Reviewing the comparison job and saved Newton trace
We continued the action plan by inspecting the cancelled comparison job, reviewing its runner and analysing the already saved Newton trace. The scheduler still records a root cancellation without a cause. Current node status cannot explain that earlier event. The owner or scheduler support must resolve it before a fresh submission.
The runner previously returned success for coarse-tolerance solves. Those results could enter the profile comparison despite falling short of fine convergence. Execution and comparison now reject them. New output records the host and actual CPU affinity. The batch script selects allowed CPU IDs and requests a separately timed warm-up step. Execution then starts again from the initial state. Final accepted-state guard diagnostics run after timing; they do not count fallback calls within solver iterations.
Focused checks exercise rejection and overlapping trace intervals: 11 performance-tool checks pass in full and Tier-A-only modesC-147. A bounded local smoke run also checks the host runner’s warm-up and final guard output. It supplies execution evidence, not a reference comparison or speed-up measurement.
The saved Newton trace has 72059 complete HLO events with about 5.369 s of covered trace timeC-148. Taking the union of event intervals avoids counting overlap twice, but every recorded operation still belongs to the same fused module and runtime source. That does not identify ensemble inference, geometry or Jacobian cost. The performance ceiling remains open.
A concrete bounded comparison request is ready for review after the scheduler issue is resolved. No new cluster job has been submitted. RF and pedestal reference selection remains the other prerequisite for completing M3.
Technical details → Performance review.
Technical details → Surrogates › performance review