Skip to content

LU2-WP04 - 5M Bronze

Decision

LU2-WP04 = USER_INTERRUPTED_INCONCLUSIVE. The deterministic 5M TET4 workload and preflight were completed, but the owner-interrupted frozen AIJ construction path did not produce a completion record before PETSc/GAMG readiness. The 5M Bronze acceptance criteria therefore remain unmet. No 5M Silver or converged-solve claim is made.

The next action is a supervised WP04 retry with progress and resource telemetry. C1 is not confirmed: the 2x reference time was exceeded, but that elapsed overrun is not a resource-limit proof after a manual interruption. Any retry must remain scoped to the frozen TET4 linear-static route and must not change WP02 tolerances or the FEM formulation.

Frozen contract

  • Freeze ID: LU2-WP02-FREEZE-bfd1975b012453a3
  • Freeze digest: bfd1975b012453a3b492cc79c968ceeba6ae6951a293e3ce65ddda548d8339a1
  • Runtime image: qf-solver-large@sha256:d6a1718001fc36772906d1a9505637bbd0a4b7e1d8ccc9afdbcb6f67b7ff6d0e
  • PETSc 3.25.1, MPICH 5.0.1, 8 MPI ranks, contiguous partition, AIJ, CG, GAMG
  • No rank, partition, matrix, solver, preconditioner or tolerance change was made

The predeclared contract is wp04_execution_contract.json.

Workload and preflight

The workload is a real structured TET4 linear-static FEM model with 5,012,640 true DOF and 9,773,946 elements. It uses a homogeneous isotropic linear-elastic material, all translations fixed at x=0, and a uniform 1,000,000 N x-direction nodal load at x=length.

Two independent generator runs produced the same input digest:

ff73bc9debd0c8e1ae7355cb6b42e62734c619efa87423e484284878449a55ec

The preflight estimated 10,135,530,000 bytes of RAM and 4,090,089,600 bytes for two model recreations on the pinned Docker route; the preflight classification was PASS. The machine-readable records are wp04_preflight.json, workload_5m_run1_build.json, and workload_5m_run2_build.json.

Controlled stop

The comparable reference was the two WP02 3M 8-rank replays. Their total times were 1568.651002509 s and 1596.150367223 s, with mean 1582.400684866 s; the 2x reference point was 3164.801369732 s. The previous record noted an elapsed time of approximately 12,500 s, while forensic Docker inspection found the same container still running after about 14,531 s. All eight Python ranks were runnable and reported approximately 99.9% CPU. No OOM, MPI error or fallback was observed.

The last confirmed phase was the per-element AIJ insertion loop. The runner creates and sets up the PETSc AIJ object before that loop, but it calls matrix.assemble(), RHS setup and ksp.setUp() only afterward. No completed raw result was written, so operator completion and GAMG readiness are NOT_PROVED. Run 2 was not started after the owner interruption. The full assembly progress counter, peak RSS and swap telemetry were not persisted by this historical run. S1 installs opt-in rank-zero progress telemetry for the supervised retry; it does not alter this historical result.

The original time-budget observation is retained in wp04_resource_guard_audit.json and the corrected forensic classification is recorded in wp04_forensic_audit.json and summarized in wp04_summary.json.

Boundary and next step

The evidence proves deterministic 5M workload construction and a passing resource preflight only. It does not prove 5M operator readiness, GAMG initialization, a converged solve, or a performance claim. CPU activity is consistent with continued work, but no assembly percentage can be recovered; therefore RESOURCE_LIMITED_PROVEN = NO and C1_TRIGGER_CONFIRMED = NO. WP05 remains locked until a supervised retry completes the Bronze contract.