SlabLU: Two-Level Sparse Direct Solver for Elliptic PDEs

Overview

SlabLU is a sparse direct solver designed for solving linear systems that arise from the discretization of elliptic partial differential equations (PDEs). By decomposing the domain into thin "slabs" and employing a two-level approach, SlabLU enhances parallelization, making it well-suited for modern multi-core architectures and GPUs.

The solver leverages the $\mathcal{H}^2$-matrix structures that emerge during factorization and incorporates randomized algorithms to efficiently handle these structures. Unlike traditional multi-level nested dissection schemes that use hierarchical matrix techniques across varying front sizes, SlabLU focuses on fronts of approximately equal size, streamlining its application and tuning for heterogeneous computing environments.

Key Features

• Two-Level Parallel Scheme: Optimized for parallel execution, particularly on systems with GPU support.
• Efficient Handling of $\mathcal{H}^2$-Matrices: Utilizes efficient matrix algebra for efficient computational performance.
• Flexible Discretization Compatibility: Supports a wide range of local discretizations, including high-order multi-domain spectral collocation methods.
• High Performance: Demonstrated ability to solve significant problems, such as a Helmholtz problem on a domain of size $1000 \lambda \times 1000 \lambda$ with $N=100M$, within 15 minutes to six correct digits.

Figure 1: Solutions of variable-coefficient Helmholtz problem on square domain $\Omega$ with Dirichlet data given by $u \equiv 1$ on $\partial \Omega$ for various wavenumbers $\kappa.$ The scattering field is $b_{\rm crystal}$, which is a photonic crystal with an extruded corner waveguide. The crystal is represented as a series of narrow Gaussian bumps with separation $s=0.04$ and is designed to filter wave frequencies that are roughly $1/s$.

Figure 2: Solutions of constant-coefficient Helmholtz problem on curved domain $\Phi$ with Dirichlet data given by $u \equiv 1$ on $\partial \Phi$ for various wavenumbers $\kappa$.

Usage

To install SlabLU for CPU-based development or testing, run

python -m pip install --upgrade pip setuptools wheel
python -m pip install --no-build-isolation -e .

To test the framework on a benchmark problem, run

python argparse_driver.py --n 1000 --disc hps --p 22 --ppw 10 --domain square --pde bfield_constant --bc free_space --solver slabLU

which solves the constant-coefficient Helmholtz equation of size $100 \lambda \times 100 \lambda$ on a unit square with $N \approx 1\rm M$ with local polynomial order $p=22$.

To solve a set of example PDEs on curved domains, run

python generate_pictures.py

which generates the figures in [1].

Citation

If you use SlabLU, please cite:

Yesypenko, Anna, and Per-Gunnar Martinsson. "SlabLU: a two-level sparse direct solver for elliptic PDEs." Advances in Computational Mathematics 50, no. 4 (2024): 90. https://doi.org/10.1007/s10444-024-10176-x

Related Papers

[1] Yesypenko, Anna, and Per-Gunnar Martinsson. "GPU optimizations for the hierarchical Poincaré-Steklov scheme." In International Conference on Domain Decomposition Methods, 519-528. Springer, 2022.