S1 - 5M Assembly Telemetry Setup¶
S1 adds an optional, low-overhead progress log for the supervised LU2-WP04 retry. It does not run the 5M workload and does not change the frozen FEM configuration, tolerances or numerical formulation.
Contract¶
The machine-readable contract is
wp04_telemetry_contract.json.
The implementation is telemetry.py,
and the WP04 runner enables it explicitly with:
--telemetry-log qualification/0_2_7/wp04_runtime/wp04_5m_progress.jsonl
Without this option, existing runs remain unchanged. The assembly-progress file
is append-only; rank 0 writes and flushes each JSONL checkpoint. A sibling file
per rank is also written for phase markers, for example
wp04_5m_progress.rank003.jsonl. These streams are independent and perform no
MPI collective.
Progress and phases¶
elements_processed is emitted after a completed insertion chunk and is
reported against the global element count. Ranks keep local counters and use a
small integer reduction only when their local checkpoint threshold is reached.
Checkpoints target approximately one percent of a local contiguous partition,
with a minimum interval of 100,000 elements. No per-element output or matrix
synchronization is introduced.
Each record includes rates, conservative ETA values when a rate exists, memory
when available, explicit NOT_MEASURED swap status, rank count and phase. The
phase vocabulary is GENERATING, ASSEMBLING, MAT_ASSEMBLY, RHS,
PCSETUP, PC_READY, PC_READY_GLOBAL, FAILED and COMPLETED. PC_READY
is emitted only after a global frozen-route consensus; PC_READY_GLOBAL makes
that scope explicit. The rank streams record setup, ownership-gather,
post-ready memory-gather, exception and finalization markers. One-million-element milestones
are recorded with their elapsed time and seconds per million.
Failure and scope¶
An open/write/flush error changes telemetry status to DEGRADED, emits a
warning, and leaves the main FEM calculation running. It is never silently
ignored and never changes the solver result. Interrupted runs retain all
already flushed lines.
S1 status is PASS. WP04 remains
USER_INTERRUPTED_INCONCLUSIVE; S1 does not trigger C1 or claim 5M readiness.
No 5M, 1M, 3M, full-regression or numerical campaign was run for S1.
Post-PC_READY diagnosis¶
The owner-terminated retry reached rank-zero local PC_READY but did not write
COMPLETED or a raw report. S2 identified a collective-order mismatch in the
old runner: rank 0 executed a timing allreduce during evidence construction
while non-root ranks had already entered finalization. The patch moves that
reduction to an all-rank phase, adds global readiness consensus and persists
per-rank markers for the next supervised retry. It does not establish 5M
Bronze readiness or a 5M solve claim.