CUDA Implementation of a Navier-Stokes Solver on Multi-GPU Desktop Platforms for Incompressible Flows

Abstract

Graphics processor units (GPU) that are traditionally designed for graphics rendering have emerged as massively-parallel "co-processors" to the central processing unit (CPU). Small-footprint desktop supercomputers with hundreds of cores that can deliver teraflops peak performance at the price of conventional workstations have been realized. A computational fluid dynamics (CFD) simulation capability with rapid computational turnaround time has the potential to transform engineering analysis and design optimization procedures. We describe the implementation of a Navier-Stokes solver for incompressible fluid flow using desktop platforms equipped with multi-GPUs. Specifically, NVIDIA's Compute Unified Device Architecture (CUDA) programming model is used to implement the discretized form of the governing equations. The projection algorithm to solve the incompressible fluid flow equations is divided into distinct CUDA kernels, and a unique implementation that exploits the memory hierarchy of the CUDA programming model is suggested. Using a quad-GPU platform, we observe two orders of magnitude speedup relative to a serial CPU implementation. Our results demonstrate that multi-GPU desktops can serve as a cost-effective small-footprint parallel computing platform to accelerate CFD simulations substantially.

Real-time flow simulation of indoor environments using lattice Boltzmann method

“…Adding a dynamic sub-grid model and implementing wall functions in LBM which will be done in the future should capture the wall boundary layers more accurately. Since our algorithm is mainly dominated by GPU memory bandwidth (Thibault and Senocak, 2009) adding more computations will not degrade the real-time capability of our method. Furthermore implementation of nonuniform lattice and extending the algorithm to multiple-GPU platform will enable real-time simulation of indoor environments with complicated geometry and large domain sizes.…”

“…We apply an interactive and real-time LBM CFD model with an integrated visualisation tool developed in (Delbosc et al, 2014) to evaluate the suitability, accuracy and usefulness of a 3D LBM based real-time, thermal and turbulent air flow solver running on a GPU platform. The implementation of LBM on the GPU is not unique in the sense that traditional CFD based methods could also be implemented on the GPU (Thibault and Senocak, 2009). But due to the local nature of the LBM algorithm along with the absence of any non-local Poisson pressure loop lends itself to be easily parallelisable compared to traditional CFD methodology on GPUs (Delbosc et al, 2014).…”

OpenACC-based GPU Acceleration of a 3-D Unstructured Discontinuous Galerkin Method

“…To list a few of the CUDAaccelerated CFD applications, Elsen et al 11 reported a 3D high-order FDM solver for large calculation on multi-block structured grids; Klöckner et al 16 developed a 3D unstructured high-order nodal DGM solver for the Maxwell's equations; Corrigan et al 10 proposed a 3D FVM solver for compressible inviscid flows on unstructured tetrahedral grids; Zimmerman et al 29 presented an SDM solver for the Navier-Stokes equations on unstructured hexahedral grids; and more as in the references. 3,4,12,7,21,23,13,19,8,14,1,9 However applying CUDA to a legacy CFD code is not likely an easy job since the developer has to define an explicit layout of the threads on the GPU (numbers of blocks, numbers of threads) for each kernel function. 15 So what if the CFD code designers have to meet specific investment requirements like (1) enable GPU computing for legacy CFD programs at a minimum extra cost in time and effort (usually a major concern for large-scale code development), (2) enable the GPU-accelerated programs running on different platforms (similar to the situation that the video game designers would like to make their products available across platforms)?…”

Scalability of Incompressible Flow Computations on Multi-GPU Clusters Using Dual-Level and Tri-Level Parallelism

“…The use of GPUs for Euler solvers and incompressible NavierStokes solvers has been well documented. [10][11][12][13][14][15][16] Thibault and Senocak 15 developed a single-node multi-GPU 3D incompressible Navier-Stokes solver with a Pthreads-CUDA implementation that targets multi-GPU desktop platforms. This work was extended in Jacobsen et al 16 where an MPI-CUDA implementation was presented and assessed on the NCSA Lincoln Tesla Cluster.…”