tomoaki/hakmem
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
d355041638d0b77d0bf2f8b4928e8201dbd52f95
hakmem/docs/archive/MAINLINE_INTEGRATION.md
Moe Charm (CI) 52386401b3 Debug Counters Implementation - Clean History 
Major Features:
- Debug counter infrastructure for Refill Stage tracking
- Free Pipeline counters (ss_local, ss_remote, tls_sll)
- Diagnostic counters for early return analysis
- Unified larson.sh benchmark runner with profiles
- Phase 6-3 regression analysis documentation

Bug Fixes:
- Fix SuperSlab disabled by default (HAKMEM_TINY_USE_SUPERSLAB)
- Fix profile variable naming consistency
- Add .gitignore patterns for large files

Performance:
- Phase 6-3: 4.79 M ops/s (has OOM risk)
- With SuperSlab: 3.13 M ops/s (+19% improvement)

This is a clean repository without large log files.

🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
2025-11-05 12:31:14 +09:00

2.0 KiB
Raw Blame History

Mainline Integration Plan (Tiny focus)

What we promoted to mainline (safe, general)

  • Safer tiny refill into SLL (already integrated)
    • sll_refill_small_from_ss(class_idx, max_take) now caps refill by the actual SLL free capacity, avoiding overtake and wasted meta->used increments.
  • Entry order consolidation in the small fast path
    • Prefer SLL → Magazine → SuperSlab in the normal (non‑bench) path; Quick/FrontCache/Ultra tiers remain opt‑in.
  • Targeted remote‑drain queue (compiled in, default OFF)
    • Per‑class Treiber queue for slabs that exceed remote thresholds; disabled by default via env knobs to preserve conservative behavior.
  • PGO recipe (opt‑in)
    • Makefile targets for PGO on tiny benches are available, but not required for normal builds.

What remains bench‑only (NOT promoted)

  • SLL‑only front (Magazine compiled out) and TLS warmup
    • These are highly benchmark‑specific and are kept in bench‑only builds.
  • Free‑side SLL‑first push without owner/stats
    • Mainline preserves learning/stats semantics; bench builds can cut them out.
  • Quick/FrontCache / 32/64 specialization hardwiring
    • Retained as A/B options; not enabled by default in mainline.

Recommended “Perf‑Main” preset (opt‑in, no bench macros)

  • Environment (Tiny‑Hot biased but general):
    • HAKMEM_TINY_TLS_SLL=1
    • HAKMEM_TINY_REFILL_MAX=96
    • HAKMEM_TINY_REFILL_MAX_HOT=192
    • HAKMEM_TINY_SPILL_HYST=16
    • HAKMEM_TINY_BG_REMOTE=0 (keep targeted remote drain off by default)
    • Keep Quick/FrontCache/Ultra OFF unless explicitly A/B tested

How to try Perf‑Main locally

  • Build benches: make bench_fast
  • Run tiny‑hot triad (no bench macros): bash scripts/run_tiny_hot_triad.sh 60000
  • Run random‑mixed matrix: bash scripts/run_random_mixed_matrix.sh 100000

Notes

  • LD_PRELOAD/app mode remains conservative (LD_SAFE staging). Tiny‑only and pass‑through modes are recommended for stability. The bench‑only optimizations are intentionally not applied in LD mode.