Distributed semi-speculative parallel anisotropic mesh adaptation

Area of Science:

• Computational Science and Engineering
• Numerical Analysis
• High-Performance Computing (HPC)

Background:

• Anisotropic mesh adaptation is crucial for simulating complex physical phenomena.
• Existing methods often rely on collective communication and global synchronization, limiting scalability on modern HPC architectures.
• Efficiently adapting large-scale meshes requires novel parallel strategies.

Purpose of the Study:

• To present a distributed memory method for anisotropic mesh adaptation.
• To avoid collective communication and global synchronization in parallel meshing.
• To leverage concurrency in emerging HPC architectures for mesh generation.

Main Methods:

• Separation of meshing functionality from performance aspects using distinct software components.
• Utilizing a multicore cc-NUMA (shared memory) mesh generation software.
• Employing a parallel runtime system for distributed memory execution on HPC clusters.
• Decomposing the mesh, adapting interface elements on a shared memory node, and freezing them during parallel adaptation of interior elements.

Main Results:

• The method successfully generates high-quality anisotropic meshes up to approximately 1 billion elements.
• Performance is comparable to state-of-the-art HPC meshing software.
• Re-designs of the shared memory software and utilization of its speculative execution model contribute to good performance.

Conclusions:

• The proposed distributed memory method offers a scalable and efficient approach for anisotropic mesh adaptation on HPC systems.
• The strategy of separating meshing logic and utilizing frozen interface elements effectively maintains mesh conformity and quality.
• This work provides valuable insights for developing future parallel meshing algorithms on advanced computing architectures.