The current evolution in the aeronautical field towards high-fidelity simulations, including multi-physics and more reliable modelling of turbulence and transition, calls for a new approach of the complete CFD-multi-physics simulation chain, with a drastic reduction of its turnaround time.
This requires revising the whole CAE chain, from pre-processing (CAD handling and mesh generation), to very fast basic CFD algorithms and to efficient, full parallel post-processing, in order to achieve a reduction of the global turn-around time by several orders of magnitude.
On a shorter term, of 24 months of the current CfP project, the following objectives were ensured, based on very recent developments performed at NUMECA Int.:
- A gain of one order of magnitude at the pre-processing level, covering automatic CAD cleaning, wrapping and parallel unstructured grid generation for arbitrary complex configurations with the software system HEXPRESS™/Hybrid.
- A gain of one order of magnitude, due to a novel convergence acceleration algorithm, allowing calculations with CFL=1000 and convergence of steady state RANS simulations, in 50 multigrid cycles.
The present proposal had an objective to respond to the CfP topic by:
- extending these capabilities to the GRA-LNC configurations;
- extending the convergence acceleration methodology to simulations with laminar-turbulent transition, and to unsteady flows;
- providing guidelines for a next generation software environment for industrial aerodynamics simulation, in response to task 2 of the CfP;
- porting of the CFD code and the convergence acceleration algorithms to GPU’s, with an expected additional gain of 1 to 2 orders of magnitude.
One therefore expected, combining the above mentioned efforts that within the framework of the project duration, a gain of 3-to 4 orders of magnitude would be achieved, in global CPU performance and turn-around time, for steady state RANS simulations in a first step.
The needs of the aeronautical industry in term of simulation lead to the current trend towards “High-Fidelity”. In particular, CFD whereby increasing complex geometries and physical effects are being simulated with up to hundreds of millions points, resulting in costly computations. Hence, there is a growing need towards further improvement of the current CFD computation chain in order to minimize computational costs and engineering time.
Many of the current industrial CFD solvers are either based on explicit Runge-Kutta time integration schemes, associated with a low level of implicitness through a residual smoothing step, or they are based on techniques such as conjugate gradient and GMRES-type methods. Each of these approaches have some advantages and de-advantages regarding the convergence efficiency, the memory requirements and the robustness. Hence, it is believed that there could be an optimum by combining explicit methods with simplified implicit steps, coupled to an effective multi-grid method. The solution adopted by NUMECA permits to relax the CFL constraint on large scale industrial structured and unstructured hexahedral and hybrid finite-volume solvers with applications to large scale 3D test cases. The achieved performance gain is significative.
In CFD, the treatment of the CAD geometries and the subsequent grid generation often require cleaning and wrapping the CAD files in order to obtain tight and continuous surfaces of the solid geometries, and remove geometrical features that are not necessary for the CFD objectives. This process can be very long, in terms of engineering and turn-around time, before even generating the mesh which can be quite cumbersome for very complex geometries. The new approach with Hexpress™/Hybrid automates all these operations, and integrates a shared -memory parallel generation of unstructured hexahedral dominant grids. This leads to a gain in turn-around time and engineering time of one order of magnitude.
But the “High Fidelity Simulation” needs by industrials must take into consideration multi-physic aspects like, for instance, Fluid-Structure Interactions (FSI) combining Computational Fluid Dynamics (CFD) and Computational Structure Mechanics (CSM). During a FSI simulation the CFD and CSM solvers are commonly performed separately, and have to share information about data at their common interface though a third interface such as MPCCI. This process is heavy and requires an double expertise.
It obviously leads to the need for a higher level of integration close to a monolithic approach in order to raise multi-physics simulations to a higher level of efficiency, but also to unify the entire simulation chain into a multi-users environment able to solve large industrial problems at the HPC scale. This project provided a state of the art and a set of specifications and guidelines towards a new generation of software environment, responding to these objectives.
Finally, the use of graphics processing units (GPUs) is a relatively new approach in HPC. Nevertheless this many core co-processor has shown to be very capable in accelerating a wide range of application codes by a factor up to 100 over a typical modern server CPU core. Nvidia is the main player in the GPU computing field with the Nvidia CUDA technology, a combination of hardware features and software tools. Middleware technologies also appear, like compilers and frameworks which permit to instrument the code with hardware-independent directives. They offer the potential advantage of generating efficient code by taking into account the underlying hardware at compile time. All these recent advancements underline the high potential of GPU computing to significantly enhance the performance of real world CFD simulations. In this context, the structured flow solver FINE™/Turbo has been accelerated. The computation of the fluxes and the residuals, one of the most time consuming computation part, has been offloaded in order to gain an order of magnitude on the global computation time.
All put together, this project proposed to improve each step of the CFD simulation chain by at least one order of magnitude and thus, improve the global simulation process by the same ratio.