Digital Drosophila Archive
49 lines
2.9 KiB
Markdown
49 lines
2.9 KiB
Markdown
# Independent CPU residency performance probe
|
|
|
|
This is a bounded engineering experiment authorized after the user asked to try
|
|
reducing repeated process startup. It does not replace or modify registered AR,
|
|
its timing/seed/dynamics, any frozen model, or any existing scientific artifact.
|
|
No GPU worker or new brain simulation is launched, and AR interim learning
|
|
outputs are not inspected. Input selection uses the two named, already sealed
|
|
AQ engineering fixtures, not outcomes from the ongoing experiment.
|
|
|
|
## Fixed experiment
|
|
|
|
Serial sequence: M, N, M, N_D0, N_frozen, M_D0, M. M is the mechanical fresh
|
|
parent; N is the natural light fresh engineering trial, whose actual prior
|
|
overlay comes from an AP learned snapshot. The three explicitly derived command
|
|
controls change only D=0/frozen over the same sealed events. These are not new
|
|
conditioning trials and do not alter the underlying files or install weights.
|
|
|
|
Compare fresh original epoch postprocessing against the exact same function
|
|
called repeatedly in a dedicated worker. Preserve all original hashes and
|
|
checkpoint validation; do not cache mutable neural state or delta. Every output
|
|
is independently checked by the original fresh-process CPU validator. A second,
|
|
separate resident validator is a clearly identified prototype: it has a new
|
|
schema/source identity, caches only SHA-checked module identity, and preserves
|
|
the original numerical and input-checking body.
|
|
|
|
Require exact four-file SHA identity, raw delta dtype/shape/byte identity,
|
|
complete update-receipt/trace-digest identity, and identical scientific verifier
|
|
fields. Compare mechanical and natural cases back to their original AQ outputs.
|
|
Check D0/frozen byte identity, and source/consumed-input hashes before and after.
|
|
Each task ACK is distinct from a true worker process exit. Both resident workers
|
|
must exit zero as observed by their parent; every fresh validator must exit zero.
|
|
|
|
One active CPU task at a time, one numerical thread, low scheduling priority,
|
|
CUDA hidden and uninitialized, at most seven rows, 180-second child-response
|
|
limit and 900-second experiment limit. Fail closed, preserve partial outputs,
|
|
no automatic retry. New output root must be an empty temporary directory outside
|
|
the registered experiment/model trees. No full-history rescan is claimed.
|
|
|
|
## Interpretation limits
|
|
|
|
Report cold and warm timings separately. Fresh reference and resident treatment
|
|
use the same unchanged inputs, original function body and complete validation.
|
|
The comparison excludes GPU/brain work, AP outer wrapper and AR ledger overhead.
|
|
It cannot establish whole-run speedup, GPU correctness, sequence generality,
|
|
learning success or readiness to replace AR. Nonzero prior delta is included,
|
|
but only two source trajectories and one compartment are covered. Broadening to
|
|
different compartments or a persistent GPU runtime requires a new independent
|
|
equivalence test and authorization before replacing a registered workflow.
|