Achieving h‐ and p‐Robust Monolithic Multigrid Solvers for the Stokes Equations

Amin Rafiei, Scott MacLachlan

TL;DR

This paper introduces and analyzes novel monolithic multigrid solvers for the Stokes equations, specifically targeting higher-order finite-element discretizations. The key contribution is the development and evaluation of extended Vanka relaxation schemes that achieve robust performance across various mesh types and polynomial orders, offering a significant step towards efficient and scalable simulations of incompressible fluid flow.

What Problem Does This Paper Address?

Simulating incompressible fluid dynamics, particularly using higher-order finite-element methods for the Stokes equations, presents a significant computational challenge. While the discretization schemes themselves have seen considerable advancement, efficiently solving the resulting large, sparse linear systems remains a bottleneck. Traditional methods often struggle with robustness and scalability as the polynomial order of the finite elements increases (p-refinement) or the mesh is refined (h-refinement). This paper addresses the need for robust and efficient preconditioners for these systems, focusing on monolithic multigrid methods with Vanka-style relaxation, which have historically been limited to lower-order discretizations.

Methods

The core of this work lies in extending Vanka relaxation schemes to higher-order finite-element discretizations for the Stokes equations. The authors propose two main approaches:

  • Composite Vanka Relaxation for Taylor-Hood Discretizations: This involves a topological extension of existing "pressure node oriented" relaxation techniques. They construct "composite" patches that combine vertex, edge, and cell-based pressure degrees of freedom, along with associated velocity degrees of freedom. This approach is implemented using the PCPATCH and ASMPatchPC frameworks in Firedrake.
  • Extended Vanka Relaxation for H(div)-L² Discretizations: For non-conforming H(div)-L² discretizations, the authors adapt and extend Vanka relaxation schemes, particularly focusing on "extended" cell-based patches. This is crucial because these discretizations have different properties and require specialized patch constructions.

The monolithic multigrid solvers are built using these relaxation schemes as smoothers within a V-cycle. They are employed as preconditioners for the FGMRES iterative solver. The paper also discusses the implementation details, including grid transfer operations and coarse-grid operators, and compares two implementation strategies: a callback-based approach (PCPATCH) and a matrix-full approach (ASMPatchPC).

Datasets

Not discussed in the paper. The authors use a manufactured solution to the Stokes problem on a unit square domain ([0,1]²) for all numerical experiments.

Key Insights and Ideas

  • Vanka Relaxation Extensibility: The paper demonstrates that Vanka relaxation, a technique previously limited to low-order methods, can be successfully extended to higher-order Taylor-Hood and H(div)-L² discretizations, leading to robust multigrid solvers.
  • Composite Patches are Key: The development of "composite" Vanka patches for Taylor-Hood elements is crucial for achieving p-robustness, especially on quadrilateral meshes.
  • Matrix-Full vs. Callback Implementations: While callback-based implementations (PCPATCH) offer flexibility, the matrix-full approach (ASMPatchPC) can provide significant performance advantages, particularly for complex patch constructions and higher-order discretizations, despite the overhead of assembling the global matrix.
  • Stopping Criteria Challenge: A significant finding is the difficulty in establishing robust and universally applicable stopping criteria for higher-order discretizations. Standard residual-based tolerances become unreliable, and achieving discretization-level accuracy requires very strict criteria that can be sensitive to polynomial order and mesh refinement.
  • Quadrilateral Meshes Advantage: For Taylor-Hood discretizations, quadrilateral meshes exhibit better p-robustness in terms of iteration counts compared to triangular meshes, although the computational cost per iteration might be higher due to larger patches.

Key Results

  • Robust Iteration Counts: For certain discretizations (e.g., (BDMk, dPk-1) on triangles and Taylor-Hood on quadrilaterals), the number of FGMRES iterations remains remarkably stable across increasing polynomial orders (k) and mesh refinements (l), indicating good h- and p-robustness.
  • Performance Comparison: The ASMPatchPC implementation generally outperforms PCPATCH in terms of CPU time, especially for higher-order discretizations and finer grids, often by around 20%.
  • Scaling Behavior: Time-to-solution for Taylor-Hood discretizations scales roughly as O(k³) with polynomial order, which is significantly better than pessimistic estimates.
  • Stopping Criterion Dependence: The required relative tolerance for achieving discretization-level accuracy varies significantly with polynomial order and mesh refinement, highlighting a critical area for future research. For instance, for (BDMk, dPk-1), the tolerance scales from h³ for k=2,3 to h⁵ for k=5,6.
  • Accuracy vs. Cost: Higher-order discretizations can achieve comparable accuracy to lower-order methods on much coarser meshes, leading to substantial reductions in time-to-solution. For example, k=4 on a 10x10 mesh can match k=2 on a 160x160 mesh.

Limitations and Future Work

  • Stopping Criteria: The paper explicitly identifies the lack of robust stopping criteria for higher-order discretizations as a major limitation, which hinders their practical application. Further research is needed to develop reliable criteria.
  • Automated Parameter Tuning: The authors suggest exploring automated parameter choices for relaxation, potentially using local Fourier analysis or other techniques, to further optimize solver performance.
  • Patch Construction and Efficiency: Investigating alternative patch constructions and improving the efficiency of solving patch systems are identified as avenues for future work.
  • Generalizations: The authors propose extending this work to algebraic multigrid settings and to more complex systems of PDEs, such as Navier-Stokes or magnetohydrodynamics.
  • Raviart-Thomas Discretizations: The (RTk, dPk-1) discretization showed poor scalability, and the paper notes that while the extended Vanka patches are used, the performance is not as favorable as with BDMk elements. Further investigation into these discretizations might be warranted.

Summary updated

Abstract

ABSTRACT The numerical analysis of higher order mixed finite‐element discretizations for saddle‐point problems, such as the Stokes equations, has been well‐studied in recent years. While the theory and practice of such discretizations is now well‐understood, the same cannot be said for efficient preconditioners for solving the resulting linear (or linearized) systems of equations. In this work, we propose and study variants of the well‐known Vanka relaxation scheme that lead to effective geometric multigrid preconditioners for both the conforming Taylor‐Hood discretizations and nonconforming ‐ discretizations of the Stokes equations. Numerical results demonstrate robust performance with respect to FGMRES iteration counts for increasing polynomial order for some of the considered discretizations and expose open questions about stopping tolerances for effectively preconditioned iterations at high polynomial order.
BibTeX
@inproceedings{rafiei2024achieving,
  author = {Amin Rafiei and Scott MacLachlan},
  title = {{Achieving h‐ and p‐Robust Monolithic Multigrid Solvers for the Stokes Equations}},
  booktitle = {Numerical},
  year = {2024},
}
Figure