[shinaoka]

Closes #6 (stage 2, proposal 3). Stage 1 is #7, released as 0.1.1. Item 4, where faer drops team_section_cost, is a separate faer PR that follows this release.

Summary

• for_class(HostClass::X86_64Avx2) gets a measured team row with ten anchors (widths 2 to 64). The data is from the comment on WidthCosts is flat above width 6: extrapolate conservatively and accept measured anchors up to 64 #6 (2026-10-29):

  • host broccoli: AMD EPYC 7702P (Zen2), 64 cores, SMT off, boost off, governor performance;
  • pool 64 pinned to CPUs 0-63, threadteam 0.1.0;
  • the Linux port of benches/dispatch_cost.rs in faer-benchmark;
  • statistic team/load10us-minus-10us, the same quantity as the existing rows, rounded to the nearest ns.

  width	2	4	6	8	12	16	24	32	48	64
  team, ns	1423	1992	2329	2738	3558	4076	4292	4838	6402	9333

  The raw output is checked in as docs/design/thresholds-data/raw/broccoli-dispatch-20261029.csv.

• The plain kinds (in pool, outside) keep the generic row on purpose. The run's plain-token numbers were a control and are not promoted to the row.

• What the row does not claim (docs/design/thresholds.md section 3.7):

  • The dispatch row is measured, but the consumer crates' rates for x86_64-avx2 are still generic: a heuristic exception to "one machine per class". The for_class rustdoc and placement.md now name it.
  • The class covers small AVX2 hosts too, and their width-2 team floor goes from 9 to 4.3 us. The row is a representative class default; PoolBuilder::dispatch_costs overrides it.
  • The intermittent ~0.62 ms section that faer#91 saw on a fresh team entry is not attributed and is not part of the steady-state row.

• Measured rows are now const items (anchored). An invalid anchor list fails the build instead of reaching the runtime expect from stage 1.

• threadteam 0.1.2 for the approved patch release.

Tests

tests/dispatch_cost.rs::avx2_class_has_the_measured_zen2_team_row checks:

• the ten anchors;
• width 10 interpolates to 3148 ns;
• width 128 extrapolates to 18666 ns;
• in pool and outside equal the generic row, including width 64.

x86_64-avx2 is removed from the generic-row loop.

Verification

The AGENTS.md pre-push checklist passes locally:

• cargo fmt --check
• cargo clippy (all features and no default features, both with -D warnings)
• cargo test (default, all features and no default features)
• loom loom_park
• cargo deny check bans
• scripts/check-no-public-rayon.sh
• cargo doc

Independent pre- and post-implementation reviews (gpt-6.1-sol) were done. The pre-review's two IMPORTANT points (mixed provenance, small hosts) are documented as above, and the post-review's minor wording points are fixed.

🤖 Generated with Claude Code