zum Inhalt springen

StrömungsRaum - Novel Exascale-Architectures with Heterogeneous Hardware Components for Computational Fluid Dynamics Simulations

Funding

The project StrömungsRaum - Novel Exascale-Architectures with Heterogeneous Hardware Components for Computational Fluid Dynamics Simulations is funded by the Bundesministerium für Bildung und Forschung (BMBF).

Project Partners

Project Description

Upcoming exascale computer architectures will be characterized by a very high number of heterogeneous hardware components, which will also include special processors or accelerators. The corresponding realization of Computational Fluid Dyanamics (CFD) applications as a central core component of today's flow simulations in the industrial environment requires highly scalable solution methods. Especially for solving the high-dimensional and transient (non)linear systems of equations, modern numerical algorithms have to be able to exploit the high peak performance of modern and future hardware. In addition, the new approaches must be implemented in a user software in such a way that they can be used by "non-HPC experts" for real applications, in particular for the simulation, control, and optimization of industry-relevant processes. Additionally, they should exploit the high performance of future exascale computers efficiently.

The open source software FEATFLOW, developed mainly at TU Dortmund University, is a powerful CFD tool and a central part of the StrömunngsRaum® platform, which has been successfully used by IANUS Simulation in the industrial environment for years. Within the scope of the overall project, FEATFLOW will be extended methodically and by hardware-related parallel implementations, such that highly scalable CFD simulations with FEATFLOW will be possible on future exascale architectures.

In the subproject here in Cologne, nonlinear Schwarz domain decomposition methods are further developed and optimized for solving CFD problems. Nonlinear Schwarz methods have the potential to improve nonlinear convergence and reduce simulation time by decomposing the nonlinear problem directly. A hardware scalable implementation optimized through tailored performance engineering will be integrated into the FROSch package, a Trilinos-based open source implementation of linear Schwarz methods. The newly developed methods will be made available via interface in FEATFLOW.

Preliminary Results

Comparison of convergence behavior of nonlinear Schwarz-methods (RASPEN approaches) and classical Newton-Krylov (NK) approach. The favorable two-level approach is H1-RASPEN, here with a second level based on MsFEM-E basis functions. Taken from "Heinlein, Alexander; Lanser, Martin; Additive and hybrid nonlinear two-level Schwarz methods and energy minimizing coarse spaces for unstructured grids. SIAM J. Sci. Comput. 42 (2020), no. 4, A2461–A2488".
Result of a CFD simulation of the fluid in an aneurysm obtained using Trilinos.