Skip to content

QF Solver 0.2.5a0 benchmark plan

Common protocol

Each benchmark stores mesh/input digests, SHA, machine/OS/Python/dependency versions, warm-up policy, repetitions, CPU and wall time, peak/estimated memory, DOF, nnz, Newton iterations, linear iterations, retries and residual histories. Performance assertions use a frozen baseline and variability band from WP0.

ID Work package Benchmark Families / path Primary outputs Reference
BM-025-00 WP0 published 0.2.4 baseline replay existing release suite numerical snapshots, coverage, cost v0.2.4a0
BM-025-01 WP1 meshed elastoplastic bar/coupon TET4/TET10/HEX8/HEX20 F-u, reaction, VM, PEEQ, energy, iterations analytical + external
BM-025-02 WP1 cyclic material/structure path load/unload/reload/reversal hysteresis path within isotropic law, state analytical law + external
BM-025-03 WP1 rollback/cutback equivalence forced large increment final state, retries, digests direct small-step run
BM-025-04 WP2 rigid rotation/objectivity TET4/HEX8 stress, force, energy exact zero response
BM-025-05 WP2 large-rotation cantilever TET4/HEX8, high order SHOULD load-displacement, energy, mesh trend published + external
BM-025-06 WP3 Euler column verified solid discretization critical factor/mode/cost analytical Euler
BM-025-07 WP3 nontrivial buckling case selected solid/plate case factor/mode/mesh trend external
BM-025-08 WP4 shallow arch snap-through common nonlinear core full lambda-u branch, limit point published + external
BM-025-17 WP4 finite-kinematic adaptive arc-length path TET4/TET10/HEX8/HEX20 total_lagrangian_j2 signed load factor, radius history, residual, steps bounded internal research contract
BM-025-09 WP5 frictionless block/plane finite sliding/open-close reaction, gap, active set, penetration analytical + external
BM-025-10 WP5 curved/recontact case finite sliding trajectory, contact pressure, convergence external
BM-025-11 WP6 J2 + geometry approved coupled model F-u, VM, PEEQ, energy external
BM-025-12 WP6 geometry + contact deformable contact load-gap-u, pressure, reaction external
BM-025-13 WP6 triple coupling SHOULD J2+geometry+contact complete histories external
BM-025-14 WP7 friction block cycle COULD stick/slip/reversal slip, traction, dissipation analytical + external
BM-025-15 WP8 HEX20 cost decomposition elastic/J2/geometric Gauss, constitutive, assembly, copy, solve frozen baseline
BM-025-16 WP9 adversarial suite every mandatory solver structured failure and state safety contract
BM-025-18 WP5/WP8 bounded finite-sliding diagnostic path TET4 opt-in common Newton projection mode, face switch, gap, penetration, sparse tangent, timing/RSS internal contract; no general contact claim

The following bounded observations are now available in the internal robustness campaign (they remain provisional evidence, not closed release gates):

  • BM-025-06: TET4/HEX8 critical-factor and sparse tangent-bracket smoke;
  • BM-025-08: proportional arc-length continuation to load factor one;
  • BM-025-08: reduced shallow-arch branch following through the analytical limit point, with equilibrium error and branch-turn diagnostics. This is algorithmic internal evidence only; the FEM snap-through case remains open;
  • BM-025-17: all four finite-kinematic solid families reach signed load factor 0.5 with the common driver and a bounded adaptive radius. This is monotone internal research evidence only; no FEM snap-through, post-buckling or external correlation claim is made. The opt-in performance path is python scripts/benchmark_nonlinear_025.py --path arc_length_finite_kinematic. The latest four-family replay is archived at results/benchmark_0_2_5/arc_length_finite_kinematic_latest.json and all four rows pass with residuals at or below 9.90e-08. HEX20 required the explicit max_arc_steps=512 benchmark budget and completed in 319 steps; this is a continuation-budget observation, not a performance target.
  • BM-025-09: common penalty contact open/close activation and Newton smoke.
  • BM-025-18: the explicit path python scripts/benchmark_nonlinear_025.py --path finite_sliding now runs a one-step common Newton contact case. The recorded smoke uses 15 DOF, one Newton iteration, one clamped projection with the serialized mode bounded_closest_point_node_to_triangle, nine sparse tangent entries and a finite residual. It is a bounded internal diagnostic, not a surface-to- surface or production finite-sliding qualification.
  • BM-025-15: explicit profiling paths for load-control, geometric_nonlinear_static, arc-length, contact and coupled execution, with elapsed time, phase timings, allocations and optional RSS. A local HEX20 load-control replay also measured the redundant geometry-validation hotspot: 114.288 s to 17.321 s total and 96.639 s to 11.573 s assembly on the same 96-DOF case. This remains provisional until repeated on a clean SHA with numerical before/after comparison.
  • VV-061 / BM-025-15 now also records component-level phase telemetry for the four-family load-control replay: element setup, element kernel, element scatter, sparse conversion, contact assembly, tangent nnz, nonlinear assembly-plan cache hits/misses, Python allocation peak and RSS samples. This identifies the current HEX20 hotspot and verifies reuse of immutable assembly objects without claiming a scaling law or a release performance target.
  • The cache replay is archived at results/benchmark_0_2_5/nonlinear_load_control_cache_latest.json. It converged for all four families with zero cache misses during Newton assembly. The setup-time reduction is exploratory, single-run evidence on a dirty worktree and does not close G08.
  • A two-repeat all-family replay is archived at results/benchmark_0_2_5/nonlinear_load_control_all_families_repeats2.json. It contains eight converged runs across TET4/TET10/HEX8/HEX20 and records phase timings, residuals, Newton counts, Python allocation peaks and RSS samples. It is still dirty-worktree characterization and does not close G08.
  • BM-025-07: the first shared solid-family CalculiX buckling probe is archived at results/vnv_0_2_5/calculix_buckling_solid_families_mode1_recorded/. The deck uses a fixed boundary, one requested mode and a Lanczos subspace bounded by the free-equation count. It is still negative evidence: TET4, TET10 and HEX8 remain outside the 10 % band and C3D20 stops natively. No buckling correlation claim is made from it.

Scale bands

  • CI-small: seconds, deterministic contract tests.
  • Engineering-medium: multi-element mesh/convergence, minutes.
  • Evidence-heavy: external correlations and refined meshes, separately marked.

No million-DOF claim is planned. This release characterizes nonlinear costs at engineering-medium scale and preserves existing sparse/HPC behavior.

Acceptance discipline

Before changing a benchmarked path, WP0 records baseline repeatability and freezes target/warning/reject bands. Faster execution cannot compensate for a failed physical/numerical metric. Performance regressions may be accepted only with an Owner-reviewed reason and an updated limitation.