Add P/D ratio ablation: 6P+2D vs 4P+4D vs Combined

6P+2D gives more GPUs to prefill, fewer to decode:
- Decode util: 7.8% (4D) -> 19.0% (2D), less waste
- TTFT: 1.99s (4P) -> 1.48s (6P), -26% from less prefill queuing
- But Combined (30.5% util, TTFT 1.01s) still best overall

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

scripts/analyze_3way.py

@@ -0,0 +1,78 @@
+"""3-way GPU utilization + latency comparison."""
+import csv, json, statistics, os
+
+def gpu_stats(path, groups):
+    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 gname, indices in groups.items():
+        vals = []
+        for i in indices:
+            vals.extend(by_gpu.get(i, []))
+        if vals:
+            s = sorted(vals)
+            p = lambda q: s[min(int(q*len(s)), len(s)-1)]
+            nz = sum(1 for v in vals if v > 0)
+            result[gname] = {
+                "mean": statistics.fmean(vals), "p50": p(.5), "p90": p(.9),
+                "max": max(vals), "active_pct": nz*100//len(vals), "n": len(rows)//8
+            }
+    return result
+
+def latency_stats(path):
+    rows = [json.loads(l) for l in open(path)]
+    ok = [r for r in rows if not 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
+    return {
+        "ok": len(ok), "total": len(rows),
+        "ttft50": p(ttfts,.5), "ttft90": p(ttfts,.9),
+        "tpot50": p(tpots,.5), "tpot90": p(tpots,.9),
+        "e2e50": p(lats,.5), "e2e90": p(lats,.9),
+    }
+
+configs = [
+    ("gpu_ab_combined", "Combined 8colo", {"All(0-7)": list(range(8))}),
+    ("gpu_ab_pdsep",    "PD-Sep 4P+4D",  {"P(0-3)": [0,1,2,3], "D(4-7)": [4,5,6,7], "All": list(range(8))}),
+    ("gpu_ab_6p2d",     "PD-Sep 6P+2D",  {"P(0-5)": list(range(6)), "D(6-7)": [6,7], "All": list(range(8))}),
+]
+
+sep = "=" * 85
+print(sep)
+print("  3-WAY COMPARISON: Combined vs 4P+4D vs 6P+2D")
+print("  Same cache-aware scheduler, 200 req, time_scale=20")
+print(sep)
+
+print("\n  GPU UTILIZATION:")
+fmt = "  %-22s %7s %7s %7s %7s %7s"
+print(fmt % ("Config / Group", "Mean%", "P50%", "P90%", "Max%", "Active"))
+print("  " + "-" * 62)
+
+for dirname, label, groups in configs:
+    gpath = "outputs/%s/gpu_util.csv" % dirname
+    if not os.path.exists(gpath):
+        continue
+    gs = gpu_stats(gpath, groups)
+    for gname, s in gs.items():
+        tag = "%s %s" % (label, gname)
+        print(fmt % (tag[:22], "%.1f" % s["mean"], "%.0f" % s["p50"],
+                      "%.0f" % s["p90"], "%.0f" % s["max"], "%d%%" % s["active_pct"]))
+
+print("\n  LATENCY:")
+fmt2 = "  %-17s %6s %8s %8s %8s %8s %8s"
+print(fmt2 % ("Config", "OK/N", "TTFT50", "TTFT90", "TPOT50", "TPOT90", "E2E50"))
+print("  " + "-" * 65)
+
+for dirname, label, _ in configs:
+    mpath = "outputs/%s/metrics.jsonl" % dirname
+    if not os.path.exists(mpath):
+        continue
+    s = latency_stats(mpath)
+    okn = "%d/%d" % (s["ok"], s["total"])
+    print(fmt2 % (label[:17], okn, "%.3f" % s["ttft50"], "%.3f" % s["ttft90"],
+                   "%.3f" % s["tpot50"], "%.3f" % s["tpot90"], "%.3f" % s["e2e50"]))