scimate-ai/hpc-skills/hpc-petsc

HPC PETSc

Treat PETSc as a layered solver toolkit. Start from the highest-level object that matches the mathematical problem and only drop lower when there is a clear reason.

Start

1. Read references/solver-stack-and-object-model.md before creating or repairing a PETSc-based solve path.
2. Read references/ksp-and-pc-matrix.md when selecting Krylov methods, direct solves, or preconditioner families.
3. Read references/snes-and-ts-patterns.md when the application is nonlinear or time-dependent.
4. Read references/options-and-ksp-snes-playbook.md when translating runtime options into code or debugging prefix handling.
5. Read references/runtime-option-matrix.md when selecting monitors, residual diagnostics, or common options-database switches.
6. Read references/dm-and-discretization-playbook.md when DM, nullspaces, multigrid layout, or field decomposition matter.
7. Read references/external-backend-matrix.md when deciding whether to route through HYPRE, direct solvers, or matrix-free paths.
8. Read references/build-and-integration.md when configuring PETSc, enabling external packages, or integrating PETSc into another codebase.
9. Read references/parallel-and-runtime-debugging.md when a distributed run shows assembly, ownership, convergence, or monitoring problems.
10. Read references/error-recovery.md when configure, setup, or solve phases fail.

Work sequence

1. Classify the problem first:

• linear system -> KSP
• nonlinear residual -> SNES
• time-dependent problem -> TS

1. Choose the data model before tuning the solver:

• Vec for distributed unknowns
• Mat or matrix-free operator for the linearization
• DM when mesh, hierarchy, or field layout must drive assembly and coarsening

1. Get a robust baseline solve working before tuning for scale.
2. Move configuration into the options database whenever practical so runs stay inspectable and reproducible.
3. Read convergence reason and monitor output before changing algorithms.

Guardrails

• Do not hand-roll nonlinear or time-stepping logic that SNES or TS already manages unless the application truly needs it.
• Do not tune preconditioners before checking operator correctness, boundary conditions, and assembly completion.
• Do not mix compiler, MPI, and external solver stacks casually across PETSc and the host application.
• Do not treat options prefixes or convergence reasons as optional diagnostics.

Additional References

Load these on demand:

• references/ksp-and-pc-matrix.md for operator-class to solver-family mapping
• references/snes-and-ts-patterns.md for nonlinear and transient workflow design
• references/runtime-option-matrix.md for monitor, logging, and option-database baselines
• references/dm-and-discretization-playbook.md for DMDA, DMPlex, field splits, and hierarchy-aware setup
• references/external-backend-matrix.md for HYPRE, direct-solver, and matrix-free integration tradeoffs
• references/build-and-integration.md for external package wiring such as HYPRE or direct-solver backends
• references/parallel-and-runtime-debugging.md for ownership ranges, assembly ordering, monitors, and distributed diagnostics
• references/error-pattern-dictionary.md for fast matching of common PETSc failure classes

Reusable Templates

Use assets/templates/ when a concrete starting point is faster than rebuilding the solve path from scratch, especially:

• ksp_poisson_minimal.c
• petsc4py_ksp_minimal.py
• petsc_build_example.sh
• petsc_ksp_slurm.sh

Outputs

Summarize:

• problem class and chosen PETSc layer
• matrix or operator strategy
• solver and preconditioner baseline
• runtime options or code-path decisions
• the exact failure phase if the workflow is being repaired