As heterogeneous supercomputing architectures leveraging GPUs become increasingly central to high-performance computing (HPC), it is crucial for computational fluid dynamics (CFD) simulations, a de-facto HPC workload, to efficiently utilize such hardware. One of the key challenges of HPC codes is performance portability, i.e. the ability to maintain near-optimal performance across different accelerators. In the context of the REFMAP project, which targets scalable, GPU-enabled multi-fidelity CFD for urban airflow prediction, this paper analyzes the performance portability of SOD2D, a state-of-the-art Spectral Elements simulation framework across AMD and NVIDIA GPU architectures. We first discuss the physical and numerical models underlying SOD2D, highlighting its computational hotspots. Then, we examine its performance and scalability in a multi-level manner, i.e. defining and characterizing an extensive full-stack design space spanning across application, software and hardware infrastructure related parameters. Single-GPU performance characterization across server-grade NVIDIA and AMD GPU architectures and vendor-specific compiler stacks, show the potential as well as the diverse effect of memory access optimizations, i.e. 0.69× - 3.91× deviations in acceleration speedup. Performance variability of SOD2D at scale is further examined on the LUMI multi-GPU cluster, where profiling reveals similar throughput variations, highlighting the limits of performance projections and the need for multi-level, informed tuning.

As heterogeneous supercomputing architectures leveraging GPUs become increasingly central to high-performance computing (HPC), it is crucial for computational fluid dynamics (CFD) simulations to maintain performance portability. In this paper, we examine the performance and scalability of CFD framework SOD2D in a crosslayer manner, i.e. across application, software and hardware infrastructure related parameters. Single-GPU performance characterization across server-grade NVIDIA and AMD GPU architectures and vendor-specific compiler stacks, show the potential as well as the diverse effect of memory access optimizations, i.e. 0.69x - 3.96x deviations in acceleration speedup. Performance variability of SOD2D at scale is then further examined on the LUMI multi-GPU cluster, showcasing analogous diverse effects on throughput, demonstrating the ineffectiveness of adopting performance projections, thus underscoring the importance and necessity of cross-layer informed performance analysis and tuning for multi-GPU configurations.

This paper focuses on using the Newton-Raphson method to solve the power-flow problems. Since the most computationally demanding part of the Newton-Raphson method is to solve the linear equations at each iteration, this study investigates different approaches to solve the linear equations on both central processing unit (CPU) and graphical processing unit (GPU). Six different approaches have been developed and evaluated in this paper: two approaches of these run entirely on CPU while other two of these run entirely on GPU, and the remaining two are hybrid approaches that run on both CPU and GPU. All six direct linear solvers use either LU or QR factorization to solve the linear equations. Two different hardware platforms have been used to conduct the experiments. The performance results show that the CPU version with LU factorization gives better performance compared to the GPU version using standard library called cuSOLVER even for the larger power-flow problems. Moreover, it has been proven that the best performance is achieved using a hybrid method where the Jacobian matrix is assembled on GPU, the preprocessing with a sparse high performance linear solver called KLU is performed on the CPU in the first iteration, and the linear equation is factorized on the GPU and solved on the CPU. Maximum speed up in this study is obtained on the largest case with 25000 buses. The hybrid version shows a speedup factor of 9.6 with a NVIDIA P100 GPU while 13.1 with a NVIDIA V100 GPU in comparison with baseline CPU version on an Intel Xeon Gold 6132 CPU.

In this paper, an automation process to perform Reynolds-Averaged Navier-Stokes (RANS) computational fluid dynamics (CFD) analysis is developed to carry out aerodynamic design and optimization. The aircraft model/geometry is defined by a Common Parametric Aircraft Configuration Schema (CPACS) file, and the analyses are facilitated using high performance computers (HPC). As the computational capability of the available HPC systems is a limiting factor in the complexity of analyses that can be performed, a detailed performance analysis of the open source CFD code SU2 is undertaken and the profiling and performance analyses for large simulations are carried out.

GPU-accelerated computing is becoming a popular technology due to the emergence of techniques such as OpenACC, which makes it easy to port codes in their original form to GPU systems using compiler directives, and thereby speeding up computation times relatively simply. In this study we have developed an OpenACC implementation of the high order finite difference CFD solver ESSENSE for simulating compressible flows. The solver is based on summation-by-part form difference operators, and the boundary and interface conditions are weakly implemented using simultaneous approximation terms. This case study focuses on porting code to GPUs for the most time-consuming parts namely sparse matrix vector multiplications and the evaluations of fluxes. The resulting OpenACC implementation is used to simulate the Taylor-Green vortex which produces a maximum speed-up of 61.3 on a single V100 GPU by compared to serial CPU version.

Due to its high performance and throughput capabilities, GPU-accelerated computing is becoming a popular technology in scientific computing, in particular using programming models such as CUDA and OpenACC. The main advantage with OpenACC is that it enables to simply port codes in their "original" form to GPU systems through compiler directives, thus allowing an incremental approach. An OpenACC implementation is applied to the CFD code Nek5000 for simulation of incompressible flows, based on the spectral-element method. The work follows up previous implementations and focuses now on the P-N-PN-2 method for the spatial discretization of the Navier-Stokes equations. Performance results of the ported code show a speed-up of up to 3.1 on multi-GPU for a polynomial order N > 11.

Airinnova is a start-up company with a key competency in the automation of high-fidelity computational fluid dynamics (CFD) analysis. Following on from our previous PRACE SHAPE project, we have continued collaborating with the PDC Center for High Performance Computing at the KTH Royal Institute of Technology (KTH-PDC), to investigate the performance analysis of the open source CFD code SU2 and further develop the automation process for the field of aerodynamic optimization and design.

Stable and conservative interface boundary conditions are developed for the unsteady compressible Navier-Stokes equations using finite difference and finite volume methods. The finite difference approach is based on summation-by-part operators and can be made higher order accurate with boundary conditions imposed weakly. The finite volume approach is an edge- and dual grid-based approach for unstructured grids, formally second order accurate in space, with weak boundary conditions as well. Stable and conservative weak boundary conditions are derived for interfaces between finite difference methods, for finite volume methods and for the coupling between the two approaches. The three types of interface boundary conditions are demonstrated for two test cases. Firstly, inviscid vortex propagation with a known analytical solution is considered. The results show expected error decays as the grid is refined for various couplings and spatial accuracy of the finite difference scheme. The second test case involves viscous laminar flow over a cylinder with vortex shedding. Calculations with various coupling and spatial accuracies of the finite difference solver show that the couplings work as expected and that the higher order finite difference schemes provide enhanced vortex propagation.