Skip to content

WP14 - Large-Scale Execution Contract

WP14 is a contract gate, not a large benchmark. It freezes the physical model, execution profile, acceptance metrics, resource rules and evidence format before WP15-WP18 runs. The authoritative machine-readable record is qualification/0_2_7/wp14_execution_contract.json.

Reference model

The G01 reference is the public structured TET4 linear-static route generated by generate-large-tet4-block:

  • 69 x 69 x 69 bricks with the historical six-TET decomposition;
  • 343,000 nodes, 1,971,054 TET4 elements and 1,029,000 true displacement degrees of freedom;
  • unit SI cube, homogeneous isotropic material (E=210 GPa, nu=0.3);
  • all translations fixed on x=0 and a real uniform nodal dead load of 1,000,000 N on x=1 in the x direction;
  • 4,900 loaded nodes and a load resultant invariant. The generated HDF5 input is hashed byte-for-byte at execution.

The four subscale cases (81, 375, 2,187 and 14,739 true DOF) use the same geometry, material, boundary conditions and loading rules. They are the declared comparison ladder for assembled versus matrix-free execution; they are not 1M evidence.

Frozen execution profile

The captured profile is Windows x86_64, AMD64 Family 25 Model 80, 12 logical CPUs, 102,963,109,888 bytes of physical memory, Python 3.13.1, NumPy 2.2.6, SciPy 1.15.2 and scipy-openblas 0.3.29. Numerical threads are fixed to one. An execution on another machine must record a separate environment and is not equivalent evidence.

The 1M reference route is explicit matrix-free CG with the nodal block-Jacobi preconditioner. The assembled subscale route is explicit SciPy CG with the Jacobi preconditioner. There is no silent backend fallback. The contract freezes rtol=1e-8, atol=0, a maximum of 10,000 iterations and the declared seed 0, one numerical thread and the predeclared residual, equilibrium and energy checks.

CG is conditional, not universal: the constrained operator must be symmetric and positive definite for this homogeneous linear-elastic scope. A rigid mode, non-positive direction, singular preconditioner block or unsupported backend is an explicit fail-closed outcome.

1M and 3M rules

1M_PASS requires at least 1,000,000 real FEM DOF, a complete iterative solve, finite outputs, residual, reaction/equilibrium and energy evidence, a matrix-free/assembled subscale comparison, two independent replays and full resource/provenance capture. Timeout, OOM or a resource-limited verdict never creates a solve claim.

The 3M ladder is deliberately separate:

  • Bronze: real 3M FEM model plus topology and resource preflight; no solve claim;
  • Silver: complete reproducible iterative solve with residual, equilibrium, energy and resource evidence; principal target;
  • Gold: PETSc/MPI distributed solve, restart and a second case or scaling point; stretch target.

Preflight estimates memory and disk before generation. The fixed safety rule is ceil(1.25 * declared peak estimate). Timeouts are 7,200 s for 1M, 14,400 s for 3M Silver and 28,800 s for 3M Gold. Partial runs retain configuration, input digest and resource/status metadata.

Evidence and dependencies

Every future run records source/configuration/input digests, hardware and software versions, backend, solver, preconditioner, topology, true DOF, nonzero/operator memory, wall time, peak RSS, iterations, residual, equilibrium, energy, UTC timestamp and a controlled verdict. Text configuration uses canonical UTF-8/LF SHA-256; HDF5 and other binary artifacts use byte-for-byte SHA-256.

WP14 is PASS because these contracts are frozen and validated. It does not claim a 1M or 3M solve. WP15 adds matrix-free numerical evidence, WP16 is the release-blocking 1M qualification, WP17 covers PETSc/MPI, and WP18 executes the Bronze/Silver/Gold ladder.