24fad8f72f
Benchmark Results: - bench_random_mixed: hakmem 56.8M, system 84.5M, mimalloc 107M - bench_tiny_hot: hakmem 81.0M, system 156.3M - bench_mid_large_mt: hakmem 9.94M, system 8.40M (hakmem wins! +18.3%) Key Findings: 1. Tiny allocations: hakmem is 0.52x slower than mimalloc (main weakness) 2. Mid/Large MT: hakmem is 1.18x faster than system (strength!) 3. Identified Tiny Front as optimization target for Phase 4 This benchmark comparison informed the Phase 4 optimization strategy: - Focus on Tiny Front bottleneck (15-20 branches) - Target: 2x improvement via PGO + Hot/Cold separation + Config optimization - Expected: 56.8M → 110M+ ops/s (closing gap with mimalloc) 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
158 lines
4.8 KiB
Markdown
158 lines
4.8 KiB
Markdown
# HAKMEM Allocator - Benchmark Comparison (Phase 3)
|
||
|
||
**測定日**: 2025-11-29
|
||
**Phase**: Phase 3 (mincore完全削除後)
|
||
**Commit**: d78baf41c - Phase 3: Remove mincore() syscall completely
|
||
|
||
---
|
||
|
||
## 📊 ベンチマーク結果サマリー
|
||
|
||
### 1. bench_random_mixed (混合ランダムサイズ: 16-1024B)
|
||
|
||
**Parameters**: 1M iterations, working set=256
|
||
|
||
| Allocator | Throughput | vs hakmem |
|
||
|-----------|------------|-----------|
|
||
| **hakmem** | **56.8M ops/s** | 1.00x |
|
||
| system (glibc) | 84.5M ops/s | **1.49x faster** |
|
||
| mimalloc | 107.0M ops/s | **1.88x faster** |
|
||
|
||
**Result**: ❌ hakmem が最下位 (system の 67%, mimalloc の 53%)
|
||
|
||
---
|
||
|
||
### 2. bench_tiny_hot (固定サイズ: 64B, ホットパス)
|
||
|
||
**Parameters**: 64B allocations, batch=100, cycles=100K (20M total ops)
|
||
|
||
| Allocator | Throughput | Latency | vs hakmem |
|
||
|-----------|------------|---------|-----------|
|
||
| **hakmem** | **81.0M ops/s** | 12.35 ns/op | 1.00x |
|
||
| system (glibc) | 156.3M ops/s | 6.40 ns/op | **1.93x faster** |
|
||
| mimalloc | N/A (lib error) | N/A | - |
|
||
|
||
**Result**: ❌ hakmem が遅い (system の 52%)
|
||
|
||
---
|
||
|
||
### 3. bench_mid_large_mt (Mid/Large マルチスレッド)
|
||
|
||
**Parameters**: Default (multithreaded allocations)
|
||
|
||
| Allocator | Throughput | vs hakmem |
|
||
|-----------|------------|-----------|
|
||
| **hakmem** | **9.94M ops/s** | 1.00x ← **Winner!** |
|
||
| system (glibc) | 8.40M ops/s | 0.85x (slower) |
|
||
| mimalloc | timeout | - |
|
||
|
||
**Result**: ✅ **hakmem が勝利！** (system より +18.3% 速い)
|
||
|
||
---
|
||
|
||
## 🎯 総合評価
|
||
|
||
### 強み (hakmem が勝つケース)
|
||
- ✅ **Mid/Large マルチスレッド**: system より +18.3% 速い
|
||
- ✅ **効率性**: IPC, Branch Miss率, Cache Miss率は優秀（前回分析より）
|
||
|
||
### 弱み (hakmem が負けるケース)
|
||
- ❌ **Small 混合ランダム**: mimalloc の 53%, system の 67%
|
||
- ❌ **Tiny ホットパス**: system の 52%
|
||
|
||
---
|
||
|
||
## 📈 Phase 2 → Phase 3 の変化
|
||
|
||
### bench_random_mixed (ws=256)
|
||
- **Phase 2**: 51.3M ops/s (commit 0ce20bb83)
|
||
- **Phase 3**: 56.8M ops/s (commit d78baf41c)
|
||
- **Change**: **+10.7% 改善** ✓ (mincore削除の効果)
|
||
|
||
### Phase 3の変更内容
|
||
```
|
||
Phase 3: Remove mincore() syscall completely
|
||
- Makefile: DISABLE_MINCORE 設定削除
|
||
- hak_free_api.inc.h: mincore ロジック削除 (~60 lines)
|
||
- external_guard_box.h: is_mapped() を常に 1 を返すように簡略化
|
||
```
|
||
|
||
**Performance Impact**: +10.7% improvement (mincore overhead eliminated)
|
||
|
||
---
|
||
|
||
## 🔍 ボトルネック分析 (from PERF_ANALYSIS_EXECUTIVE_SUMMARY.md)
|
||
|
||
### Top 3 問題点
|
||
|
||
#### 1位: SuperSlab 初期化の Page Fault (23.83% CPU時間)
|
||
- `shared_pool_acquire_slab()` → memset(1MB-2MB) → page fault
|
||
- **原因**: mmap が既にゼロページを返すのに、4つの memset() を実行
|
||
- **対策**: Lines 912-915 in hakmem_tiny_superslab.c をコメントアウト
|
||
- **推定効果**: +10-15% throughput
|
||
|
||
#### 2位: mincore() syscall overhead (17.94% CPU時間) ← **Phase 3で解決済み**
|
||
- **Phase 2**: 毎回のfreeでmincore()呼び出し (TLSキャッシュあり)
|
||
- **Phase 3**: 完全削除 → +10.7% 改善確認済み ✓
|
||
|
||
#### 3位: getenv() overhead (9.86% CPU時間)
|
||
- 毎回のmalloc/freeでgetenv()呼び出し
|
||
- **対策**: TLSキャッシュ、または __attribute__((constructor))で初期化
|
||
|
||
---
|
||
|
||
## 🚀 Phase 4 最適化ターゲット
|
||
|
||
### Priority 1: SuperSlab memset() 削除 (Expected: +10-15%)
|
||
```bash
|
||
# Lines 912-915 in hakmem_tiny_superslab.c をコメントアウト
|
||
# mmap(MAP_ANONYMOUS) は既にゼロ初期化済み
|
||
```
|
||
|
||
### Priority 2: getenv() TLS キャッシュ (Expected: +5-8%)
|
||
```bash
|
||
# malloc/free の高頻度 getenv() 呼び出しをTLSキャッシュ化
|
||
```
|
||
|
||
### Priority 3: Branch Optimization/PGO (Expected: +3-5%)
|
||
```bash
|
||
# Profile-Guided Optimization + LIKELY/UNLIKELY hints
|
||
```
|
||
|
||
**Total Expected**: +18-28% improvement → **Target: 67-73M ops/s** (bench_random_mixed)
|
||
|
||
---
|
||
|
||
## 📌 次のアクション
|
||
|
||
### Option A: Phase 4 最適化開始
|
||
1. SuperSlab memset() 削除 (+10-15%)
|
||
2. getenv() TLSキャッシュ化 (+5-8%)
|
||
3. PGO/branch optimization (+3-5%)
|
||
|
||
### Option B: 詳細プロファイル (Phase 3 baseline)
|
||
```bash
|
||
# 現在のPhase 3ベースラインでperf record
|
||
perf record -F 9999 -g ./bench_random_mixed_hakmem 5000000 256 42
|
||
perf report --stdio --no-children --sort symbol
|
||
```
|
||
|
||
### Option C: mimalloc との徹底比較
|
||
- mimalloc のビルド修正（libmimalloc.so.1 パス設定）
|
||
- 全ベンチマークで3回実行して統計取得
|
||
|
||
---
|
||
|
||
## 🔗 参考ドキュメント
|
||
|
||
- `CHECKPOINT_PHASE2_COMPLETE.md` - Phase 2 Box化完了
|
||
- `PHASE2_PERF_ANALYSIS.md` - Phase 2 パフォーマンス分析
|
||
- `PERF_ANALYSIS_EXECUTIVE_SUMMARY.md` - ボトルネック分析
|
||
- `perf_phase2_*.txt` - Phase 2 詳細プロファイル結果
|
||
|
||
---
|
||
|
||
Generated: 2025-11-29
|
||
Phase: 3 (mincore removal complete)
|
||
Next: Phase 4 optimization targets (+18-28% expected)
|