Fair A/B: Elastic P2P wins on ALL metrics vs baseline (fresh restart)

Same-condition comparison (both fresh restart, same trace, same params):
  Baseline (combined):  TTFT=2.383/27.622  TPOT90=0.117  E2E=10.232
  Elastic P2P (cap=4):  TTFT=1.315/13.179  TPOT90=0.075  E2E=5.708
  Delta:                -45%  / -52%        -36%          -44%

Key finding: TPOT p90 dropped 36% — confirming heavy prefill DOES
disrupt decode in combined mode, and elastic offload effectively
isolates it. Previous comparisons missed this because baselines
were run under different conditions (stale instances, different time_scale).

GPU util: elastic uses less GPU (15.8% vs 28.7%) but achieves better
latency — higher efficiency through better cache distribution.

APC: elastic has more balanced per-instance APC (36-38% prefix + 30-35%
external) vs baseline's skewed distribution (3.8% - 68.3%).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
2026-05-22 15:48:51 +08:00
parent 76ee28a40f
commit 1e8628581b
2 changed files with 219 additions and 0 deletions

118
scripts/compare_ab_final.py Normal file
View File

@@ -0,0 +1,118 @@
"""Final A/B comparison: baseline (dash0) vs elastic (dash1).
Both fresh restart, same trace, same params. GPU util + APC + latency."""
import json, csv, statistics, os, urllib.request
def lat(path):
rows = [json.loads(l) for l in open(path)]
ok = [r for r in rows if not r.get("error")]
err = [r for r in rows if r.get("error")]
ttfts = sorted([r["ttft_s"] for r in ok if r.get("ttft_s")])
tpots = sorted([r["tpot_s"] for r in ok if r.get("tpot_s") and r["tpot_s"]>0])
lats = sorted([r["latency_s"] for r in ok])
p = lambda v,q: v[min(int(q*len(v)),len(v)-1)] if v else 0
ok_inp = sorted([r["input_length"] for r in ok])
err_inp = sorted([r["input_length"] for r in err])
return {"ok": len(ok), "n": len(rows),
"t50": p(ttfts,.5), "t90": p(ttfts,.9),
"p50": p(tpots,.5), "p90": p(tpots,.9),
"e50": p(lats,.5), "e90": p(lats,.9),
"inp50": p(ok_inp,.5), "err_inp50": p(err_inp,.5) if err_inp else 0}
def gpu_per_inst(path):
if not os.path.exists(path):
return {}
rows = list(csv.DictReader(open(path)))
by_gpu = {}
for r in rows:
g = int(r["gpu"])
by_gpu.setdefault(g, []).append(float(r["util_pct"]))
result = {}
for g, vals in sorted(by_gpu.items()):
nz = sum(1 for v in vals if v > 0)
result[g] = {"mean": statistics.fmean(vals), "active": nz*100//len(vals)}
return result
def get_apc(host, port_start=8000, n=8):
"""Get APC from vLLM log files."""
results = {}
for i in range(n):
for log_prefix in ["/tmp/ab_base_", "/tmp/ab_elastic_"]:
logfile = "%s%d.log" % (log_prefix, i)
try:
import subprocess
r = subprocess.run(["ssh", "-o", "ConnectTimeout=5", host,
"grep 'Prefix cache hit rate' %s 2>/dev/null | tail -1" % logfile],
capture_output=True, text=True, timeout=10)
line = r.stdout.strip()
if "Prefix cache hit rate:" in line:
import re
pch = re.search(r"Prefix cache hit rate: ([0-9.]+)", line)
ech = re.search(r"External prefix cache hit rate: ([0-9.]+)", line)
results[i] = {
"prefix": float(pch.group(1)) if pch else 0,
"external": float(ech.group(1)) if ech else 0,
}
except:
pass
return results
sep = "=" * 80
print(sep)
print(" A/B COMPARISON: Baseline (dash0) vs Elastic P2P (dash1)")
print(" Both: fresh restart, 200 req, time_scale=20, 8 sessions")
print(sep)
# Latency
print("\n LATENCY:")
fmt = "%-30s %7s %8s %8s %8s %8s %8s %8s"
print(fmt % ("Config", "OK/N", "TTFT50", "TTFT90", "TPOT50", "TPOT90", "E2E50", "inp_p50"))
print("-" * 80)
for path, label in [
("outputs/ab_baseline/metrics.jsonl", "Baseline (combined)"),
("outputs/ab_elastic/metrics.jsonl", "Elastic P2P (cap=4)"),
]:
if os.path.exists(path):
s = lat(path)
print(fmt % (label, "%d/%d" % (s["ok"],s["n"]),
"%.3f" % s["t50"], "%.3f" % s["t90"], "%.3f" % s["p50"],
"%.3f" % s["p90"], "%.3f" % s["e50"], str(s["inp50"])))
# Delta
b = lat("outputs/ab_baseline/metrics.jsonl") if os.path.exists("outputs/ab_baseline/metrics.jsonl") else None
a = lat("outputs/ab_elastic/metrics.jsonl") if os.path.exists("outputs/ab_elastic/metrics.jsonl") else None
if b and a:
print()
for label, bv, av in [("TTFT p50",b["t50"],a["t50"]),("TTFT p90",b["t90"],a["t90"]),
("TPOT p90",b["p90"],a["p90"]),("E2E p50",b["e50"],a["e50"])]:
d = (av/bv-1)*100 if bv > 0 else 0
print(" %s: %.3f -> %.3f (%+.1f%%)" % (label, bv, av, d))
# GPU utilization
print("\n GPU UTILIZATION:")
for path, label in [
("outputs/ab_baseline/gpu_util.csv", "Baseline"),
("outputs/ab_elastic/gpu_util.csv", "Elastic"),
]:
gi = gpu_per_inst(path)
if gi:
means = [gi[g]["mean"] for g in sorted(gi.keys())]
actives = [gi[g]["active"] for g in sorted(gi.keys())]
print(" %s:" % label)
for g in sorted(gi.keys()):
print(" GPU%d: mean=%5.1f%% active=%2d%%" % (g, gi[g]["mean"], gi[g]["active"]))
print(" Aggregate: mean=%.1f%% imbalance=%.1fx" % (
statistics.fmean(means), max(means)/max(min(means),0.1)))
# APC from vLLM logs
print("\n PREFIX CACHE HIT RATE (from vLLM logs):")
for host, label, prefix in [("dash0", "Baseline", "/tmp/ab_base_"), ("dash1", "Elastic", "/tmp/ab_elastic_")]:
apc = get_apc(host)
if apc:
prefixes = [v["prefix"] for v in apc.values()]
externals = [v.get("external", 0) for v in apc.values()]
print(" %s:" % label)
for i in sorted(apc.keys()):
ext = " ext=%.1f%%" % apc[i]["external"] if apc[i].get("external") else ""
print(" inst_%d: prefix=%.1f%%%s" % (i, apc[i]["prefix"], ext))
print(" Avg prefix: %.1f%% Avg external: %.1f%%" % (
statistics.fmean(prefixes), statistics.fmean(externals)))

View File

@@ -0,0 +1,101 @@
"""Plot per-GPU utilization timeline for elastic vs baseline."""
import csv, json, sys, os
def load_gpu(path):
"""Load GPU util CSV, return {gpu_id: [(timestamp, util%)]]}."""
by_gpu = {}
with open(path) as f:
for r in csv.DictReader(f):
g = int(r["gpu"])
t = float(r["timestamp"])
u = float(r["util_pct"])
by_gpu.setdefault(g, []).append((t, u))
# Normalize timestamps to start at 0
if by_gpu:
t0 = min(pts[0][0] for pts in by_gpu.values())
for g in by_gpu:
by_gpu[g] = [(t - t0, u) for t, u in by_gpu[g]]
return by_gpu
def print_timeline(by_gpu, label, max_time=None):
"""Print ASCII timeline of GPU utilization."""
print(f"\n{'='*70}")
print(f" {label}")
print(f"{'='*70}")
if not by_gpu:
print(" No data")
return
# Bucket into 10s windows
window = 10.0
if max_time is None:
max_time = max(t for pts in by_gpu.values() for t, _ in pts)
n_windows = min(int(max_time / window) + 1, 40) # cap at 40 columns
for gpu in sorted(by_gpu.keys()):
pts = by_gpu[gpu]
buckets = [[] for _ in range(n_windows)]
for t, u in pts:
b = min(int(t / window), n_windows - 1)
buckets[b].append(u)
avgs = [sum(b)/len(b) if b else 0 for b in buckets]
# ASCII bar: . = 0-10%, o = 10-30%, O = 30-60%, # = 60-100%
bar = ""
for a in avgs:
if a < 1: bar += " "
elif a < 10: bar += "."
elif a < 30: bar += "o"
elif a < 60: bar += "O"
else: bar += "#"
mean = sum(a for a in avgs) / len(avgs) if avgs else 0
print(f" GPU{gpu}: |{bar}| mean={mean:.0f}%")
print(f" Time: {'0':>1}{'':>{n_windows-6}}{int(max_time)}s")
print(f" Legend: ' '=0% .=1-10% o=10-30% O=30-60% #=60-100%")
# Per-GPU stats
print(f"\n Per-GPU mean utilization:")
for gpu in sorted(by_gpu.keys()):
pts = by_gpu[gpu]
vals = [u for _, u in pts]
mean = sum(vals) / len(vals)
nz = sum(1 for v in vals if v > 0)
print(f" GPU{gpu}: mean={mean:.1f}% active={nz*100//len(vals)}% samples={len(vals)}")
# Load and compare
configs = [
("outputs/baseline_dash1/gpu_util.csv", "Baseline (8 combined, dash1)"),
("outputs/elastic_v4/gpu_util.csv", "Elastic P2P v4 (dash0)"),
]
for path, label in configs:
if os.path.exists(path):
by_gpu = load_gpu(path)
print_timeline(by_gpu, label)
else:
print(f"\n {label}: {path} NOT FOUND")
# Imbalance metric
print(f"\n{'='*70}")
print(f" LOAD IMBALANCE ANALYSIS")
print(f"{'='*70}")
for path, label in configs:
if not os.path.exists(path):
continue
by_gpu = load_gpu(path)
means = []
for gpu in sorted(by_gpu.keys()):
vals = [u for _, u in by_gpu[gpu]]
means.append(sum(vals) / len(vals))
if means:
avg = sum(means) / len(means)
max_m = max(means)
min_m = min(means)
imbalance = max_m / max(min_m, 0.1)
print(f" {label}:")
print(f" Per-GPU means: {['%.1f' % m for m in means]}")
print(f" Avg={avg:.1f}% Min={min_m:.1f}% Max={max_m:.1f}% Imbalance={imbalance:.1f}x")