A multi-GPU accelerated solver for the three-dimensional two-phase incompressible Navier-Stokes equations

Abstract

The use of graphics hardware for general purpose computations allows scientists to enormously speed up their numerical codes. We presently investigate the impact of this technology on our computational fluid dynamics solver for the three-dimensional two-phase incompressible Navier-Stokes equations, which is based on the level set technique and applies Chorin's projection approach. To our knowledge, this is the first time, that a two-phase solver for the Navier-Stokes equations profits from the computation power of modern graphics hardware. As part of our project, a Jacobi preconditioned conjugate gradient solver for the pressure Poisson equation and the reinitialization of the level set function of our CPU based code were ported to the graphics processing unit (GPU). They are implemented in double precision and parallelized by the Message Passing Interface (MPI). We obtain speedups of 16.2 and 8.6 for the Poisson solver and the reinitialization on one GPU in contrast to a single CPU. Our implementation scales close to perfect on multiple GPUs of a distributed memory cluster. This results in excellent speedups of 115.8 and 53.7 on eight GPUs of our cluster. Furthermore our whole multi-GPU accelerated solver achieves an impressive speedup of 69.6 on eight GPUs/CPUs.

Simulation of bevel gear cutting with GPGPUs—performance and productivity

“…As mentioned in [19], in numerous publications researchers neglected including CPU-GPU data transfer into their time measurements, or neglected tuning the CPU program version and leveraging multithreading on a multicore CPU architecture for reasonable speedups. For instance, results are just compared to a single CPU in [7], while [19] and [5] look at the real performance of conventional multicore platforms. We support the latter approach by investigating an additional OpenMP implementation that uses all available cores on the platforms.…”

“…Here, we examine a particular module of KegelSpan and focus on the comparison of a variety of parallel programming models for GPGPU and multicore CPUs on different types of hardware platforms. Current GPU implementations (such as [6,7]) concentrate on the CUDA programming paradigm for GPGPUs. But other programming models for GPGPU have emerged and disregarding their existence without specific reason is not longer justified.…”

Cpu/Gpu Computing for an Implicit Multi-Block Compressible Navier-Stokes Solver on Heterogeneous Platform

“…A test case is the solution for the steady flow around NACA0012 airfoil at a Mach number of 0.3, the Reynolds number is 6 1.86 10  , the angle of attack is 3.59  . Fig.…”

“…Jespersen et al 5 accelerated Jacobi iteration section from the CFD code OVERFLOW using a GPU and shown a speedup by a factor between 2.5 and 3 compared to a single CPU. Griebel et al 6 implemented and optimized a two-phase solver for the Navier-Stokes equations using the Runge-Kutta time integration on a multi-GPU platform and achieved an impressive speedup of 69.6 on eight GPUs/CPUs. Jacobsen et al 7 utilized the MPI-CUDA programming pattern to implement a Jacobi iterative solver for the incompressible Navier-Stokes equations on the Lincoln GPU cluster with 128 GPUs and obtained a speedup of 130 over the CPU solution using Pthreads on two quad-core Intel Xeon processors.…”

GPU-accelerated Large-Eddy Simulation of Turbulent Channel Flows

“…There is a need for GPU clusters and multi-GPU parallel CFD models to study turbulent flows that are common in engineering practice. For the near future, dual-level parallelism that interleaves CUDA with Message Passing Interface (MPI) appear to be an adequate choice to address multi-GPU parallelism 10,15,16 . Jacobsen and Senocak 23 investigated trilevel parallelism using CUDA, MPI, and OpenMP for clusters with multiple GPUs per node.…”