Files
agentic-kvc/microbench/fresh_setup/PD_DISAGG_INVESTIGATION.md
Gahow Wang 4a93096c1e Add PD_DISAGG_INVESTIGATION.md — living TODO for proving H1–H4
We don't have paper-grade evidence yet that PD-disagg fails in agentic.
MB1+MB2 corrected accounting puts phase-isolation cost-benefit on
PD-disagg's side; the only direct support is colleague's one data point
on a patched dash0 build (TTFT p50 62×, success 52%) and the f4b
geometric capacity argument. To close §3.2 properly we need fresh-venv
empirical replication PLUS system-level instrumentation that tells the
reviewer *which* component is the bottleneck — not just headline
latency.

This document tracks the four candidate failure hypotheses (H1 D-pool
capacity, H2 static-partition mismatch, H3 cache reuse + P-pool hotspot,
H4 end-to-end throughput loss), their current evidence status, and the
phased experiment plan to address each.

Key findings already recorded:
- Phase 0 TODO 0.1 (find standard PD-disagg deployment) is done — vLLM
  ships an official example at
  examples/online_serving/disaggregated_serving/mooncake_connector/
  with a kv_producer+kv_consumer launcher and a Mooncake-aware proxy
  that supports arbitrary P:D ratios via env vars. Per user direction,
  we will NOT polish PD-disagg policy ourselves; we use the official
  recipe as the "PD-disagg" baseline in §3.2 / §5.2.
- Phase 1 (MB5+3 combined: PD ratio sweep with D-pool occupancy logging)
  is the critical path. Designed to either confirm H1 with system
  breakdown evidence (D-pool ≥ 90% for ≥ 30% of trace + queue depth
  spike) or falsify it (some ratio matches 8C colo, in which case §3.2
  needs rewriting).
- D-pool occupancy timeline is the single most important new
  instrumentation — turns "PD-disagg is 10× worse" into "PD-disagg is
  10× worse BECAUSE the D pool sits at >90% for X% of the trace".

Configurations to run on dash1 8-GPU first:
  8C (colo baseline), 6P+2D, 4P+4D, 2P+6D  ×  3 reps  ×  w600 trace.

Open question still in the doc: vLLM 0.18.1 had an AttributeError on
self.bootstrap_server in kv_consumer mode when we hit it during MB2
sanity; likely the issue was bad kv_transfer_params from our side
(missing transfer_id, wrong field names), which we have since fixed.
Official proxy uses the same handshake we now have, so it should
just work. If not, single-line patch to initialize self.bootstrap_server
= None for consumer mode.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
2026-05-27 22:24:31 +08:00

9.5 KiB
Raw Blame History

PD-disagg 在 agentic 下到底为什么不 work — 系统行为级论证

Living TODO 文档。记 H1/H2/H3/H4 四条假设的实验状态,以及每个实验需要的 instrumentation、figures、判定标准。所有 microbench 文件名前缀沿用 MB


当前状态2026-05-27

问题陈述:之前关于 "PD-disagg 在 agentic 失败" 的证据只有 colleague 一个数据点TTFT p50 62×、success 52%,在 dash0 patched 栈上。MB1+MB2 算正确账后 phase-isolation 反而 favor PD-disagg所以我们目前没有 paper-grade 实证证据。本文档跟踪需要的实验和 system-level 分析。

核心 gap:我们只看到 "PD-disagg 表面性能差 10×" 的 headline 数字,但没有 system-level breakdown 告诉 reviewer "为什么差,差在哪个组件"。需要的是 D-pool occupancy / scheduler queue depth / KV transfer queue / GPU SM utilization 等系统指标的时间序列,能直接指出 bottleneck。


4 条独立失败假设

假设 描述 当前证据 状态
H1 D 池容量天花板 p99 单请求 11.5 GiB / D-instance 38 GiB pool4P+4D 减半系统 decode 容量;长 context overflow → 排队 colleague 旧数据 + f4b 几何论证 need fresh-venv replication + system breakdown
H2 静态分池 mismatch 工作负载 P:D 比例随时间漂移;静态 ratio 无法适应 not started
H3 Cache reuse 退化 + P 池 hotspot round-robin P 丢 affinityAPC 跌sticky P 复现 §3.3 hot pin logical extension not started
H4 端到端 throughput 输 各 phase 看着 OK 但 wall-clock 输dispatch coupling 放大) not started

Phase 0 — Infrastructure必须先 done

TODO 0.1: 找 vLLM 标准 PD-disagg 部署方案 DONE

结论:用 vLLM 仓库 ship 的官方 example third_party/vllm/examples/online_serving/disaggregated_serving/mooncake_connector/

  • run_mooncake_connector.sh —— 参数化 P/D GPU 列表、ports、bootstrap直接支持任意 P:D 比例e.g., PREFILL_GPUS=0,1,2,3 DECODE_GPUS=4,5,6,7 起 4P+4D
  • mooncake_connector_proxy.py —— 官方 FastAPI proxyround-robin P + round-robin D每个请求 fire-and-forget 到 P 做 do_remote_decode={transfer_id},并行用 P 的 (bootstrap_addr, engine_id) 触发 D 做 do_remote_prefill={remote_bootstrap_addr, remote_engine_id, transfer_id}

部署形态P 实例用 kv_role:kv_producer(带 bootstrap_serverD 实例用 kv_role:kv_consumer无 bootstrap_server,正是之前我们撞到 AttributeError 的那个 mode

  • 修改 run_mooncake_connector.sh 适配我们环境:模型路径 /home/admin/cpfs/wjh/models/Qwen/Qwen3-Coder-30B-A3B-InstructTP=1所需 vLLM 启动 flags--max-model-len 200000 等)
  • 重新验证 kv_consumer 模式:之前我们碰到的 AttributeError 可能是 kv_transfer_params 不对missing transfer_id 或字段名错MB2 调试后我们已经 nail down 正确 handshake官方 proxy 用的就是同样的 handshake所以这条很可能直接 work
  • 如果 kv_consumer 仍然报 AttributeError加一个 minimal patch 到 vLLM之前看过 self.bootstrap_server 在 consumer 模式没被初始化)—— 单行修复

TODO 0.2: 包装官方 launcher

直接基于 run_mooncake_connector.sh 改:

  • microbench/fresh_setup/mb5_pd_launch.sh —— 直接 sourceable参数 PREFILL_GPUS / DECODE_GPUS / PREFILL_PORTS / BOOTSTRAP_PORTS / DECODE_PORTS / PROXY_PORT,启动 P + D + 官方 proxy
  • 注入正确模型路径 + vLLM flags--max-model-len 200000 --gpu-memory-utilization 0.9 --enable-prefix-caching --max-num-batched-tokens 8192
  • 配置覆盖:先做 4 个 — 8C colo + 6P+2D + 4P+4D + 2P+6D
  • colo baseline 8C 不走这个 launcher沿用 dash0 现有 8-instance unified setup最公平的对照

TODO 0.3: System-level instrumentation

不止看 latency / success rate要记 system 行为时间序列才能 attribute bottleneck。

必须采集的 metrics(每秒一次,每实例):

  • vllm:gpu_cache_usage_perc — KV pool 占用率(核心 H1 证据)
  • vllm:num_requests_running — 并发 in-flight 数
  • vllm:num_requests_waiting — scheduler 排队数
  • vllm:time_to_first_token_seconds (histogram) — 每 instance TTFT 分布
  • vllm:time_per_output_token_seconds (histogram) — TPOT 分布
  • D 池请求 admission control 拒绝事件vllm 拒收新请求时的事件)—— 看 vllm 是否暴露 metric
  • Mooncake 侧:send_blocks 事件MB2 instrument 已存在);可选 transfer queue depth

实现

  • metrics_sampler.py:周期性 GET /metrics 解析 prometheus 文本,输出 jsonl
  • 每个 instance 一个采样进程 / 或者一个集中采样进程拉所有 instance
  • 输出 schema: {t_unix, instance_id, role, kv_pool_perc, num_running, num_waiting, ...}

TODO 0.4: D 池 occupancy timeline 可视化

  • plot_d_pool_timeline.pyheatmap 或 stacked area
    • x 轴trace replay 时间
    • y 轴:每个 D-instanceheatmap或 KV pool 总占用比stacked area
    • 色彩:占用率 0100%
    • 标 90% 红线("vllm stops admitting new requests" 阈值,参考 colleague 旧数据)
  • Output: 每个 config 一张图stacked 起来对比

Phase 1 — MB5+3: PD ratio sweep + D-pool occupancy 实测

Question: 任何 static P:D ratio 能否打过 8C colo如果不能瓶颈在哪

矩阵

configs = [8C, 6P+2D, 4P+4D, 2P+6D]
trace = traces/w600_r0.0015_st30.jsonl (1.2k req, ~13 min if not stalled)
reps = 3

Primary metrics (per config, mean of 3 reps)

  • TTFT / TPOT / E2E (mean / p50 / p90 / p99)
  • Success rate
  • Wall-clock time to drain trace(→ dispatch coupling amplification
  • APC (effective prefix cache hit ratio)
  • System throughput (req/s steady state)

System breakdown (the kill-shot evidence)

  • D-pool occupancy timeline per D-instance
  • Scheduler queue depth timelinewaiting requests over time
  • TTFT per request scattercolored by which-instance— 看 D 池满了的时候 TTFT 是不是直接挂掉
  • Admission denial events(如果 vllm 暴露)

Output figures

  • mb5_latency_bars.png — config × TTFT/TPOT/E2E p90 bar
  • mb5_success.png — success rate per config
  • mb5_wallclock.png — 实测 trace replay 时间 vs 8C colo
  • mb5_d_pool_timeline.png — 4 configs × 8 instances heatmap
  • mb5_queue_depth_timeline.png — 同上结构
  • mb5_diagnostic_summary.png — 1 张总图把上面 5 个塞进去给 paper 用

判定标准H1 何时被 confirmed

H1 算 被证实 如果同时满足:

  1. PD-disagg 在所有非 8C config 上 TTFT p90 ≥ 2× 8C colo
  2. D-pool occupancy timeline 显示 D 池 ≥ 90% 持续超过 trace 时长的 30%
  3. Scheduler queue depth 在 PD-disagg 下显著高于 8C colo
  4. 至少一个 PD config 出现 success rate 下降

H1 算 被证伪 如果存在某 PD ratio 使所有指标和 8C colo 差不多 —— 那 §3.2 要重写。


Phase 2 — System analysis钉死 bottleneck

Phase 1 数据出来后,写 system-level analysis

  • 对 4P+4D 配置:哪条 D-instance 哪段时间到 90%?对应那时进来的请求是不是大 KV 的?
  • 比对 8C 和 4P+4D8C 同样时间窗内每个 instance 的 occupancy 是什么样?是不是因为有 prefill workload 帮忙 churn 走?
  • Wall-clock difference attribution用 system metrics 把 PD-disagg 比 8C 慢的 X 秒分解成 (queue_wait + decode + transfer + ...)
  • 看 vllm log 里有没有 "scheduler skipped due to no available KV blocks" 之类事件

预期写一段:

"在 4P+4D 配置下D 池在 trace 第 X 秒到第 Y 秒持续超过 90% 占用,这段时间 prefill 完成的请求(来自 P 池)平均排队 Z 秒才能进入 decode导致 TTFT p90 从 8C 的 18s 涨到 60s。在 8C 配置下相同时刻 KV 池只到 60%,因为 prefill 完成后 immediately decode 而不需要排队 admission。"


Phase 3 — MB4 cache-reuse × PD-routingsecondary

只在 H1 没被钉死的情况下才优先做;否则当 supporting 实验。

  • 4P+4D 固定,扫 P-routinground-robin / session-sticky / kv-aware
  • APC, P-pool per-worker hotspot, end-to-end TTFT
  • 预期round-robin 丢 ~23pp APCsticky P 池复现 §3.3 hot pin

Phase 4 — 可选 scale-out16 GPU on dash1+dash2

  • Phase 1 结论稳定后才动
  • 配置16C / 8P+8D / 4×(2P+2D)(节点内 disagg + 跨节点协调)
  • 用 cost model如有外推 80 GPU

时间表(粗估)

  • Phase 0infrastructure: 2-3 天 — 调研 vllm 标准部署 + 改 launcher + 加 metrics 采样
  • Phase 1MB5+3 sweep: 1 天跑 + 1 天分析 = 2 天
  • Phase 2system analysis 写 paper section: 1-2 天
  • Phase 3MB4: 2-3 天(可选)

总计:~1 周钉死 H1 + Phase 2 写作。


Notes / open questions

  • vLLM 0.18.1 kv_consumer mode 的 bootstrap_server AttributeError 这次必须正面处理 —— 因为 standard PD-disagg 部署典型用 producer/consumer 角色,不是 kv_both
  • D-pool occupancy timeline 是否要补到 MB3 的实验范围里?我倾向 yesPhase 1 自然就 cover
  • Cost modeltask #12是否还需要做取决于 Phase 1 结论是否需要 analytical 外推