Files
agentic-kvc/RESULTS_SUMMARY.md
Gahow Wang abde010b64 Add RESULTS_SUMMARY.md — concise Chinese summary of current findings
One-page distillation of what the paper can claim today, with figure /
data path next to each row. Sections:

  1. Workload 性质 — intra-session reuse, skew, KV footprint
  2. Dispatch Coupling — agentic vs chatbot inter-turn gap regime
  3. 现有调度三类失败 — load-balance / static PD-disagg / pure sticky
  4. PD-disagg cost vs benefit — MB2 (transfer 9.7 GB/s ceiling,
     topology-independent) + MB1 (decode halted during prefill 15-200x),
     joined into the §3.2 cost > benefit headline for any KV ≥ 80 MiB
  5. EAR 实证状态 — Pillar 1 (affinity) validated, Pillar 2 (migration)
     substrate validated + strategy-layer pending
  6. 已能写的 paper 主张(按 confidence 排序)
  7. 待做(MB3-5, migration e2e, wall-clock sweep, scale-out)

Designed to be the one doc to read when re-entering the project after
a break.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
2026-05-27 21:38:28 +08:00

6.7 KiB
Raw Blame History

目前已成立的结论2026-05-27

EAR 项目目前能用实测数据支撑的论点汇总。每条都标了对应的图/数据路径。


1. Workload 性质§2

Production trace = Qwen3-Coder agentic1.3 M sessions / 2.1 M reqs / 7200 s。

性质 数据 实证
KV 复用几乎全在 session 内 intra 93.2% / cross 5.7% / shared 1.1%;理论 APC 上界 79.6% figs/f2a_reuse_topology.png
Session 极度偏斜 top 1%/5%/10%/25%/50% = 46.5%/66.5%/74.6%/87.5%/96.0% input mass figs/f2b_session_skew.png
单请求 KV 已经很大 p50 1.8 GiB / p90 8.0 / p95 9.6 / p99 11.5 GiBKV pool 38 GiB/instance0.4 × H20 96 GiB→ p99 req 只能装 3 个/instance figs/f2c_kv_footprint_cdf.png

结论cache 是 session-local 的scheduling 必须保留 session affinity单 request KV 接近 pool 上限,PD-disagg 4P+4D 让系统 decode 容量直接减半

2. Dispatch Coupling§2.3

数据 Agentic (Qwen3-Coder) Chatbot (qwen3-max)
Inter-turn T_external p50 1.6 s 7.2 s
gap < 1 s 比例 39% 4%
gap < 5 s 比例 67% 29%
p99 738 s 43 s

参考图:figs/f3a_inter_turn_gap.png

结论agentic 有一段 chatbot 没有的 sub-second tool-call mode39% vs 4%)。当 W_turn ≳ T_external(任何 W_turn > 1 s 的 scheduler 在 agentic 上都满足这条件Little's Law L = Λ · N · (W_turn(L) + T_external) 进入闭环 regimescheduler 的 ε 退步通过 KV 竞争反馈环被放大成 wall-clock 数倍差距。实测lmetric 跑 600 s trace 用 49 min wall-clock = 8x amplification

3. 现有调度的三类失败§3

Baseline 失败模式 数据
load-balance / LMetric 丢 locality lmetric APC 56.9%vs 上界 79.6%LMetric 比 load_only 只好 +3.3pp,因为 cache 信号在乘性 score (pending+inputhit) × num_req 里被 num_req 吞掉
静态 PD-disagg D 侧 KV 容量墙 + transfer 成本 见 §4 cost-vs-benefit
Pure sticky 全员被 hot session 拖累,不是单一热点 sticky median worker 20.3 s vs unified 10.3 ssystem e2e p90 sticky 34.6 s vs unified 18.0 s用 max/median ratio 衡量是误导§3.3 用 absolute per-worker latency

参考图:figs/f4a_apc_loss.pngfigs/f4b_pdsep_kv_wall.pngfigs/f4c_per_worker_ttft.pngfigs/f6_e2e_latency_bars.pngfigs/f6_e2e_latency_full_grid.png

4. PD-disagg 在 agentic 下输——cost vs benefit§3.2

由两个独立 microbench 钉死(全用 vanilla vLLM 0.18.1 + Mooncake 0.3.11fresh venv无 patch)。

4.1 MB2 — KV transfer cost

dash1 GPU 0+1intra-node和 dash1 ↔ dash2inter-node, 200 Gbps RoCE扫 9 个 size × 5 reps。

路径 稳态带宽(≤ 3 GiB p99 agentic 请求11.5 GiBtransfer 时间
Intra-node 9.7 GB/s p50 1.9 s · min 1.5 s · max 10 s
Inter-node 10.0 GB/s(差 <3% p50 1.7 s · min 1.3 s · max 9.2 s

新发现intra/inter 几乎重合 → Mooncake batch_transfer_sync_write 永远走 RDMA NIC包括 intra-node loopback,不走 NVLink。200 Gbps NIC 是天花板,PD-disagg 的 transfer cost 与拓扑无关

参考图:figs/mb2_transfer_time_compare.pngfigs/mb2_transfer_bw_compare.png、doc analysis/mb2/README.md

4.2 MB1 — Phase interferencechunked-prefill on, 默认 baseline

dash1 GPU 0 单 instanceDconcurrent decodes× Pprefill size扫描。

D=8最 agentic-realistic的结果

Prefill prefill_ttft per-stream TPOT during penalty
2k tok 143 ms 32 ms 4×
8k 583 ms 114 ms 15×
32k 4.5 s 388 ms 52×
65k 15.6 s 757 ms 99×
131k 57 s 1419 ms 183×

baseline TPOT 7.7 ms。Decode 在大 prefill 期间基本被 halted。chunked-prefill 已经默认开启PD-disagg 在它之上能额外提供的 phase isolation = decode 在 prefill 期间被 halted 的那部分时间

参考图:figs/mb1_interference.png、doc analysis/mb1/README.md

4.3 联合结论

Per-request
Max PD-disagg benefit(救回来的 decode 时间) decode 时长 = 50200 msagentic tool-call output
PD-disagg costMB2 transfer p50 80 MiB ≈ 8 ms · 3 GiB ≈ 320 ms · 11.5 GiB ≈ 1.9 sp99 实测最差 10 s
Cost / Benefit 每个 KV ≥ 80 MiB 的请求都输trace 平均 KV 192 MiB → 已经输

结论:在 agentic 上 PD-disaggregation 是结构性失败的。Chunked-prefill 默认已经在 colocation 内做了 first-order phase isolationPD-disagg 在此之上能额外补的decode 短时段没被 prefill 挤)小于它新带来的(每个 routed 请求都付 KV transfer。这个结论与拓扑无关intra-node 和 inter-node 一样)。

参考图:figs/pd_cost_vs_benefit.png§3.2 headline

5. EAR 设计的实证状态§4

Pillar 已实证 待实证
Affinity-default routing (Pillar 1) Current unified 算法 = LMetric + high-cache affinityAPC 79.4%(达到 79.6% 上界 97%TTFT p90 7.3 smedian worker p90 10.3 s
Hot-triggered session migration (Pillar 2) substrate 已通:kv_both connector 在 trace replay 上 net positiveTTFT p90 18.6%DR-fix 后 36.6%),原 elastic_migration_v2 paper 的 "+45% kv_both penalty" obsolete e2e 策略层trigger 阈值 + target selection 在反馈环里)未直接验证

6. 已经能写的 paper 主张(按 confidence 排序)

  1. Agentic vs chatbot 在调度上是不同 regimedispatch coupling + sub-second tool-call mass—— 实证完整
  2. PD-disaggregation 在 agentic 下输cost > benefit跨拓扑—— MB1 + MB2 实证完整
  3. 三类现有调度 baseline 各自的失败模式 —— 实证完整
  4. Affinity-default 调度current unified达到 APC 上界per-worker latency 也压倒 sticky —— 实证完整
  5. Hot-triggered migration 修复 sticky 的 hot pin —— design 完整、e2e 待验证

7. 待做

  • MB3-5end-to-end PD-disagg deploymentD-pool runtime occupancy、cache reuse × PD interaction、PD ratio sweep。这些是 §5 完整实验矩阵的事
  • EAR Pillar 2 migration e2e validation(在 connector_tax DR-fix 之上重测)
  • §5.4 wall-clock amplification sweep5 baseline × 3 runs钉死 dispatch coupling 论证的实证 closure
  • Scale-out 验证dash1+dash2 = 16 GPU等 dash0 + 3-node 可用时扩到 80 GPU