Add per-request breakdown profiling, identify KV cache memory bottleneck

Breakdown profiling at proxy level captures:
  t_proxy_recv → t_prefill_sent → t_prefill_done → t_decode_sent → t_first_token

Key finding: 87.7% of TTFT is spent in kv+decode phase, NOT prefill.
Root cause: decode instance KV cache memory saturation (97.1% usage).

With 6P+2D config, 2 decode GPUs have only ~56GB total KV cache.
Large agentic requests (avg 33.6k tokens) fill this quickly.
Small requests (49 tokens, prefill=0.044s) wait 114s for KV cache
to be freed by large requests completing decode.

vLLM log confirms: Running=0, Waiting=6, KV cache=97.1%
GPU is idle but requests queue for KV cache memory, not compute.

This is the fundamental bottleneck of single-machine PD separation
for long-context agentic workloads: concentrating decode onto fewer
GPUs creates a KV cache memory wall.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
2026-05-22 00:13:50 +08:00
parent c7afdc5074
commit ce616f46d1
3 changed files with 182 additions and 7 deletions

View File

@@ -0,0 +1,54 @@
"""Analyze per-request breakdown data from the proxy."""
import json, statistics, sys
url = sys.argv[1] if len(sys.argv) > 1 else "http://localhost:9090/breakdown"
if url.startswith("http"):
import urllib.request
data = json.loads(urllib.request.urlopen(url, timeout=10).read())
else:
data = json.load(open(url))
print("Total records: %d" % len(data))
results = []
for d in data:
keys = ["t_proxy_recv", "t_prefill_sent", "t_prefill_done", "t_decode_sent", "t_first_token"]
if not all(k in d for k in keys):
continue
results.append({
"input": d["input_length"],
"prefill": d["t_prefill_done"] - d["t_prefill_sent"],
"proxy_gap": d["t_decode_sent"] - d["t_prefill_done"],
"kv_decode": d["t_first_token"] - d["t_decode_sent"],
"ttft": d["t_first_token"] - d["t_proxy_recv"],
})
results.sort(key=lambda x: x["input"])
print("Complete breakdown: %d" % len(results))
if not results:
print("No complete records yet")
sys.exit(0)
print()
print(" %8s %9s %9s %9s %9s" % ("input", "prefill", "proxy", "kv+dec", "TTFT"))
print(" %8s %9s %9s %9s %9s" % ("-----", "-------", "-----", "------", "----"))
for r in results[:25]:
print(" %8d %9.3f %9.3f %9.3f %9.3f" % (
r["input"], r["prefill"], r["proxy_gap"], r["kv_decode"], r["ttft"]))
print()
for key in ["prefill", "proxy_gap", "kv_decode", "ttft"]:
vals = sorted([r[key] for r in results])
p = lambda q: vals[min(int(q * len(vals)), len(vals) - 1)]
print(" %s: p50=%.3fs p90=%.3fs mean=%.3fs" % (
key, p(.5), p(.9), statistics.fmean(vals)))
# Fraction of TTFT by stage
print()
print(" TTFT breakdown (fraction of total):")
for key in ["prefill", "proxy_gap", "kv_decode"]:
fracs = [r[key] / r["ttft"] * 100 for r in results if r["ttft"] > 0.01]
if fracs:
print(" %s: mean=%.1f%% of TTFT" % (key, statistics.fmean(fracs)))

View File

@@ -196,25 +196,42 @@ async def _handle_combined(api, req_data, token_ids, input_length, session_id, h
return StreamingResponse(generate(), media_type="text/event-stream")
async def _send_prefill_async(p_inst, api, prefill_data, p_headers, token_ids, input_length):
import time as _time
# Per-request breakdown log (append-only)
_breakdown_log: list[dict] = []
async def _send_prefill_async(p_inst, api, prefill_data, p_headers, token_ids,
input_length, breakdown):
"""Fire-and-forget prefill: send and don't block caller."""
try:
resp = await p_inst.client.post(api, json=prefill_data, headers=p_headers)
breakdown["t_prefill_done"] = _time.monotonic()
resp.raise_for_status()
await resp.aclose()
p_inst.record_prefix(token_ids)
except Exception:
pass
breakdown["t_prefill_done"] = _time.monotonic()
breakdown["prefill_error"] = True
finally:
p_inst.ongoing_tokens -= input_length
async def _handle_pd_sep(api, req_data, request_id, token_ids, input_length,
session_id, headers):
"""PD-Sep mode. --fire-and-forget controls prefill waiting behavior."""
"""PD-Sep mode with per-stage breakdown profiling."""
breakdown = {
"request_id": request_id,
"input_length": input_length,
"t_proxy_recv": _time.monotonic(),
}
p_inst, _ = pick_instance(prefill_instances, token_ids, session_id,
input_length, session_affinity)
d_inst = min(decode_instances, key=lambda x: x.ongoing_tokens)
breakdown["p_inst"] = p_inst.url
breakdown["d_inst"] = d_inst.url
prefill_data = req_data.copy()
prefill_data["kv_transfer_params"] = {
@@ -228,24 +245,27 @@ async def _handle_pd_sep(api, req_data, request_id, token_ids, input_length,
p_headers = {**headers, "X-data-parallel-rank": "0"}
p_inst.ongoing_tokens += input_length
breakdown["t_prefill_sent"] = _time.monotonic()
if global_args.fire_and_forget:
# Fire-and-forget: send prefill async, immediately proceed to decode
asyncio.create_task(_send_prefill_async(
p_inst, api, prefill_data, p_headers, token_ids, input_length))
p_inst, api, prefill_data, p_headers, token_ids, input_length, breakdown))
else:
# Await: block until prefill completes, then send decode
try:
resp = await p_inst.client.post(api, json=prefill_data, headers=p_headers)
breakdown["t_prefill_done"] = _time.monotonic()
resp.raise_for_status()
await resp.aclose()
p_inst.record_prefix(token_ids)
except Exception as e:
breakdown["t_prefill_done"] = _time.monotonic()
breakdown["prefill_error"] = True
_breakdown_log.append(breakdown)
raise HTTPException(status_code=502, detail=f"Prefill failed: {e}")
finally:
p_inst.ongoing_tokens -= input_length
# Stream decode
# Send decode
d_inst.ongoing_tokens += input_length
parsed = urllib.parse.urlparse(str(p_inst.client.base_url))
bootstrap_addr = f"http://{parsed.hostname}:{p_inst.bootstrap_port}"
@@ -258,18 +278,32 @@ async def _handle_pd_sep(api, req_data, request_id, token_ids, input_length,
"transfer_id": f"xfer-{request_id}",
}
breakdown["t_decode_sent"] = _time.monotonic()
async def generate():
first_token = True
try:
async with d_inst.client.stream("POST", api, json=decode_data, headers=headers) as resp:
resp.raise_for_status()
async for chunk in resp.aiter_bytes():
if first_token:
breakdown["t_first_token"] = _time.monotonic()
first_token = False
yield chunk
finally:
breakdown["t_done"] = _time.monotonic()
d_inst.ongoing_tokens -= input_length
_breakdown_log.append(breakdown)
return StreamingResponse(generate(), media_type="application/json")
@app.get("/breakdown")
async def get_breakdown():
"""Return per-request breakdown data for analysis."""
return _breakdown_log
def parse_args():
p = argparse.ArgumentParser(description="Unified cache-aware global scheduler")
p.add_argument("--port", type=int, default=8000)

87
scripts/profile_fnf.py Normal file
View File

@@ -0,0 +1,87 @@
"""Deep profile: why fire-and-forget TTFT is 5x worse than await."""
import json, statistics
await_rows = [json.loads(l) for l in open("outputs/gpu_ab_6p2d/metrics.jsonl")]
fnf_rows = [json.loads(l) for l in open("outputs/gpu_ab_6p2d_fnf/metrics.jsonl")]
await_ok = [r for r in await_rows if not r.get("error")]
fnf_ok = [r for r in fnf_rows if not r.get("error")]
# Match by request_id
await_by_id = {r["request_id"]: r for r in await_ok}
fnf_by_id = {r["request_id"]: r for r in fnf_ok}
common = set(await_by_id.keys()) & set(fnf_by_id.keys())
print("=" * 75)
print(" PROFILE: Fire-and-Forget vs Await-Prefill (same 6P+2D instances)")
print("=" * 75)
print(f" Common requests: {len(common)}")
# Per-request comparison
diffs = []
for rid in common:
a = await_by_id[rid]
f = fnf_by_id[rid]
if a.get("ttft_s") and f.get("ttft_s") and a["ttft_s"] > 0:
diffs.append({
"id": rid, "input": a["input_length"],
"a_ttft": a["ttft_s"], "f_ttft": f["ttft_s"],
"ratio": f["ttft_s"] / a["ttft_s"],
"a_e2e": a["latency_s"], "f_e2e": f["latency_s"],
"a_tpot": a.get("tpot_s", 0), "f_tpot": f.get("tpot_s", 0),
"a_out": a.get("actual_output_tokens", 0) or 0,
"f_out": f.get("actual_output_tokens", 0) or 0,
})
diffs.sort(key=lambda x: x["input"])
print("\n Per-request (sorted by input_length):")
hdr = "%8s %10s %10s %7s %10s %10s %8s %8s" % (
"input", "await_TTFT", "fnf_TTFT", "ratio", "await_E2E", "fnf_E2E", "a_TPOT", "f_TPOT")
print(" " + hdr)
print(" " + "-" * len(hdr))
for d in diffs[:25]:
print(" %8d %10.3f %10.3f %6.1fx %10.3f %10.3f %8.4f %8.4f" % (
d["input"], d["a_ttft"], d["f_ttft"], d["ratio"],
d["a_e2e"], d["f_e2e"], d["a_tpot"], d["f_tpot"]))
# Statistics
if diffs:
ratios = [d["ratio"] for d in diffs]
ratios.sort()
p = lambda v, q: v[min(int(q*len(v)), len(v)-1)]
print("\n TTFT ratio (FnF / Await):")
print(" p10=%.2fx p50=%.2fx p90=%.2fx mean=%.2fx" % (
p(ratios,.1), p(ratios,.5), p(ratios,.9), statistics.fmean(ratios)))
faster = sum(1 for r in ratios if r < 1.0)
print(" FnF faster: %d/%d (%.0f%%)" % (faster, len(ratios), faster*100/len(ratios)))
# Bucket by input size
print("\n TTFT ratio by input size bucket:")
buckets = [(0, 5000, "<5k"), (5000, 20000, "5-20k"), (20000, 50000, "20-50k"), (50000, 999999, ">50k")]
for lo, hi, label in buckets:
subset = [d for d in diffs if lo <= d["input"] < hi]
if subset:
rs = [d["ratio"] for d in subset]
a_ttfts = [d["a_ttft"] for d in subset]
f_ttfts = [d["f_ttft"] for d in subset]
print(" %6s: n=%3d await_TTFT=%.3f fnf_TTFT=%.3f ratio=%.2fx" % (
label, len(subset), statistics.fmean(a_ttfts), statistics.fmean(f_ttfts),
statistics.fmean(rs)))
# TPOT comparison
a_tpots = [d["a_tpot"] for d in diffs if d["a_tpot"] > 0]
f_tpots = [d["f_tpot"] for d in diffs if d["f_tpot"] > 0]
if a_tpots and f_tpots:
print("\n TPOT comparison:")
print(" Await: mean=%.4f p50=%.4f" % (statistics.fmean(a_tpots), sorted(a_tpots)[len(a_tpots)//2]))
print(" FnF: mean=%.4f p50=%.4f" % (statistics.fmean(f_tpots), sorted(f_tpots)[len(f_tpots)//2]))
# Also look at non-common requests (FnF only failures)
fnf_err = [r for r in fnf_rows if r.get("error")]
await_err_ids = {r["request_id"] for r in await_rows if r.get("error")}
fnf_only_err = [r for r in fnf_err if r["request_id"] not in await_err_ids]
print("\n Errors unique to FnF: %d" % len(fnf_only_err))
for r in fnf_only_err[:5]:
print(" input=%d err=%s" % (r["input_length"], r["error"][:60]))