# HP-only fresh wire-floor baseline

2026-10-11. The user restricted this continuation to HP. There is no SCAN
connection, job, or configuration operation in this generation. This is an
engine-free ordinary-loopback diagnostic, separate from the existing physical
llfix reference results and the frozen Stage 2 rankings.

## Frozen workload and functional proof

Source `55fe6661b9efa7f0ec25ef36410022cbc5fbe00d` is retained on
`s/codex-llfix-reference-hp-epoch-source-20261011`. The dedicated
[HP declaration](../epoch_baseline/PROFILE_HP.md) preserves the existing
SCAN controller and profile bytes. All native floor sources remain unchanged.
Seven of twelve controller function ASTs are unchanged. The five changed
functions select HP paths/placement and add host/policy and complete-pair checks;
the compiler command, payload, native timing boundaries and sample counts are
unchanged.

The 14-file capsule binds twelve runtime inputs plus the cold-control and
retention helpers. Its SHA256 is
`8535bdb2cad76bcfdd71b13f8a758467de14e849ccbb5cd489510a0a5194a5ca`.
Fresh GCC 11.5/C++17/O3/march=native preparation uses one compiler process under
the high-priority shared lease. Functional execution uses a high-priority isol
lease: server CPU 9, client CPU 14, observers/controller CPUs 0–7, as declared by
HP's installed load profile. Both actual socket pairs pass independent cold
audit: 1000 closed samples, and 1000 each of paced service/response/scheduler
samples, with 100 warmups. Four altered controls refuse changed count, wrong
client affinity, failed native return and truncated raw. The original vector
still passes. No skips or SDK inputs.

Every timed attempt uses high-priority isol-no-shared, 120 seconds of quiet
preflight, the unchanged 5-percent process and load 1 limits, continuous pidstat
and a two-second tail. Three closed and three paced pairs each plan 100,000
samples and 1,000 warmups; paced arrivals remain 50,000/s, window 64, 128-byte
touched payload, original buffers/quickack. Shedding is retained as incomplete;
there is no rate substitution or reconstructed latency.

## Timing decisions

Quiet1 refused during preflight:
VS Code PID 221861 reached 23.6 percent of one CPU. Load stayed 0.00–0.05, but
this does not waive the process gate. Zero timed pairs ran in that attempt.
Quiet2 passes the unchanged preflight, all six actual pairs and independent
cold raw/activity audit. All twelve timed raw series have exactly 100,000
samples, with zero lost/clamped records: 1,200,000 samples in total. Paced
offered/admitted/sent/completed counts all equal 100,000 per pair, with zero
shed, deficit or warmup deficit. Complete activity coverage contains 15
one-second sample times, with no unowned process above the existing gate in
any pair window. Both normalized live policy fingerprints match.

Each p50/p99/mean below is the median of three individual run statistics.
Maximum is the largest actual maximum across the three runs. Quantiles are
not pooled; three repeats do not establish confidence intervals.

| Probe | Total samples | p50 us | p99 us | Mean us | Largest maximum us |
| --- | ---: | ---: | ---: | ---: | ---: |
| Closed round trip | 300,000 | 11.441 | 12.218 | 11.454887 | 25.736 |
| Paced round trip from actual issue | 300,000 | 10.070 | 10.806 | 10.086378 | 17.665 |
| Paced response from planned arrival | 300,000 | 10.122 | 10.856 | 10.137576 | 17.721 |
| Paced scheduler delay | 300,000 | 0.051 | 0.102 | 0.051312 | 1.445 |

All paced pairs complete at 49,999.998–49,999.999 per second. Their observed
maximum in-flight count is one, against the declared upper bound of 64.
Closed summaries preserve TSC calibration; the three paced probes preserve
monotonic nanoseconds. No cross-domain duration subtraction is performed.
This qualifies the captured diagnostic sitting under its source and quiet
checks; physical, SDK, Stage 2 ranking, vendor and release admission stay false.

## Busy polling and frequency comparison

Read-only live checks confirm the exact installed profile bytes and
`net.core.busy_read=50`, `net.core.busy_poll=50`. The load profile also declares
`KERNEL_BUSY_POLL_USECS=50` for its explicit per-socket load-harness setting.
The plain floor does not call setsockopt(SO_BUSY_POLL); that load setting is not
an extra flag secretly applied to this workload. Ordinary loopback does not
measure the effect of physical Mellanox IRQ placement or Onload.

After quiet1, HP reports intel_pstate no_turbo=1, min/max performance percentages
100/100, turbo_pct=0, and a 3.10 GHz hardware/scaling ceiling. Cpupower reports
boost unsupported/inactive. Idle scaling_cur_freq=1.20 GHz is an idle snapshot,
not evidence of the frequency sustained during a native hot loop. No turbo,
governor, BIOS, sysctl, IRQ, persistence or service setting is changed.

Llfix's [published setup](https://www.llfix.net/) says maximum CPU frequency,
isolated cores, SMT disabled and interleaved memory, using two Gold 6134 CPUs,
an SFN8522 PLUS, RHEL 9.4 and GCC 11.4.1. Its
[latency article](https://akinocal1.substack.com/p/how-to-achieve-p90-sub-microsecond)
describes 1,000 warmups, one million sends and RDTSCP probes including encoding,
memory-mapped persistence and local send/enqueue. Its
[server article](https://akinocal1.substack.com/p/building-a-high-throughput-fix-server)
describes the separate multi-client loopback receive workload. These are
different from this floor and from a wire-to-wire or order/ack latency test.

The Gold 6134 has a 3.20 GHz base and up to 3.70 GHz turbo according to
[Intel](https://www.intel.com/content/www/us/en/products/sku/120493/intel-xeon-gold-6134-processor-24-75m-cache-3-20-ghz/specifications.html).
The vendor does not disclose its sustained benchmark clock. Ideal scaling
between 3.20 and 3.10 GHz explains approximately 3 percent, whereas the retained
[large-buffer HP Onload comparison](../../../../internal/docs/bench/RECORD_2026-10-08_llfix-hp-load-os.md)
has 642.5 ns p50 versus 573 ns, a 12.13 percent gap. Sustained vendor turbo could
explain more, but neither hardware maximum nor governor name establishes the
actual speed or causality. The vendor's pinned native output used a 3.2 GHz TSC
conversion; that is not an observation of sustained core frequency. Original
RDTSCP/CPU-MHz and separately calibrated clocks retain their own identities;
no frequency rescaling is applied to make the distributions agree.

The retained HP p99 point estimate 2620.5 ns is close to the published 2663 ns;
the means, rare tails, hardware and effective buffers remain different. Those
physical comparisons are diagnostic and do not become vendor reproductions.
They are not fresh wire-floor results from this record.

## Retention and verification

HP and the local replica independently stream all 80 sealed files, matching
their SHA256, size and mode. The archive is 2,511,114 bytes and remains at
`~/archive/libhft/codex-llfix-reference-20261011/hp-epoch-baseline1.tar.gz` on
both hosts. Archive SHA256:
`7805a95aa0c896b5ef6ae02ff79b98e926227aa575dca848efa8ae196753ab35`.
Seal SHA256:
`48302ef64c511b0ce02ecb7d7db2d9d31169369cefc4642ee4837be8597585d4`.
The Git selection retains 64 original metadata/log/raw files; raw files use
lossless gzip and their decompressed bytes must match the sealed originals.
The [independent verifier](evidence/hp-epoch-baseline-20261011/verify.py)
recomputes ordered tails, counts, quantiles/means/maxima, activity decisions
and policy fingerprints, and checks source and replica identities. Actual
HP high-priority shared execution passes both the initial review and a final
review that additionally recomputes the published aggregate matrix and checks
the exact frequency readback. Their complete 83/84-file source selections and
compressed review archives remain on both hosts. Original passing receipts
stay separate; the final review does not rewrite the earlier source.

After exact sealed-input/replica checks and a visible-consumer refusal guard,
[HP cleanup](evidence/hp-epoch-baseline-20261011/hp-cleanup.json) removes only
the completed source, output, operations and two review materializations plus
their staged tar files. Both complete evidence archives and review archives
remain. Local staging is also removed after its original source and archive
hashes match. No installed payload, cache, operator wrapper or another
session's path/process is changed. Both HP CPU pools are free afterward.

Initial staging refused because
this session's HP build parent no longer existed; creating that owning parent
and retransferring the identical capsule resolved it. The failure is retained.
Final evidence keeps raw order, deficits, summary extrema, source/compiler/
binary identities and all unsuccessful attempts. Derived plots are unnecessary;
retained raw is losslessly compressed rather than rounded to plot resolution.

The publication check originally compared the entire historical SCAN source
inventory against today's tree. It refused at the SCAN load profile: another
session's `e49628bfc` profile change already existed before this HP source pin.
The [original refusal](evidence/hp-epoch-baseline-20261011/input-comparison-failed.json)
is retained. The
[corrected integration check](evidence/hp-epoch-baseline-20261011/input-comparison-integrated.json)
confirms all 84 qualified HP review inputs unchanged, the other eleven SCAN
sources unchanged, and the concurrent profile preserved. That new current
profile is not backdated into the old frozen SCAN campaign. Raw SSH queue logs
retain their original CRLF bytes; whitespace checking allows CR at end of line
without rewriting sealed evidence.