers. In this paper is presented how to accelerate LBM implementation on commodity computers using characteristics of OpenCL specification. Simulation is executed simultaneously on multiple heterogeneous devices. Four different approaches for several commodity computer configurations are presented. Obtained results are compared for different types of commodity computers and advantages and disadvantages are discussed. In this paper it presented which LBM OpenCL code implementation, among four different presented, shows best simulation performance and should be used when solving similar CFD problems.