Gahow Wang
ef9e0102ec
Re-runs the elastic_migration_v2 trace (w600 r0.0015 st30, 1214 reqs,
274 sessions, 8×TP1 vLLM + cache_aware_proxy) with three configs:
- plain unified
- unified + Mooncake kv_both
- unified + Mooncake kv_both + DR-fix (env-gated O(|cache|) hash sync removal)
TTFT p90: 11.97 s → 9.74 s (−18.6%) → 7.58 s (−36.6% vs plain)
E2E p90: 23.48 s → 21.25 s (−9.5%) → 17.93 s (−23.6% vs plain)
Two findings:
1. The "+45% kv_both penalty" claim from elastic_migration_v2 is OBSOLETE
on current codebase — kv_both is now *faster* than plain at p90.
Likely fixed by e3a1d70 (RDMA-READ → bootstrap PUSH refactor) and
the connector-mode delay_free_blocks extending cross-turn prefix
cache hits on a 93%-intra-session-reuse trace.
2. DR-fix removes another 22% from TTFT p90 by skipping the
O(|cache|) hash sync in build_connector_meta. Cache-sweep with
DR-fix shows slope drops from +94.5 to +2.3 μs/1k blocks.
Adds:
- run_trace_replay_drfix.sh: A/B/C harness (env CT_DR_FIX gates patch)
- analyze_trace_replay.py: TTFT/TPOT/E2E delta analysis
- REPORT_TRACE_REPLAY.md: summary + reproduction
- results/20260526_1627_drfix/: cache-sweep with DR-fix
- results/trace_replay_20260526_1652/: full trace-replay A/B/C
Implication for EAR paper: the kv_both substrate is no longer the
bottleneck blocking session migration. The prior 4 migration reverts
were dominated by transfer overhead that has now been characterized
and (partially) removed.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
116 lines
4.1 KiB
Python
Executable File
116 lines
4.1 KiB
Python
Executable File
#!/usr/bin/env python3
|
|
"""Compute TTFT/TPOT/E2E mean/p50/p90/p99 from trace-replay metrics.jsonl.
|
|
|
|
Usage:
|
|
python analyze_trace_replay.py --root <outroot>
|
|
where <outroot>/{unified,unified_kv_both,unified_kv_both_drfix}/metrics.jsonl
|
|
each contain one JSONL object per request (with ttft_s, tpot_s, latency_s).
|
|
"""
|
|
|
|
import argparse
|
|
import json
|
|
import statistics
|
|
from pathlib import Path
|
|
|
|
|
|
def pct(xs, p):
|
|
if not xs:
|
|
return None
|
|
xs = sorted(xs)
|
|
k = max(0, min(len(xs) - 1, int(p / 100.0 * (len(xs) - 1))))
|
|
return xs[k]
|
|
|
|
|
|
def summarise(rows):
|
|
ok = [r for r in rows if not r.get("error")]
|
|
ttft = [r["ttft_s"] * 1000 for r in ok if r.get("ttft_s") is not None]
|
|
tpot = [r["tpot_s"] * 1000 for r in ok if r.get("tpot_s")]
|
|
e2e = [r["latency_s"] * 1000 for r in ok if r.get("latency_s") is not None]
|
|
return {
|
|
"n_total": len(rows),
|
|
"n_ok": len(ok),
|
|
"n_err": len(rows) - len(ok),
|
|
"ttft_mean_ms": statistics.mean(ttft) if ttft else None,
|
|
"ttft_p50_ms": pct(ttft, 50),
|
|
"ttft_p90_ms": pct(ttft, 90),
|
|
"ttft_p99_ms": pct(ttft, 99),
|
|
"tpot_mean_ms": statistics.mean(tpot) if tpot else None,
|
|
"tpot_p50_ms": pct(tpot, 50),
|
|
"tpot_p90_ms": pct(tpot, 90),
|
|
"tpot_p99_ms": pct(tpot, 99),
|
|
"e2e_mean_ms": statistics.mean(e2e) if e2e else None,
|
|
"e2e_p50_ms": pct(e2e, 50),
|
|
"e2e_p90_ms": pct(e2e, 90),
|
|
"e2e_p99_ms": pct(e2e, 99),
|
|
}
|
|
|
|
|
|
def main():
|
|
ap = argparse.ArgumentParser()
|
|
ap.add_argument("--root", type=Path, required=True)
|
|
ap.add_argument("--configs", nargs="+",
|
|
default=["unified", "unified_kv_both", "unified_kv_both_drfix"])
|
|
args = ap.parse_args()
|
|
|
|
results = {}
|
|
for cfg in args.configs:
|
|
p = args.root / cfg / "metrics.jsonl"
|
|
if not p.exists():
|
|
print(f"MISSING: {p}")
|
|
continue
|
|
rows = [json.loads(l) for l in open(p) if l.strip()]
|
|
results[cfg] = summarise(rows)
|
|
|
|
def fmt(v):
|
|
return f"{v:.0f}" if v is not None else "-"
|
|
|
|
def pctd(a, b):
|
|
if a is None or b is None or a == 0:
|
|
return "-"
|
|
return f"{(b/a-1)*100:+.1f}%"
|
|
|
|
print(f"{'metric':<14}", end="")
|
|
for cfg in args.configs:
|
|
print(f"{cfg[:20]:>22}", end="")
|
|
print()
|
|
print("-" * (14 + 22 * len(args.configs)))
|
|
|
|
for m in ["n_ok", "ttft_mean_ms", "ttft_p50_ms", "ttft_p90_ms", "ttft_p99_ms",
|
|
"tpot_mean_ms", "tpot_p50_ms", "tpot_p90_ms", "tpot_p99_ms",
|
|
"e2e_mean_ms", "e2e_p50_ms", "e2e_p90_ms", "e2e_p99_ms"]:
|
|
print(f"{m:<14}", end="")
|
|
for cfg in args.configs:
|
|
if cfg not in results:
|
|
print(f"{'-':>22}", end="")
|
|
else:
|
|
print(f"{fmt(results[cfg][m]):>22}", end="")
|
|
print()
|
|
|
|
# Tax tables
|
|
if len(args.configs) >= 2 and all(c in results for c in args.configs):
|
|
plain = args.configs[0]
|
|
for ref_cfg in args.configs[1:]:
|
|
print(f"\n=== {ref_cfg} vs {plain} ===")
|
|
for m in ["ttft_p50_ms", "ttft_p90_ms", "ttft_p99_ms",
|
|
"tpot_p50_ms", "tpot_p90_ms", "tpot_p99_ms",
|
|
"e2e_p50_ms", "e2e_p90_ms", "e2e_p99_ms"]:
|
|
a = results[plain][m]; b = results[ref_cfg][m]
|
|
print(f" {m:<14} {fmt(a):>10} → {fmt(b):>10} ({pctd(a, b)})")
|
|
if "unified_kv_both" in results and "unified_kv_both_drfix" in results:
|
|
print(f"\n=== DR-fix improvement: unified_kv_both_drfix vs unified_kv_both ===")
|
|
for m in ["ttft_p50_ms", "ttft_p90_ms", "ttft_p99_ms",
|
|
"tpot_p50_ms", "tpot_p90_ms", "tpot_p99_ms",
|
|
"e2e_p50_ms", "e2e_p90_ms", "e2e_p99_ms"]:
|
|
a = results["unified_kv_both"][m]
|
|
b = results["unified_kv_both_drfix"][m]
|
|
print(f" {m:<14} {fmt(a):>10} → {fmt(b):>10} ({pctd(a, b)})")
|
|
|
|
# Save machine-readable
|
|
with open(args.root / "trace_replay_summary.json", "w") as f:
|
|
json.dump(results, f, indent=2)
|
|
print(f"\nWrote {args.root}/trace_replay_summary.json")
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|