20-Agent Deep Audit Report
Date: 2026-04-19 21:55
Scope: Method "Stall-Aware Embedding Layout + E+L0 Dual-Path Routing + Host LFU L0"
Wave 1: 12 agents(narrative / prior art / code / fairness)
Wave 2: 8 agents(statistics / simulation / data / traceability)
💀 論文不可送審的 Catastrophic Blockers(需立刻修)
1. RESULTS.md Optane row 數字錯 → 結論 INVERT
- Claim "MaxEmbed 5.62 ms/batch, CV 5.4%, 贏 E+L0 2.31x"
- 真實:MaxEmbed overall_us = 15.80 ms/batch, CV 21.3%
- E+L0 在 Optane 也贏 9.5x,不是輸 2.31x
- Agent 13: parse unit swap(5.62 是 per-sample μs,5.4% 是 ops/s CV)
- 立刻修 RESULTS.md + 所有引用該結論的地方
2. Paper 捏造 stall + scheduling 實驗數字
- Agent 3 + 20 雙重確認:
- 所有 stall numbers(80.10 / 11.47 / 6.02 / 4.04 / 5.08 / 7.80 / 10.38 / 14.41 / 3844 ms)都是 analytic 推導,但 paper 寫成 "MQSim cycle-accurate 輸出"
- 隱含 per-miss latency 在不同 step 不一致(7.56/37/47/9.34μs 不可能同時為真)
- Tagged flash 234,895 完全虛構(實際 92,331)
- Scheduling flash 5,160/5,224/26,132/31,600 虛構(沒跑過 miss_first/hit_first/spread_miss 實驗)
- SE spread-miss 29,532 MISLABELED(實際是 cog_v2_se K10000)
- 行動:重跑 scheduling 實驗或改寫 paper 去掉假數字
3. Paper Narrative 3 docs 互相矛盾 + E+L0 核心未寫
- Agent 1 + 19 雙重確認:
paper_story.md推 miss-firstalgorithm_design.md推 COG v2 spread-misspage_oriented_embedding_spec.md是舊 project- E+L0 Dual-Path Routing + Host LFU L0(宣稱的核心)3 個 doc 都沒寫
dual_path_pollution_model.md是 100% vapor claim:Score_MEM / PC(p) / ConflictRisk / 最佳 split k* / frequency-sketch admission filter 全無 code 實作,code 用 hardcoded EWMA_THETA=0.926f(archived doc previously labelled the admission filter "TinyLFU" — that claim is now withdrawn; the shipped L0 is aging LFU, Arlitt et al. 2000)- 行動:完全重寫 paper_story + algorithm_design,用 E+L0 + Host LFU L0 當核心
4. MaxEmbed Tail 比較不成立
- Agent 14:MaxEmbed 報的是 windowed avg 100-200 batches(CLT 壓低 variance ~10x)
- 我方 E+L0 是 true per-batch
- Paper claim "MaxEmbed tail 窄 p99/mean=1.04" 是 averaging artifact
- 真實 MaxEmbed tail 未知
batch_latencies.csv根本沒存,每 run 被覆蓋- 行動:不能 claim tail,或重跑兩邊都 dump per-batch CSV
5. Dataset Scale 兩個 dataset 不對
- CriteoTB: 宣稱 882M/4.37B,實際 126.6M/450M(-85% / -90%),HF 10% subsample
- Alibaba-iFashion: 宣稱 4.46M items,實際 2.18M(-51%)referenced in 999K outfits
- 行動:修 paper wording 或下 Tianchi 370GB full
🔴 Code Bugs(修不難,立刻動)
| # | Bug | File:Line | Agent | Severity |
|---|---|---|---|---|
| 6 | Mode 2 timer 包 posix_memalign + pthread_create |
race_trace_v3.c:580 |
7 | 🔴 |
| 7 | nhit = nm Mode 2 永遠 report miss=0,CSV 說謊 |
race_trace_v3.c:589 |
7 | 🔴 |
| 8 | Mode 2 無 _mm_lfence(Mode 4 有)→ OoO overlap 讓 Mode 2 假快 |
race_trace_v3.c |
7 | 🔴 |
| 9 | l0_hit_primary_or_replica 定義但 unused(replicas dead code) |
race_trace_v3.c |
7 | 🔴 |
| 10 | LFU tie-break 永遠 bias way-0(comment 說 xorshift 但沒實作) | race_trace_v3.c:252-258 |
8 | 🔴 |
| 11 | LFU hash pg % sets 非均勻(table_offset bit clustering) |
race_trace_v3.c:231 |
8 | 🔴 |
| 12 | LFU 16-bit counter 無 aging → scan resistance failure | race_trace_v3.c:254 |
8 | 🟡 |
| 13 | Criteo encoding insertion-order(paper 是 descending-freq) | criteo_raw_to_maxembed.py |
18 | 🔴 |
| 14 | L0 預設 16384 (64MB),但 paper 聲稱 4096/16MB | race_trace_v3.c L0_CAPACITY |
8 | 🟡 |
| 15 | Alibaba v1 converter(outfit-based)需 delete | convert_alibaba_ifashion.py |
18 | 🟡 |
🔴 Baseline Fairness(MaxEmbed 被削弱)
Agent 11 發現 4 個參數不對齊 paper AE:
| 參數 | Paper AE | 我方 | 影響 |
|---|---|---|---|
-n threads |
64 | 1 | 64x 少 concurrency |
-b batch_size |
32 (Criteo) | 16 | 減半 |
-c cache_ratio |
0.01-0.4 sweep | 0.0039(低於 paper range) | off-regime |
rep_ratio |
sweep 5 個 | single | 選最差 rep_ratio? |
行動:補 anchor run at -n 64 -b 32 -c 0.01,讓 MaxEmbed 跑在它被調過的 regime
📚 Must-cite Competitors(Agent 4, 5, 6)
| Paper | Venue | 重疊 |
|---|---|---|
| CARINA | SIGMOD'25 | ~70% overlap on LFU filter for DLRM CXL(最嚴重) |
| Fleche | EuroSys'22 | "probability-based filter policy" 已用 filter cache 術語 |
| RomeFS | SoCC'24 | Dual-path CXL.mem + CXL.io(非 DLRM,但協議層已被做) |
| Hotline | MLSys'23 | DLRM + DRAM/SSD tiering,直接 predecessor |
| FlexEMR | EuroSys'24 | 3-tier DRAM/CXL/SSD recsys |
| UGACHE | SOSP'23 | Near-optimal hotness policy for embeddings |
| Merlin HPS | NVIDIA | 3-tier GPU/CPU/SSD 已建立 pattern |
| RecShard | ASPLOS'22 | HBM+DRAM embedding sharding |
| RecSSD | ASPLOS'21 | SSD embedding caching |
| TPP / Pond / TMO / HeMem | ASPLOS'22-23, SOSP'21 | Generic OS tiering(orthogonal) |
我方剩下的 novelty:CXL-SSD × DLRM × stall-aware 三維交集(single-device dual-path + recsys-specific + offline+online hybrid)
🟡 Math / Statistical Rigor(Agent 2, 13, 14)
| Issue | Agent |
|---|---|
dual_path_math_theory.md 基本空 |
2 |
| Threshold 推導 0.64 vs 0.761 矛盾,中間推導缺失 | 2 |
ĥₚ, α, f̄, effective_parallelism 都 undefined |
2 |
| Independence assumption 被 PC(p) 破壞(coupling) | 2 |
| n=5 對 CV>10% 不夠,CV 21% 要 n≥17,CV 50% 要 n≥95 | 13 |
| E+L0 Optane ConfigD CV=49.6%(run3 spike 39.5ms)系統擾動 | 13 |
| Warmup 沒 drop,B0 冷啟 3.4ms 主導 p99 | 14 |
| 5x 的 500 樣本只有 ~100 獨立(trace replay) | 14 |
| 1/100 sampling rate,p9999 無法取得 | 14 |
🟡 Simulation Fidelity(Agent 15, 16)
| Issue | Status |
|---|---|
| FEMU direct-mapped vs CMM-H 8-way SA | 🟡 config 一行可修 buffer_way=3 |
| FEMU FIFO vs CMM-H LRU + MRU | 🔴 FEMU 無 LRU 實作,需寫 lru.c |
| FEMU MEM MISS (60μs) < NAND (68μs) inverted | 🔴 已知 bug,絕對數不可信 |
| FEMU buffer 5MB vs CMM-H 48GB(1:10000) | 🟡 residency-budget framing 可辯護 |
| SPDK TCP loopback 加 10-15μs/IO vs PCIe 5μs | 🟡 conservative 偏向(compressed speedup) |
| SPDK p99 = avg(真 P4510 p99 ≈ 2.5x avg) | 🟡 對我方 claim 保守 |
SPDK ch_xfer_lat = 0(真 ONFI 3-4μs) |
🟡 minor |
📊 Verdict
Paper 目前狀態:不可直接送審
主要問題:
- 核心結論數字錯(Optane invert)
- Paper 捏造 stall + scheduling 實驗數字
- Narrative 描述的方法跟實際 code 不一致
- Tail latency 比較不合法
- Dataset 規模 2 個不對 paper
建議修復順序(優先級最高 → 最低):
- 🚨 修 RESULTS.md Optane row + 重新算勝負表
- 🚨 重寫 paper_story.md / algorithm_design.md 為 "E+L0 + Host LFU L0"
- 🚨 刪 paper_story 捏造 stall 數字 + 改寫 scheduling table(或 rerun 補實驗)
- 🚨 改 pollution_model 為 "future work / analytic bound" 或實作
- 🔴 修 code bugs(Mode 2 timer, nhit=nm, LFU tie-break, encoding)
- 🔴 補 MaxEmbed
-n 64 -b 32 -c 0.01anchor run - 🔴 加 CARINA / Fleche / RomeFS / Hotline 等 related work
- 🟡 補 Criteo TB Tianchi full + Alibaba items 正確標註
- 🟡 修 FEMU buffer_way=3 + 考慮 LRU policy
- 🟡 補 tail latency 正確 instrumentation
預估整體修復 1-2 週(主要 paper 重寫 + rerun scheduling 實驗 + bug fix)。
📋 20 Agents 各自 Report Links
| Agent | Focus | Flag level |
|---|---|---|
| 1 | Narrative audit | 🔴 |
| 2 | Math rigor | 🔴 |
| 3 | Claims vs data | 💀 |
| 4 | CXL+DLRM prior art | 🔴 |
| 5 | L0 filter prior art | 🔴 |
| 6 | Dual-path prior art | 🔴 |
| 7 | race_trace_v3.c audit | 🔴 |
| 8 | LFU L0 impl | 🔴 |
| 9 | FEMU mods | 🟡 |
| 10 | Timer placement | 🟡 |
| 11 | MaxEmbed baseline | 🔴 |
| 12 | HW param P4510 | 🟡 |
| 13 | Variance CV | 💀 |
| 14 | Tail latency | 💀 |
| 15 | CMM-H fidelity | 🔴 |
| 16 | SPDK TCP overhead | 🟡 |
| 17 | Dataset scale | 🔴 |
| 18 | Preprocess alignment | 🔴 |
| 19 | Claim ↔ code | 💀 |
| 20 | Figure → data | 💀 |
💀 = catastrophic blocker
🔴 = significant flag
🟡 = concern, fixable