Files
SkillFactory/docs/contact_cycle_sim_performance_v042_result.md

9.5 KiB
Raw Permalink Blame History

0.4.2 application-loop / live-GUI performance result

Simulation only, accepted_skill=false, zero training/search. Qualification scope is current-source analytic S1, not a new complete matrix/u256 or hardware qualification. Original0.4.0/0.4.1 deliveries and historical evidence remain sealed.

Measured result

Case Protocol cProfile Active loop wall s Physics ticks / wall s Render submissions / wall s Sim seconds / wall s
baseline-profile02 x_bias16,1cycle/4500ticks on 113.5415 39.633 1.982 0.07927
sensor-profile03 same on 92.1030 48.858 2.443 0.09772
combined-profile04 same on 75.7835 59.379 2.969 0.11876
display-S1-05 S1,6cycles/27000ticks off 764.8630 35.300 35.300 0.07060
headless-S1-06 same S1 off 294.6793 91.625 0 0.18325

Startup, reset and final viewing are excluded from these active-loop times. The like-profiled same-seed/same-protocol baseline→combined comparison is 1.4982× throughput /33.25% lower wall time, not a claim that every workload is1.5× faster. All three4500tick trace files are identical SHA256 2ce511b4c82b6555db30a99bb70dd0231b43ed06ebf7657d57f93c7b1a950e94 (authoritative digest in performance-result.json).

The genuine moving Kit-window capture at tick2519 shows 33.98 FPS. The full S1 loop submitted27000 current-state renders, including25650 added renders and the unchanged1350 interval20 renders. Submissions35.30/s are not independently measured display-present timestamps. The120s original-speed owned-X11-window video samples at30Hz; that recording rate is not the renderer FPS. GUI measurement includes recording and early concurrent CPU regression-test overhead.

Whole-device active-loop sampling: GPUutil mean39.74%, power133.43W, SMclock2887.82MHz, unchanged powerlimit360W. These include baseline desktop clients; not process-exclusive metrics. No artificial workload, extra environments, GPU power/clock cap or frame interpolation was used.

Important: fluid GUI is not1× real-time physics. With a render per verified new state, the GUI run advances only0.07060sim-seconds/wallsecond (54sim-seconds take764.86s). Extra rendering deliberately trades physical throughput for responsiveness. Headless is faster at0.18325×, still below1×. Single-environment Python/CUDA small-operation and mandatory synchronous checking overhead remains; many-environment training amortizes that overhead and its aggregate throughput is not comparable. No GPU100%/1× promise is made.

What changed

  • Bound cProfile/perf-counter/runtime metadata observation and device-wide read-only nvidia-smi sampling.
  • Epoch sensor proxies avoid repeated getters within one scene-update/reset epoch; update/reset delegates to the original guarded SDK instance and invalidates even on zero-dt/errors.
  • Exact device/dtype-grouped synchronous snapshot copies; all original expressions/full500Hz samples remain, with independently owned CPU fields.
  • Read-only accessor reuse is confined to ONE original safety or snapshot invocation, restoring original methods/presence in finally; never caches physics/control/latch/history/RNG updates.
  • The same six finite-difference FK evaluations are batched; epsilon, IK iteration/solve/limits unchanged.33sample strict float32-representation qualification passes on this native CUDA stack and is exactly bound between PREP/RUN. This is not an all-input proof; inexact samples explicitly retain the original Jacobian.
  • combined_display draws each new verified physical state after safety and synchronous evidence. Original configured render_interval20 remains; those existing renders are not duplicated. No physical step, time scaling, teleport, reset or target-domain change is added.
  • A specific expected Kit instance must remain running/notclosed/notpaused before/after added rendering, before subsequent physics, after controls/original interval20 renders and during/following final viewing. SDK-swallowed removal failures fail closed, not as successful paints.

Correctness evidence

  • GUI and headless original campaigns/workers succeed:6cycles27000actualticks each,5seams each, preclose identity verification, unchanged original functional gates, actual callerwait0 and strict shutdown handoff/cleanup.
  • Both70tensor complete native S1 traces are byte-identical to each other and prior0.4.1 GUI11. All three whole-file SHA256: d0934245e50f3d7e174b24c109f167236c4693a023d1da01bbadfd71e41b4dba.
  • GUI/headless effective-config differences are confined to /fields/mapping/sim/fields/mapping/visualizer_cfgs/list;41authority inputs and30base binding-source inputs match between the two runs.
  • Both independent zero-step CPU physical audits pass all6original functional windows and exact fixed-world targets. This audit is separate from original runtime continuity/integrity verification.
  • Original physical/controller/protocol files and thresholds unchanged;500Hz simulated physics/100Hz controls retained;27000rollout calls,0initialization physics,22constructor reset calls/14pose writes, no rollout reset/reanchor. Genuine retained final scene: carried-state/counters equal and hold_no_physics=true; actual WM_DELETE followed by wait0/window absent.
  • CPU full suite 1557passed/6skipped. Targeted77passes include signed-zero representation gate, epoch update/reset/error boundaries, snapshot expressions/ownership, per-tick failure/evidence ordering, swallowed visualizer removal, original interval20 failure before next physics/final completion, closed/paused/replaced visualizer, headless no-painter and GUI-close semantics.
  • Physical diagnostics unchanged: RootXYmax0.080493420m, height error0.051248193m, tilt0.055240028rad, gapmin−0.000029793009m, effort/velocity ratio0.214641616, joint excess0, guideZ0. Historical strict30mm RootXY failure is not fixed or reclassified.

Failures / evidence limits retained

  • baseline-profile01 fails PREP before physics because SimulationApp lacks get_app_version; corrected to omni.kit.app.get_app. Its approved source and original failure remain.
  • Review findings (numeric-vs-bit equality; SDK-swallowed renderer removal) were fixed before their relevant native runs; reports and negative tests retained. First presentation CPU collection failed on absent pxr; only the CPU harness import was corrected, not a runtime fallback.
  • First independent headless06 postexit observer raises GPU admission mismatch without reporting PID sets. Exact transient cause is unknown. Original campaign remains its original success. After actual callerwait0 and a read-only clean nvidia-smi observation, the unchanged strict verifier passes a second observation. No new Kit/physics, whitelist expansion, foreign termination or retroactive explanation. Both observations remain evidence.
  • New-cadence native early-abort, injected visualizer-failure, startup-close, GUI-safe-stop and600s automatic viewing-timeout are not exercised; related CPU negatives are not native acceptance. No current-source nominal20/S2/S3/u256 qualification, global determinism, whole hidden PhysX-state completeness, hardware/sim2real or generic adapter claim.
  • Shared/native cache monitoring remains sampled and incomplete, not a hard filesystem quota. SO/AV standalone packages unavailable; no unrequested USD operation/install performed. This is application optimization, not USD geometry tuning.

Artifacts / accounting

Cleanup note: one-shot helpers, intermediate reviews/observations and repeated CPU receipts formerly at this root are now archived under their original member paths; see cleanup inventory and recovery. Final receipts, comparisons, physical audits, raw runs, approved snapshots and sealed deliveries remain in place. The original accounting below describes the completed task, not a new qualification of the cleaned workspace.

Root performance_v042/: original six attempts,41-input approved source snapshots, original receipts/traces, profiles/GPUCSV, genuine GUI PNG/video, audits, review findings/resolutions, CPUlogs/XML and performance-result.json. No large raw logs are presented as substitutes for compact evidence.

Task totals before packaging: 11/12Kit,73500/120000reserved calls,67500actual calls,0optimizersteps,57owned process identities checked/none alive; source41 qualifies12cycles, prior three successful one-cycle benchmarks add3cycles. Protected4427historical files rehashed unchanged in postexit checks. Final inventory/protected/process checks and delivery sealing remain separate receipts; elapsed/storage final accounting must be read from task-complete.json. Delivery is performance_v042/deliveries/go2_fl_contact_cycle_sim-0.4.2/ with adjacent tar.gz/sha256; actual source41bytes equal the recorded staging qualification, without claiming another GPU run at the relocated delivery path.

The first relocated package CPU suite found39failures (36missing historical storage document fixture,3missing saved u256 validation trace),1517passes/7skips. After adding those immutable original fixtures,1556passed/7skipped; the additional skip was missing recorded P0replay inputs. Original P0initial/episode and P1failuretrace are now embedded, with ONLY the bundled test's P1fixture locator made relative; original assertions/hash/tolerances and all41runtime authority inputs unchanged. Final relocated package suite 1557passed/6skipped; source/frozen validation passes. All three CPU receipts and repair inventories remain; these historical test fixtures are not a live authority, training run or promotion of old P1failure.