Stage‑2 codec microbench report

Qualified host-local codec ranking. All 25 selected adapter windows had a passing 120-second quiet-host preflight and no competing process sample at or above 5% CPU in the five-second all-PID monitor. Benchmark processes and adjacent boundary samples were reviewed by process ownership. Rows were selected from 2 same-manifest segments. Campaign evidence and selection.

Synthetic in-memory FIX parser/builder benchmark. Generated 2026-10-03T03:00:43Z.

Batch averages. Each measured sample averages 32 operations. Percentiles describe per-operation batch averages; they do not timestamp each message separately.

Checked converted values. D/8 parse-only and parse-extract both check all application fields. W parse-only checks the top of book; W parse-extract checks all twenty entries. Values come from the public parser/field API and are compared with the independent converted-values-v2 oracle. Llfix supplies builds only; Falcon supplies D/8 only.

Host-local evidence. This page reports one machine and two selected timing segments. It presents retained summary percentiles only; it is not a high-definition CDF or HdrHistogram report. Do not merge or rank host pages without an explicit matched-hardware/campaign decision.

Process affinity. Adapter processes were launched with taskset on requested CPUs 18,19 (effective preflight affinity: 18,19). The launcher and builds used CPUs 0,1,2,3,4,5,6,7,8,9. CPU pinning and load readings alone do not verify that the host was quiet.

Prepared binaries. Compilation completed before this run from commit 4375a8fe04c0b7490a61413062d9b56868b13381 (tracked changes at preparation: no). The verified stage2_prepared_binaries.json SHA-256 is 8ec4385339f7e90a68b631c6155b725321f92a0e68fd20a0f8e592e417c2e1f2.

Two clock domains. Latency uses only relative monotonic clocks inside each process: C++: std::chrono::steady_clock; Rust: std::time::Instant; Java: System.nanoTime; Go: time.Since(monotonic time.Time); C#: System.Diagnostics.Stopwatch.GetTimestamp. The separate absolute capture envelope is 1790985176644749496–1790996313022230925 ns since Unix epoch and is never subtracted to produce a latency. Cross-host one-way latency is not qualified. The machine-readable companion is stage2_clock_contract.json.

Field-work parity. All selected negative fixtures were rejected through their field-work checks.

Adapter build and run status

Timed-run conditions

One-minute host load averages immediately before and after each adapter run. The full readings are in stage2_run_conditions.json. Check competing processes separately.

Headline p50 matrix nanoseconds; lower is better; click a column heading to sort

Parity mutation audit

Each probe changes field values while retaining the FIX BodyLength and checksum. A pass means the adapter rejected the stale 9001/9002 value through its own field-work check.

All operation summaries

Filter