Phase 4 E2: Alloc Per-Class FastPath - NEUTRAL (-0.21%)
A/B Test Results (Mixed, 10-run, 20M iters): - Baseline (DUALHOT=0): 45.40M ops/s (mean), 45.51M ops/s (median) - Optimized (DUALHOT=1): 45.30M ops/s (mean), 45.22M ops/s (median) - Improvement: -0.21% mean, -0.62% median Decision: NEUTRAL (within ±1.0% noise threshold) Action: FREEZE as research box (default OFF, no promotion) Key Findings: - C0-C3 fast path adds branch overhead without measurable benefit - Unlike FREE path (+13%), ALLOC path already has optimized route caching - Phase 3 C3 static routing eliminated route lookup overhead - Additional per-class specialization doesn't reduce existing cost Root Cause: - Free DUALHOT skips expensive policy_snapshot() + tiny_route_for_class() - Alloc DUALHOT adds C0-C3 branch but route already cached (Phase 3 C3) - Net effect: Branch cost ≈ Route savings → neutral Conclusion: Alloc route optimization has reached diminishing returns Cumulative Status: - Phase 4 E1: +3.92% (GO, research box) - Phase 4 E2: -0.21% (NEUTRAL, frozen) Files: - CURRENT_TASK.md: Updated with E2 results - docs/analysis/PHASE4_E2_ALLOC_PER_CLASS_FASTPATH_AB_TEST_RESULTS.md: Full A/B test report 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Haiku 4.5 <noreply@anthropic.com>
This commit is contained in:
Moe Charm (CI)
@ -1,6 +1,43 @@
|
||||
# 本線タスク（現在）
|
||||
|
||||
## 更新メモ（2025-12-14 Phase 4 E1 Complete - ENV Snapshot Consolidation）
|
||||
## 更新メモ（2025-12-14 Phase 4 E2 Complete - Alloc Per-Class FastPath）
|
||||
|
||||
### Phase 4 E2: Alloc Per-Class FastPath ⚪ NEUTRAL (2025-12-14)
|
||||
|
||||
**Target**: C0-C3 dedicated fast path for alloc (bypass policy route for small sizes)
|
||||
- Strategy: Skip policy snapshot + route determination for C0-C3 classes
|
||||
- Reuse DUALHOT pattern from free path (which achieved +13% for C0-C3)
|
||||
- Baseline: HAKMEM_ENV_SNAPSHOT=1 enabled (E1 active)
|
||||
|
||||
**Implementation**:
|
||||
- ENV gate: `HAKMEM_TINY_ALLOC_DUALHOT=0/1` (already exists, default: 0)
|
||||
- Integration: `malloc_tiny_fast_for_class()` lines 247-259
|
||||
- C0-C3 check: Direct to LEGACY unified cache when enabled
|
||||
- Pattern: Probe window lazy init (64-call tolerance for early putenv)
|
||||
|
||||
**A/B Test Results** (Mixed, 10-run, 20M iters, HAKMEM_ENV_SNAPSHOT=1):
|
||||
- Baseline (DUALHOT=0): **45.40M ops/s** (mean), 45.51M ops/s (median), σ=0.38M
|
||||
- Optimized (DUALHOT=1): **45.30M ops/s** (mean), 45.22M ops/s (median), σ=0.49M
|
||||
- **Improvement: -0.21% mean, -0.62% median**
|
||||
|
||||
**Decision: NEUTRAL** (-0.21% within ±1.0% noise threshold)
|
||||
- Action: Keep as research box (default OFF, freeze)
|
||||
- Reason: C0-C3 fast path adds branch overhead without measurable gain on Mixed
|
||||
- Unlike FREE path (+13%), ALLOC path doesn't show significant route determination cost
|
||||
|
||||
**Key Insight**:
|
||||
- Free path benefits from DUALHOT because it skips expensive policy snapshot + route lookup
|
||||
- Alloc path already has optimized route caching (Phase 3 C3 static routing)
|
||||
- C0-C3 specialization doesn't provide additional benefit over current routing
|
||||
- Conclusion: Alloc route optimization has reached diminishing returns
|
||||
|
||||
**Cumulative Status**:
|
||||
- Phase 4 E1: +3.92% (GO, research box)
|
||||
- Phase 4 E2: -0.21% (NEUTRAL, frozen)
|
||||
|
||||
### Next: Phase 4 E3 - TBD (consult perf profile or pursue other optimization vectors)
|
||||
|
||||
---
|
||||
|
||||
### Phase 4 E1: ENV Snapshot Consolidation ✅ COMPLETE (2025-12-14)
|
||||
|
||||
@ -28,10 +65,6 @@
|
||||
|
||||
**Key Insight**: Shifting from shape optimizations (plateaued) to TLS/memory overhead yields strong returns. ENV snapshot consolidation represents new optimization frontier beyond branch prediction tuning.
|
||||
|
||||
### Next: Phase 4 E2 - Alloc Per-Class Fast Path
|
||||
- 指示書（SSOT）: `docs/analysis/PHASE4_E2_ALLOC_PER_CLASS_FASTPATH_NEXT_INSTRUCTIONS.md`
|
||||
- 設計メモ: `docs/analysis/PHASE4_E2_ALLOC_PER_CLASS_FASTPATH_1_DESIGN.md`
|
||||
|
||||
### Phase 4 Perf Profiling Complete ✅ (2025-12-14)
|
||||
|
||||
**Profile Analysis**:
|
||||
|
||||
@ -0,0 +1,134 @@
|
||||
# Phase 4 E2: Alloc Per-Class FastPath - A/B Test Results
|
||||
|
||||
## Test Configuration
|
||||
|
||||
**Date**: 2025-12-14
|
||||
**Baseline**: HAKMEM_ENV_SNAPSHOT=1 (E1 enabled, new baseline from E1)
|
||||
**Profile**: MIXED_TINYV3_C7_SAFE
|
||||
**Workload**: Mixed 16-1024B allocation/free pattern
|
||||
**Parameters**: 20M iterations, ws=400, 1 thread
|
||||
**Runs**: 10 iterations per configuration
|
||||
|
||||
## Hypothesis
|
||||
|
||||
**Strategy**: Apply DUALHOT pattern to alloc path (C0-C3 dedicated fast path)
|
||||
- Free path achieved +13% with C0-C3 DUALHOT (skip policy + route)
|
||||
- Alloc path should benefit similarly by bypassing route determination
|
||||
- Expected gain: +1-3% (conservative, based on free path success)
|
||||
|
||||
**Implementation**:
|
||||
- ENV gate: `HAKMEM_TINY_ALLOC_DUALHOT=0/1` (default: 0)
|
||||
- Integration point: `malloc_tiny_fast_for_class()` lines 247-259
|
||||
- C0-C3 check: Direct to LEGACY unified cache (skip policy snapshot + route lookup)
|
||||
|
||||
## Test Results
|
||||
|
||||
### Baseline (HAKMEM_TINY_ALLOC_DUALHOT=0)
|
||||
|
||||
```
|
||||
Run 1: 45,565,309 ops/s
|
||||
Run 2: 45,083,995 ops/s
|
||||
Run 3: 45,204,517 ops/s
|
||||
Run 4: 45,610,342 ops/s
|
||||
Run 5: 45,201,090 ops/s
|
||||
Run 6: 45,658,791 ops/s
|
||||
Run 7: 45,447,571 ops/s
|
||||
Run 8: 44,603,716 ops/s
|
||||
Run 9: 45,707,506 ops/s
|
||||
Run 10: 45,900,099 ops/s
|
||||
```
|
||||
|
||||
**Statistics**:
|
||||
- Mean: **45,398,294 ops/s**
|
||||
- Median: **45,506,440 ops/s**
|
||||
- StdDev: **379,641 ops/s** (0.84% CV)
|
||||
|
||||
### Optimized (HAKMEM_TINY_ALLOC_DUALHOT=1)
|
||||
|
||||
```
|
||||
Run 1: 45,007,059 ops/s
|
||||
Run 2: 46,283,502 ops/s
|
||||
Run 3: 44,590,279 ops/s
|
||||
Run 4: 45,187,469 ops/s
|
||||
Run 5: 45,769,085 ops/s
|
||||
Run 6: 44,955,784 ops/s
|
||||
Run 7: 45,516,635 ops/s
|
||||
Run 8: 44,959,890 ops/s
|
||||
Run 9: 45,523,059 ops/s
|
||||
Run 10: 45,256,729 ops/s
|
||||
```
|
||||
|
||||
**Statistics**:
|
||||
- Mean: **45,304,949 ops/s**
|
||||
- Median: **45,222,099 ops/s**
|
||||
- StdDev: **485,566 ops/s** (1.07% CV)
|
||||
|
||||
## Performance Delta
|
||||
|
||||
- **Mean gain**: **-0.21%** (45.40M → 45.30M ops/s)
|
||||
- **Median gain**: **-0.62%** (45.51M → 45.22M ops/s)
|
||||
- **Variance increase**: StdDev increased from 0.38M to 0.49M
|
||||
|
||||
## Decision
|
||||
|
||||
**Result**: ⚪ **NEUTRAL** (-0.21% within ±1.0% noise threshold)
|
||||
|
||||
**Criteria**:
|
||||
- GO: mean gain >= +1.0%
|
||||
- NEUTRAL: mean gain in [-1.0%, +1.0%]
|
||||
- NO-GO: mean gain <= -1.0%
|
||||
|
||||
**Action**: **FREEZE** as research box (default OFF, no promotion)
|
||||
|
||||
## Analysis
|
||||
|
||||
### Why NEUTRAL (not GO)?
|
||||
|
||||
1. **Minimal performance impact**: -0.21% is within measurement noise
|
||||
2. **Free path comparison**: Free DUALHOT achieved +13%, alloc shows -0.21%
|
||||
3. **Branch overhead**: C0-C3 check adds branch cost without measurable benefit
|
||||
4. **Route caching effectiveness**: Phase 3 C3 static routing already optimized route lookup
|
||||
|
||||
### Root Cause Analysis
|
||||
|
||||
**Why alloc path differs from free path**:
|
||||
|
||||
1. **Free path (DUALHOT=1, +13% gain)**:
|
||||
- Skips expensive `policy_snapshot()` TLS read
|
||||
- Skips `tiny_route_for_class()` lookup
|
||||
- Direct to `tiny_legacy_fallback_free_base()`
|
||||
- Large savings: ~10-15 cycles per C0-C3 free
|
||||
|
||||
2. **Alloc path (DUALHOT=1, -0.21% neutral)**:
|
||||
- Route already cached via `tiny_static_route_ready_fast()` (Phase 3 C3)
|
||||
- Policy snapshot lightweight (TLS cache hit)
|
||||
- C0-C3 check adds branch before route lookup
|
||||
- Net effect: Branch cost ≈ Route lookup savings → neutral
|
||||
|
||||
### Key Insight
|
||||
|
||||
**Alloc route optimization has reached diminishing returns**:
|
||||
- Phase 3 C3 (static routing) already eliminated route lookup overhead
|
||||
- Additional C0-C3 specialization adds branch without reducing existing cost
|
||||
- Unlike free path, alloc path doesn't have "expensive operation to skip"
|
||||
|
||||
**Conclusion**: Per-class specialization effective only when bypassing measurable overhead. Alloc path route caching already optimal.
|
||||
|
||||
## Next Steps
|
||||
|
||||
**Freeze E2**:
|
||||
- Keep `HAKMEM_TINY_ALLOC_DUALHOT` as research box (default OFF)
|
||||
- No promotion to MIXED_TINYV3_C7_SAFE preset
|
||||
- Implementation retained for potential future workload-specific use
|
||||
|
||||
**Future Optimization Vectors**:
|
||||
- Consult perf profile for new hot spots (E1 changed baseline)
|
||||
- Consider non-routing optimizations (e.g., TLS layout, memory access patterns)
|
||||
- Explore workload-specific specialization (C6-heavy may differ from Mixed)
|
||||
|
||||
## References
|
||||
|
||||
- Design: `docs/analysis/PHASE4_E2_ALLOC_PER_CLASS_FASTPATH_1_DESIGN.md`
|
||||
- Instructions: `docs/analysis/PHASE4_E2_ALLOC_PER_CLASS_FASTPATH_NEXT_INSTRUCTIONS.md`
|
||||
- Phase 4 E1 (ENV Snapshot): +3.92% GO (research box)
|
||||
- Phase 3 C3 (Static Routing): +2.20% ADOPT (default ON)
|
||||
Reference in New Issue
Block a user