Migration transfer-cost study: KV transfer is slow on busy GPUs

MIGRATION_TRANSFER_COST.md: under real load, migration KV transfer runs at
~3 GB/s vs ~10 GB/s idle. Decomposed (instruments + MB6 microbench) into
~55% RDMA-actual (HBM/PCIe contention with running kernels: 7.6->4.0 GB/s)
+ ~45% control-plane GIL starvation during long prefills. Reproduced on a
fresh upstream venv (byte-identical transfer path) -> upstream/hardware
inherent, not our patch. Layerwise is the wrong lever; the tax is structural
on a loaded agentic cluster. Includes mb6_transfer_under_load + run_mb6,
instrument_dst_migration/mooncake, and the dst/transfer decomposition analyzers.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
2026-05-29 11:53:01 +08:00
parent 67fcec7933
commit 1262c9c22e
11 changed files with 2055 additions and 0 deletions

View File

@@ -0,0 +1,178 @@
# Why KV-transfer is slow during migration under real load
**Question.** EAR's unified+A+B routing beats migration (v3) on agentic
workloads. We wanted to know whether *layerwise* KV transfer would shrink
migration's overhead enough to make it viable. Investigating that led to a
sharper question: **in a real (loaded) cluster, when we migrate, the KV
transfer is already slow — the effective bandwidth is far below the
~10 GB/s wire rate. Why?**
This doc answers that with instrumented measurements.
**TL;DR.** Migration fires precisely when instances are *busy* (that's the
trigger). But on a busy instance, KV transfer runs at **~3 GB/s instead of
~10 GB/s**, because:
1. **The RDMA write itself slows ~2× under compute load** — GPU-direct RDMA
(`batch_transfer_sync_write`) contends with the running attention/MLP
kernels for **HBM and PCIe bandwidth**. (idle 7.6 GB/s → busy 4.0 GB/s)
2. **The connector's Python control plane gets GIL-starved** — mooncake's
ZMQ handshake + transfer orchestration run on asyncio threads inside the
engine process; when the engine's main thread is doing a long forward
pass (e.g. a 100k-token prefill), those threads stall for *seconds*.
Both are **inherent to upstream vLLM 0.18.1 + mooncake** (reproduced on a
clean fresh venv; the transfer path is byte-identical to upstream — our
patches did not cause this), and both get **worse**, not better, with
layerwise transfer. So the bandwidth gap is not a layerwise problem; it is a
*transfer-on-a-busy-GPU* problem.
---
## 1. Evidence chain
Three independent measurements, all on dash0 (8×H100, Qwen3-Coder-30B-A3B,
TP=1), Mooncake `kv_both`.
### 1a. Instrumented v3 trace replay — where does migration time go?
Run `outputs/b3_v3_fullbreak_20260528_0338/`. Instruments:
`instrument_dst_migration.py` (dst scheduler lifecycle) +
`instrument_mooncake.py` (connector internals: `send_blocks` RDMA,
`receive_kv` window, `ready_wait`).
25 migrations fired over the trace. Dst-side migration overhead
(`T_kv_pull` = scheduler marks `WAITING_FOR_REMOTE_KVS``finished_recving`):
| component | share | what it is |
|---|---:|---|
| RDMA-actual (`batch_transfer_sync_write`) | **55%** (55.2 s) | the real RDMA write |
| dst control-plane gap | **45%** (45.4 s) | scheduler↔receiver_loop dispatch + completion propagation |
| `ready_wait` (src KV not committed) | 0% | 25/25 already committed — **ruled out** |
- Pure RDMA aggregate rate: **2.03 GB/s** (vs MB2 idle 9.7 GB/s).
- RDMA rate **collapses with transfer size**: <3 GiB 49.5 GB/s,
>5 GiB → 0.92.6 GB/s.
- The control-plane gap is **bimodal**: median 0.04 s, but a handful of
requests stall ~10 s. Those are small-KV transfers (0.18 s of actual RDMA)
whose `T_kv_pull` is 811 s — i.e. the dst's `receiver_loop` thread was
GIL-starved for ~10 s while the engine did a big forward pass.
> Earlier (pre-instrumentation) we wrongly attributed ~90% of migration
> overhead to "dst scheduler queueing" by estimating transfer at clean wire
> speed. With real instrumentation, dst *scheduler admission* is ~0
> (`T_admission_post_kv` = 0.003 s); the time is the transfer phase (RDMA +
> connector control plane), both degraded by instance busy-ness.
### 1b. MB6 controlled microbench — does busy-ness cause it?
`microbench/fresh_setup/mb6_transfer_under_load.py` + `run_mb6.sh`: 2
instances, transfer a fixed-size KV (prefill on A → migrate to B) while
holding *N* background decode streams on both. Sweep N.
Effective transfer bandwidth (65k-token KV ≈ 6 GiB), main venv:
| background load | 65k transfer | eff bandwidth |
|---|---:|---:|
| **0 (idle)** | 747 ms | **8.76 GB/s** |
| 8 (4/instance) | 2423 ms | 4.53 GB/s |
| **24 (12/instance)** | 2015 ms | **3.33 GB/s** |
Monotonic degradation with load. **The busy level (3.3 GB/s) matches the
v3 trace's 3.3 GB/s median exactly** — because agentic instances run
~10+ concurrent requests, i.e. the bg=24 regime.
Decomposing the 65k transfer into RDMA-actual vs control-plane:
| bg | RDMA rate | control-plane share |
|---|---:|---:|
| 0 (idle) | 7.56 GB/s | 13% |
| 8 | 4.07 GB/s | 11% |
| 24 (busy) | 3.97 GB/s | 15% |
In the clean microbench the **RDMA write itself is the dominant degrading
term** (7.6 → 4.0 GB/s). The ~10 s control-plane stalls seen in the trace
(1a) don't reproduce here because steady decode forward passes are short;
they require the long (100k-token) prefills that the real trace has.
### 1c. Fresh-venv comparison — is it our patch?
Same MB6 sweep on `agentic-kv-fresh/.venv` (clean upstream-style 0.18.1):
| bg | 65k eff (fresh) | 65k eff (main/patched) |
|---|---:|---:|
| 0 | 8.73 GB/s | 8.76 GB/s |
| 8 | 4.52 GB/s | 4.53 GB/s |
| 24 | 3.27 GB/s | 3.33 GB/s |
**Identical within noise.** Plus a static check: the v3 transfer path
(`send_kv_to_decode`, `_send_blocks`/`batch_transfer_sync_write`,
`_build_transfer_params`) is **byte-identical** to pristine upstream 0.18.1
(commit `445e491`); `receive_kv_from_single_worker` differs only by a 4-line
error branch. Our mooncake commits (`a7df84b` direct-read,
`ea51497` partial-prefill, `e3a1d70` read→push) only touch a *separate*
`direct_read` path that v3 does **not** use (v3 requests carry no
`direct_read` flag → normal push path).
**The slowdown is upstream/hardware-inherent, not introduced by us.**
---
## 2. Root cause
Migration in agentic serving transfers KV **between instances that are
concurrently busy with compute** — by construction, since v3 migrates *away
from* a busy host. On a busy instance:
- **HBM/PCIe contention (the dominant, irreducible part).** Mooncake's
transfer is GPU-direct RDMA: the NIC DMAs KV straight out of / into GPU
HBM. While the GPU runs attention+MLP kernels, those kernels saturate HBM
bandwidth, so the NIC's RDMA gets a smaller slice. Effective transfer
bandwidth roughly halves (7.6 → 4.0 GB/s at our load), and degrades
further for large multi-segment transfers.
- **Control-plane GIL starvation (secondary, bursty).** The connector runs
its ZMQ handshake + `send_kv_to_decode`/`receive_kv` orchestration on
asyncio threads (`sender_loop`/`receiver_loop`) *inside the engine
process*. A long forward pass (100k-token prefill) holds the GIL for
seconds, stalling those threads → multi-second dispatch gaps even when the
actual transfer is 0.2 s.
MB2 measured 9.7 GB/s precisely because both endpoints were **idle**. The
real-workload gap is the difference between "idle benchmark" and "transfer
while the GPU is doing the day job."
---
## 3. Implication: layerwise is the wrong lever; migration's tax is largely irreducible
| lever | effect on the gap |
|---|---|
| **Model-level layerwise transfer** (push each layer's KV during prefill) | **Worse.** Prefill is the most HBM-intensive phase, so per-layer transfers contend *harder* for HBM (Cause 1); and they multiply the control-plane round-trips (Cause 2). |
| **Control-plane fix** (move mooncake orchestration off the GIL-contended threads / separate process) | Addresses only the bursty ~10 s stalls (~15% in the clean case, up to ~45% of the trace tail). Does **not** touch the HBM-contention half. |
| **Reduce bytes** (cache-aware target so less KV moves) | Helps linearly; v3 Mechanism B already does some. Orthogonal. |
| **Migrate to/from idle instances** | Would restore ~10 GB/s — but defeats the purpose (we migrate *because* the host is busy). |
The dominant cost (RDMA contending with compute for HBM on busy instances)
is a **hardware reality**, not a software bug we can patch away, and not
something layerwise improves. This reinforces
[UNIFIED_ABLATION.md](UNIFIED_ABLATION.md): the unified no-migration path
(A+B'+RaceFix) remains the right default; migration's transfer tax is
structural on a loaded agentic cluster.
---
## 4. Repro / artifacts
- Instrumented v3 breakdown: `outputs/b3_v3_fullbreak_20260528_0338/unified_v3/`
(`transfer_decomp.txt`, `dst_migration_breakdown.{csv,png}`,
`transfer_rootcause.png`)
- MB6 main: `outputs/mb6_agentic-kv_20260528_0552/mb6_result.json`
- MB6 fresh: `outputs/mb6_fresh_20260528_0559/mb6_result.json`
- Instruments: `microbench/fresh_setup/instrument_dst_migration.py`,
`microbench/fresh_setup/instrument_mooncake.py`
- Microbench: `microbench/fresh_setup/mb6_transfer_under_load.py` +
`run_mb6.sh` (`VENV=… bash run_mb6.sh`)
- Analyzers: `analyze_dst_migration.py`, `analyze_transfer_decomp.py`
All instruments apply/revert cleanly via `--apply`/`--revert`; both venvs
were restored after the runs.