d0b931b197
Step 1 & 2 Complete: - Implemented: core/front/malloc_tiny_fast.h prefetch (lines 264-267, 331-334) - LEGACY path prefetch of g_unified_cache[class_idx] to L1 - ENV gate: HAKMEM_TINY_PREFETCH=0/1 (default OFF) - Conditional: only when prefetch enabled + route_kind == LEGACY - A/B test (Mixed 10-run): PREFETCH=0 (39.33M) → =1 (39.20M) = -0.34% avg - Median: +1.28% (within ±1.0% neutral range) - Result: 🔬 NEUTRAL (research box, default OFF) Decision: FREEZE as research box - Average -0.34% suggests prefetch overhead > benefit - Prefetch timing too late (after route_kind selection) - TLS cache access is already fast (head/tail indices) - Actual memory wait happens at slots[] array access (after prefetch) Technical Learning: - Prefetch effectiveness depends on L1 miss rate at access time - Inserting prefetch after route selection may be too late - Future approach: move prefetch earlier or use different target Next: Phase 3 C2 (Metadata Cache Optimization, expected +5-10%) 🤖 Generated with Claude Code Co-Authored-By: Claude Haiku 4.5 <noreply@anthropic.com>
104 lines
2.9 KiB
Markdown
104 lines
2.9 KiB
Markdown
# Phase 3: Cache Locality 最適化（開始指示）
|
||
|
||
## 目標
|
||
|
||
**現状**: Mixed ~35.2M ops/s (B3+B4 後)
|
||
**目標**: 57-68M ops/s (+12-22%)
|
||
|
||
## Phase 3 構成
|
||
|
||
### C3（優先度: 🔴 最高）: Static Routing
|
||
|
||
**背景**:
|
||
- Mixed の perf top では malloc/policy_snapshot が hot
|
||
- 現在: 毎回 malloc 時に policy snapshot + learner evaluation → 大きな overhead
|
||
- 案: malloc_tiny_fast() 呼び出し前に "static route" を init 時決定
|
||
|
||
**設計メモ**: `docs/analysis/PHASE3_C3_STATIC_ROUTING_1_DESIGN.md`
|
||
|
||
**実装ステップ**:
|
||
|
||
1. **Profiling（現状把握）**
|
||
```bash
|
||
HAKMEM_PROFILE=MIXED_TINYV3_C7_SAFE ./bench_random_mixed_hakmem 1000000 400 1
|
||
perf record -F 99 --call-graph dwarf -- HAKMEM_PROFILE=MIXED_TINYV3_C7_SAFE ./bench_random_mixed_hakmem 1000000 400 1
|
||
perf report --stdio
|
||
# → malloc/policy_snapshot/learner がどの程度か確認
|
||
```
|
||
|
||
2. **Static Route Detection (init 時)**
|
||
- malloc_tiny_fast() が呼ばれる前に route を "決定"
|
||
- 対象: C0-C7 の class 別に「LEGACY が dominant か」を判定
|
||
- ENV gate: `HAKMEM_TINY_STATIC_ROUTE=1/0` (default 0)
|
||
|
||
3. **Route Bypass**
|
||
```c
|
||
// 現在（毎回評価）:
|
||
route = g_policy_learner->get_route(class_idx); // 高コスト
|
||
|
||
// C3 Static（init 時決定）:
|
||
if (static_route_enabled()) {
|
||
route = g_static_route[class_idx]; // cached, no learner
|
||
} else {
|
||
route = learner_route(...); // 従来通り
|
||
}
|
||
```
|
||
|
||
4. **期待**: +5-8%
|
||
5. **A/B Test**: Mixed 10-run + C6-heavy 5-run
|
||
|
||
### C1（優先度: 🟡 中）: TLS Cache Prefetch
|
||
|
||
**狙い**: policy ではなく、実際に alloc が触る **TLS cache**（Unified Cache）をプリフェッチ
|
||
|
||
**設計メモ**: `docs/analysis/PHASE3_C1_TLS_PREFETCH_1_DESIGN.md`
|
||
|
||
**実装**:
|
||
```c
|
||
// malloc_tiny_fast_for_class() 内で、LEGACY route のときだけ:
|
||
__builtin_prefetch(&g_unified_cache[class_idx], 0, 3);
|
||
```
|
||
|
||
**期待**: +2-4%
|
||
|
||
### C2（優先度: 🟡 中）: Slab Metadata Cache Optimization
|
||
|
||
**狙い**: hot metadata（policy, slab descriptor）をより近い場所に配置
|
||
|
||
**期待**: +5-10%
|
||
|
||
## 実装フロー
|
||
|
||
```
|
||
1. C3（Static Routing）実装 & A/B test
|
||
├─ GO: default 化
|
||
└─ NO-GO: freeze
|
||
|
||
2. C1（TLS Prefetch）追加実装
|
||
└─ Cumulative test
|
||
|
||
3. C2（Metadata Optimization）
|
||
└─ Final A/B
|
||
```
|
||
|
||
## 安全性確認
|
||
|
||
- **LD mode**: static route は LD 環境では disable（policy learner は LD で active）
|
||
- **Lock depth**: malloc 側なので不要
|
||
- **Rollback**: ENV gate で即時 OFF 可能
|
||
|
||
## 次のアクション
|
||
|
||
**今すぐ**:
|
||
1. `perf record -F 99 ./bench_random_mixed_hakmem` を実行して hot spot 特定
|
||
2. policy_snapshot / learner evaluation の overhead 定量化
|
||
3. C3 static route detection の実装開始
|
||
|
||
**以降**:
|
||
- A/B テスト（Mixed 10-run）で +5-8% 確認
|
||
- C1/C2 の段階的導入
|
||
|
||
---
|
||
|
||
**背中が見える段階から、さらに奥深く。**
|