Monolithic Multi-level Overlapping Schwarz Solvers for Fluid Problems

\firstname Stephan \lastname Köhler, \firstname Oliver \lastname Rheinbach · 2025

Read Monolithic Multi-level Overlapping Schwarz Solvers for Fluid Problems by \firstname Stephan \lastname Köhler, \firstname Oliver \lastname Rheinbach (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-04356.txt.

Introduction

Stephan Köhler and Oliver Rheinbach report large tests of fluid solvers in a 2025 paper. Their tools are Schwarz methods, which split a hard flow problem into patches and solve them side by side. Added coarse levels help the method scale to huge machines. A coarse space called GDSW shapes the design, and the code runs through two linked software libraries.

The team ran the solvers on up to 32,768 processor ranks. Test cases were a clean pipe flow in a unit cube and a complex die shape used in forming work. Both a two-level and a three-level form of the solver held up at that scale. The runs show that patch solvers with coarse help can serve the next class of large computers. The work sits in a German project on new machines for flow studies.

Catalog record

CREATOR\firstname Stephan \lastname Köhler, \firstname Oliver \lastname Rheinbach
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE24 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-04356

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Monolithic Multi-level Overlapping Schwarz Solvers for Fluid Problems \firstname Stephan \lastname Köhler, \firstname Oliver \lastname Rheinbach Abstract Additive overlapping Schwarz Methods are iterative methods of the domain decomposition type for the solution of partial differential equations. Numerical and parallel scalability of these methods can be achieved by adding coarse levels. A successful coarse space, inspired by iterative substructuring, is the generalized Dryja–Smith–Widlund (GDSW) space. In [ 8 ] , based on the GDSW approach, two-level monolithic overlapping Schwarz preconditioners for saddle point problems were introduced. We present parallel results up to 32 768 MPI for the solution of incompressible fluid problems for a Poiseuille flow example on the unit cube and a complex extrusion die geometry ranks using a two- and a three-level monolithic overlapping Schwarz preconditioner. These results are achieved through the combination of the additive overlapping Schwarz solvers implemented in the Fast and Robust Overlapping Schwarz (FROSch) library [ 4 ] , which is part of the Trilinos package ShyLU [ 6 ] , and the FEATFLOW library [ 13 ] using a scalable interface for the efficient coupling of the two libraries. This work is part of the project StroemungsRaum - Novel Exascale-Architectures with Heterogeneous Hardware Components for Computational Fluid Dynamics Simulations, funded by the German Bundesministerium für Forschung, Technologie und Raumfahrt BMFTR (formerly BMBF) as part of the program on New Methods and Technologies for Exascale Computing (SCALEXA). \TitleLanguage [EN] \AbstractLanguage [EN] 1 Introduction We consider monolithic overlapping Schwarz preconditioners of the generalized Dryja-Smith-Widlund (GDSW) type [ 2 , 3 ] for the application to fluid problems. This work is part of the StroemungsRaum project, financed by the BMFTR (formerly BMBF) under the SCALEXA program. SCALEXA is the German initiative to develop software in high-performance computing (HPC) in the exascale computing era. StroemungsRaum aims to extend the CFD software FEATFLOW [ 13 ] with highly scalable domain decomposition methods and multigrid solvers tailored for heterogeneous CPU-GPU systems. By leveraging parallel near-hardware implementations, the project targets efficient CFD simulations on future exascale architectures. This paper is concerned with the enhancement of the FEATFLOW software library by scalable domain decomposition methods based on the FROSch [ 4 ] framework with a focus on GDSW with three and more levels [ 5 ] . FROSch is part of the ShyLU package [ 6 ] within the Trilinos software library [ 7 ] , and provides a flexible infrastructure for implementing advanced overlapping domain decomposition preconditioners. In the numerical simulation of fluid dynamics problems, the underlying discretized systems typically exhibit a saddle point structure. As the geometric complexity increases and finer mesh resolutions are required to resolve detailed flow features or boundary layers, the resulting linear systems can become extremely large. For such systems, direct solvers become infeasible due to their high computational cost and memory consumption. This motivates the use of iterative methods, particularly Krylov subspace methods, combined with multilevel preconditioners from multigrid and domain decomposition. Domain decomposition methods, such as GDSW in the FROSch framework, offer a promising path to scalable preconditioning by enabling the solution process to be distributed across multiple compute nodes, exploiting both spatial locality and parallelism. In this context, multi-level strategies are essential to maintain scalability on emerging architectures with large core counts. FEAT3 is a modern C++ finite element software developed at TU Dortmund as a modular successor to earlier versions of the FEAT family. It emphasizes highly scalable geometric multigrid solvers, enabling efficient simulation of large-scale PDE systems on modern HPC architectures. Within the StroemungsRaum project, FEAT3 is used in collaboration with the industrial partner IANUS Simulation, who applies it to deliver simulation-as-a-service solutions. In [ 8 ] the monolithic two-level overlapping Schwarz preconditioner with a generalized Dryja–Smith–Widlund (GDSW) coarse spaces was introduced and its scalability and robustness was demonstrated, see also [ 9 ] . It is well known that we an increasing number of subdomains the size of the GDSW coarse problem grows and hence the factorization of the coarse problem will eventually be infeasible. To overcome this problem a multi-level version of overlapping Schwarz preconditioners was introduced in [ 10 ] , see also [ 11 , 12 ] . In this work, we extend the multi-level approach to the monolithic overlapping Schwarz preconditioner and show weak scaling resulting up to 32 768 MPI ranks for a Poiseuille flow problem on the unit square and up to 4 096 MPI ranks for Poiseuille flow on the geometry of an extrusion die. The paper is arranged as follows. In section 2 we introduce our model Stokes problem and give a brief overview of the monolithic overlapping Schwarz preconditioner and the extension to the monolithic multi-level preconditioner. In section 3 , we discuss briefly our implementation and in section 4 , we show our numerical results. 2 Stokes Problem For our monolithic multi-level overlapping Schwarz preconditioner, we consider the Stokes equation over a computational domain \Omega\subset\mathbf{R}^{d} , where d=2,3 . The weak formulation of the Stokes equation is given by: Find u\in V_{g} and p\in Q such that \displaystyle\mu\int_{\Omega}\nabla u\colon\nabla v\,\mathrm{d}x \displaystyle-\int_{\Omega}\mathrm{div}v\,p\,\mathrm{d}x \displaystyle= \displaystyle\int_{\Omega}fv\mathrm{d}x \displaystyle\forall v\in V_{0}, \displaystyle-\int_{\Omega}\mathrm{div}u\,q\,\mathrm{d}x \displaystyle= \displaystyle 0 \displaystyle\forall q\in Q_{0}, \displaystyle u=g, \displaystyle x\in\partial\Omega_{D}, \displaystyle\frac{\partial u}{\partial n}-pn=0, \displaystyle x\in\partial\Omega_{N}, where \Omega_{D} denotes the Dirichlet boundary, \Omega_{N} denotes the Neumann boundary, and n the outward pointing normal vector. Note that in our numerical experiments, we only consider problems with an inflow and outflow boundary conditions and therefore the pressure is uniquely determined. We could handle a non-unique pressure in the same way as considered in [ 8 ] . For the discretization, we use the FEATFLOW library [ 13 ] and Q2 elements for the velocity part u and P1 discontinuous for pressure part p . The discretized system can be written as \mathcal{A}:=\begin{bmatrix}A&B^{T}\\[2.15277pt] B&0\end{bmatrix}\begin{bmatrix}u\\[2.15277pt] p\end{bmatrix}=\begin{bmatrix}F\\[2.15277pt] 0\end{bmatrix}=:\mathcal{F}, (2.1) where the matrix A corresponds to the velocity and B^{T} is the coupling between the pressure and velocity. 2.1 Monolithic Overlapping Schwarz Preconditioner For overlapping Schwarz methods, the computational domain \Omega is decomposed into N nonoverlapping subdomains \Omega_{i} , i=1,\ldots,N and we denote the corresponding subdomain with k layers of overlap with \Omega_{i}^{\prime} . We denote the local saddle point problem of \Omega_{i}^{\prime} with \mathcal{A}_{i} , i.e., \mathcal{A}_{i} is the restriction of \mathcal{A} onto the overlapping subdomain \Omega_{i}^{\prime} . The GDSW coarse spaces for ( 2.1 ), see [ 8 ] , can be written as \mathcal{B}_{\mathrm{G}DSW}=\Phi\mathcal{A}_{0}^{-1}\Phi^{T}+\sum_{i=1}^{N}\mathcal{R}_{i}^{T}\mathcal{A}_{i}^{-1}\mathcal{R}_{i}, (2.2) where \displaystyle\mathcal{A}_{0}:=\Phi^{T}\mathcal{A}\Phi, \displaystyle\mathcal{R}_{i}:=\begin{bmatrix}R_{i,u}&0\\[2.15277pt] 0&R_{i,p}\end{bmatrix}, (2.3) the columns of \Phi are the coarse basis functions, R_{i,u} denotes the restriction of the velocity onto the i -th subdomain, and R_{i,p} the restriction of the pressure [ 8 , 9 ] . Note that A_{0} is invertible since we have a unique pressure due to the outflow boundary condition and the local problems \mathcal{A}_{i} are invertible since they correspond to problems with Dirichlet boundary conditions for the velocity and pressure. The coarse basis functions are the energy minimizing extensions of the restriction of the null space of the operators onto the interface \Gamma and can be constructed the following: We define the interface \Gamma as \Gamma=\bigcup_{i=i}^{N}\partial\Omega_{i}\setminus\partial\Omega_{D} . Furthermore, we divide the interface into vertices and edges and, in 3D, faces. By reordering of the degrees of freedom (dofs), we can write \mathcal{A} as \mathcal{A}=\begin{bmatrix}\mathcal{A}_{II}&\mathcal{A}_{I\Gamma}\\[2.15277pt] \mathcal{A}_{\Gamma

Read the complete text