WP17 - Final PETSc/MPI large solver path closure¶
Decision¶
WP17 is PASS_WITH_LIMITATIONS. This closeout consumes the controlled
WP17-R, WP16 retry and WP18 Silver evidence; it does not claim a new heavy
benchmark. The earlier WP17 and WP17-R PARTIAL records remain preserved as
historical checkpoints and are not rewritten.
The closed scope is the explicit PETSc 3.25.1 / MPICH 5.0.1 route using CG, GAMG, AIJ storage, two MPI ranks and the pinned container image. It is a bounded TET4 linear-static large-model path, not a universal PETSc, HPC or performance claim.
Frozen contract and runtime¶
The WP14 contract remains authoritative. Its acceptance tolerance is 1e-8;
the supplemental PETSc route uses the predeclared internal solver target
1e-10 and an explicit unpreconditioned KSP stopping norm. No acceptance
tolerance was changed and no post-result tuning was used.
| Item | Recorded value |
|---|---|
| Container | qf-solver-large@sha256:d6a1718001fc36772906d1a9505637bbd0a4b7e1d8ccc9afdbcb6f67b7ff6d0e |
| PETSc / petsc4py | 3.25.1 |
| MPI | MPICH 5.0.1 / MPI 5.0 |
| Backend / solver / PC | PETSc / CG / GAMG |
| Matrix format | AIJ |
| MPI size / threads | 2 / 1 |
| Fallback | explicit; no silent fallback |
| Host PETSc | UNAVAILABLE |
The normal SciPy and matrix-free paths remain available. PETSc is not a mandatory runtime dependency; a requested unavailable backend must fail closed rather than silently changing route.
Closure evidence¶
| Evidence | Result | Key metrics |
|---|---|---|
| Same-configuration subscale comparison | PASS |
displacement 8.341e-15, equilibrium 9.279e-10, energy 1.598e-12 |
| WP16 official 1M retry | PASS |
1,029,000 true DOF, two replays, residual 9.704e-11, equilibrium 5.339e-10, energy 8.484e-14 |
| Supplemental strict PETSc 1M route | PASS |
431 iterations, about 186 s total per replay |
| WP18 Silver | PASS |
3,000,000 true DOF, 619 iterations, two replays, about 2,660 s total, approximately 10.08 GB peak RSS |
The official WP16 and Silver evidence includes complete model, load, BC, solve, reaction/equilibrium, energy, finite-output and replay checks. Their source SHA, input digest, backend, preconditioner, environment and runtime records remain in the referenced JSON artifacts.
Performance boundary¶
On the recorded 1M case, the old matrix-free baseline used 1,052 iterations
and 1371.06 s; the strict PETSc CG/GAMG route used 431 iterations and about
185 s, an observed speedup of about 7.4x. This is a measurement for the
declared case and environment only. It is not a promise of universal speedup,
hardware independence or general 1M/3M scaling.
The AIJ route uses approximately 3.52 GiB at 1M DOF and 10.08 GB at 3M
DOF. WP18 Gold restart/checkpoint and distinct-second-case evidence remain
unattempted and unclaimed. GPU, general HPC and alternate preconditioner
qualification remain outside this closeout.
Owner boundary¶
The closeout does not change any FEM formulation, WP14 criterion, public
default backend or capability maturity. It closes the PETSc/MPI path as a
reproducible, bounded engineering route with explicit resource limitations.
The machine-readable closeout is
qualification/0_2_7/wp17_final_state.json; historical records are listed in
that file for provenance.
FULL_REGRESSION = SKIPPED_BY_POLICY; no functional source changed in this
documentary closeout.