[Feat] (tools): the cross-host acceptance driver and its report reducer, and a note that netbench's nc sink understates the link

2026-11-07 01:53:24 +09:00 · 2026-09-28 23:22:05 -04:00
parent b4f0c53426
commit b60da6729d
3 changed files with 164 additions and 0 deletions
@@ -76,8 +76,19 @@ server hang at eglMakeCurrent. Exit 0 = end to end, 1 = middlebox,
 2 = filtered. Fix: `ip route replace <phone>/32 via <LAN gateway> dev eth0`
 as root in WSL (temporary; delete it afterwards).

+## crosshost_accept.sh + crosshost_report.py — P12 exit gate (b)
+
+A WSL client replays the openra and rd12 traces against the phone's render server over
+TCP + stream (ssim gate, then `PRESENT_CREDIT` 1/2/3 benchmarks). Adds a temporary /32 route
+around the host's TUN proxy, checks the path with `tcp_path_check.sh`, and removes the route and
+stops the server on exit. `crosshost_report.py` reduces the output to a table. Results and the
+P6.5 link numbers: `docs/Disaggregated/notes/p12/CROSSHOST-ACCEPTANCE.md`.
+
 ## netbench.sh — Wi-Fi TCP throughput baseline

+**Understates the link**: the `toybox nc` sink capped at 18-22 MiB/s while the render server
+read 52 MB/s on the same Wi-Fi. Use the server's `P65ServerFrame` read rate as the throughput number.
+
 WSL -> phone bulk transfer over `toybox nc` sink, 3 reps; refuses to run
 unless `tcp_path_check.sh` passes. The `nc` listener must keep stdin open or it
 exits on EOF and RSTs the connection, and `pkill` must be anchored
@@ -0,0 +1,88 @@
+#!/bin/bash
+# P12 exit gate (b): a client on ANOTHER machine (WSL on the Windows host) replays the in-world
+# Minecraft trace against the phone's render server over TCP + stream data plane, and the P6.5
+# link numbers are recorded. Run from Git Bash. Results: $PERF_WORK/crosshost/<run>/.
+#
+#   bash crosshost_accept.sh                 # ssim gate for both traces, then credit 1/2/3 benchmarks
+#   RUNS="ssim:openra bench:rd12:1" bash crosshost_accept.sh
+#
+# The WSL default route may be a TUN proxy (see tcp_path_check.sh). This script adds a temporary
+# /32 route around it for the phone, and removes it on exit.
+set -u
+export MSYS_NO_PATHCONV=1
+HERE="$(cd "$(dirname "$0")" && pwd)"
+REPO="$(cd "$HERE/../../.." && pwd)"; REPO_WIN="$(cygpath -m "$REPO")"
+SER="${SER:-2f7cbe2e}"
+PKG="${PKG:-top.mobilegl.plugin.trace}"
+PHONE="${PHONE:-192.168.21.181}"
+GATEWAY="${GATEWAY:-192.168.31.1}"
+TOKEN="${TOKEN:-crosshost-token-2026-09}"
+WSL="${WSL_DISTRO:-Ubuntu}"
+PERF_WORK="${PERF_WORK:-$REPO/.trace-work/perf-compare}"
+RUNS="${RUNS:-ssim:openra ssim:rd12 bench:openra:1 bench:openra:2 bench:openra:3 bench:rd12:1 bench:rd12:2 bench:rd12:3}"
+TRACES='$HOME/wifi-perf/traces'
+WSL_OUT='$HOME/wifi-perf/out/crosshost'
+LOCAL_OUT="$PERF_WORK/crosshost"
+log() { echo "[$(date +%H:%M:%S)] $*"; }
+
+cleanup() {
+  wsl.exe -d "$WSL" -u root -- ip route del "$PHONE/32" via "$GATEWAY" dev eth0 2>/dev/null
+  python "$REPO_WIN/tools/trace_replay/tcp_device_server.py" stop --serial "$SER" --package "$PKG" >/dev/null 2>&1
+  log "cleanup: temporary route removed, server stopped"
+}
+trap cleanup EXIT
+
+wsl.exe -d "$WSL" -u root -- ip route replace "$PHONE/32" via "$GATEWAY" dev eth0 || exit 1
+bash "$HERE/tcp_path_check.sh" || { log "FATAL: path is not end to end"; exit 1; }
+
+adb -s "$SER" logcat -c
+python "$REPO_WIN/tools/trace_replay/tcp_device_server.py" start --serial "$SER" --package "$PKG" \
+  --backend DirectGLES --listen tcp://0.0.0.0:40613 --token "$TOKEN" >/dev/null 2>&1 || { log "FATAL: server start"; exit 1; }
+for i in $(seq 1 30); do
+  adb -s "$SER" logcat -d -s MobileGL:* | tr -d '\r' | grep -q 'listening on tcp://0.0.0.0:40613' && break; sleep 1
+done
+adb -s "$SER" logcat -d -s MobileGL:* | tr -d '\r' | grep -q 'listening on tcp://0.0.0.0:40613' || { log "FATAL: server not listening"; exit 1; }
+log "server listening on $PHONE:40613 (phone $(adb -s "$SER" shell dumpsys battery | tr -d '\r' | sed -n 's/.*temperature: \([0-9]*\).*/\1/p' | head -1) x0.1C)"
+
+mkdir -p "$LOCAL_OUT"
+for run in $RUNS; do
+  IFS=: read -r mode wl credit <<<"$run"
+  [ "$mode" = ssim ] && credit=1
+  name="$mode-$wl${credit:+-c$credit}"; [ "$mode" = ssim ] && name="ssim-$wl"
+  case "$wl" in
+    openra) trace=openra; golden=openra.0000031249.png; target=31249; geom="--width 640 --height 480 --crop-x 1 --crop-y 1 --crop-width 638 --crop-height 478"; tmo=600 ;;
+    rd12)   trace=minecraft-1.21.4-rd12-odinlite-in-world; golden=minecraft-1.21.4-rd12-odinlite-in-world.0004660351.png; target=4660351; geom="--width 854 --height 480"; tmo=1500 ;;
+    *) log "unknown workload $wl"; continue ;;
+  esac
+  if [ "$mode" = ssim ]; then
+    mode_args="--target-call $target --golden $TRACES/$golden --ssim-threshold 0.99 $geom"
+  else
+    mode_args="--benchmark --benchmark-result=$WSL_OUT/$name/benchmark.json"
+  fi
+  # One session at a time: no stray client of ours, and the previous session child reaped, or the
+  # supervisor answers Refuse{Busy} and the run measures nothing.
+  wsl.exe -d "$WSL" -- pkill -f mobilegl_trace_replay 2>/dev/null
+  for i in $(seq 1 60); do
+    n=$(adb -s "$SER" shell 'ps -A -o NAME | grep -c "^libMobileGLServer.so"' | tr -d '␍')
+    [ "${n:-0}" -le 1 ] && break; sleep 1
+  done
+  [ "${n:-0}" -le 1 ] || { log "FATAL: a session child is still alive after 60 s"; exit 1; }
+  log "RUN $name"
+  adb -s "$SER" logcat -c
+  wsl.exe -d "$WSL" -- bash -c "
+    set -u; cd ~/MobileGL; rm -rf $WSL_OUT/$name; mkdir -p $WSL_OUT/$name
+    export MOBILEGL_TRANSPORT=spawn MOBILEGL_IPC_CONTROL=tcp://$PHONE:40613 MOBILEGL_IPC_DATA=stream
+    export MOBILEGL_IPC_TOKEN=$TOKEN MOBILEGL_IPC_RUN_AHEAD=1 MOBILEGL_IPC_COLD_START_MS=120000
+    export MOBILEGL_IPC_PRESENT_CREDIT=$credit MOBILEGL_PIPE_STATS=1
+    SECONDS=0
+    timeout $tmo ./build-split/tools/trace_replay/mobilegl_trace_replay --trace $TRACES/$trace.trace \
+      --backend DirectGLES --mobilegl-library ./build-split/libMobileGL.so --pbuffer-surface \
+      --output $WSL_OUT/$name $mode_args > $WSL_OUT/$name/mobilegl.log 2>&1
+    echo rc=\$? secs=\$SECONDS > $WSL_OUT/$name/rc.txt"
+  mkdir -p "$LOCAL_OUT/$name"
+  cp -r "//wsl.localhost/$WSL/home/swung/wifi-perf/out/crosshost/$name/." "$LOCAL_OUT/$name/" 2>/dev/null
+  adb -s "$SER" logcat -d > "$LOCAL_OUT/$name/phone-logcat.txt" 2>/dev/null
+  log "  $(cat "$LOCAL_OUT/$name/rc.txt" 2>/dev/null) $(grep -o '"ssim": [-0-9.]*' "$LOCAL_OUT/$name/result.json" 2>/dev/null | head -1) $(grep -o '"fps": [0-9.]*' "$LOCAL_OUT/$name/benchmark.json" 2>/dev/null)"
+  sleep 3
+done
+log "ALL DONE"
@@ -0,0 +1,65 @@
+#!/usr/bin/env python3
+"""Reduce crosshost_accept.sh output (.trace-work/perf-compare/crosshost/<run>/) to a table."""
+import json, os, re, statistics, sys
+
+root = sys.argv[1] if len(sys.argv) > 1 else os.path.join(
+    os.path.dirname(__file__), "..", "..", "..", ".trace-work", "perf-compare", "crosshost")
+
+def load(path):
+    try:
+        with open(path, encoding="utf-8") as f:
+            return json.load(f)
+    except Exception:
+        return None
+
+def kv(line):
+    return dict(re.findall(r"(\w+)=([^\s]+)", line))
+
+rows = []
+for name in sorted(os.listdir(root)):
+    d = os.path.join(root, name)
+    if not os.path.isdir(d):
+        continue
+    rc = open(os.path.join(d, "rc.txt")).read().strip() if os.path.exists(os.path.join(d, "rc.txt")) else "?"
+    res = load(os.path.join(d, "result.json")) or {}
+    bench = load(os.path.join(d, "benchmark.json")) or {}
+    summary = {}
+    frame_bytes = {}
+    client = os.path.join(d, "mobilegl.client.log")
+    if os.path.exists(client):
+        for line in open(client, encoding="utf-8", errors="replace"):
+            if "P65LinkMetrics kind=summary" in line:
+                summary = kv(line)
+            elif "P65LinkMetrics kind=frame " in line:
+                f = kv(line)
+                frame_bytes[int(f["frame"])] = int(f["stage_bytes"])
+    ft = bench.get("frameTimesMs") or []
+    # Steady state = every frame after the first (the load frame carries the whole texture upload).
+    steady = [k for k in sorted(frame_bytes) if 2 <= k <= len(ft)]
+    steady_s = sum(ft[k - 1] for k in steady) / 1000 if steady else 0
+    steady_mb = sum(frame_bytes[k] for k in steady) / 1e6 if steady else 0
+    rows.append({
+        "run": name, "rc": rc,
+        "ssim": res.get("ssim") if "ssim" in name else None,
+        "frames": bench.get("totalFrames"),
+        "fps": bench.get("fps"), "mean_ms": bench.get("meanFrameMs"),
+        "median_ms": bench.get("medianFrameMs"), "p95_ms": bench.get("p95FrameMs"),
+        "first_frame_s": round(ft[0] / 1000, 1) if ft else None,
+        "steady_fps": round(1000 / statistics.median(ft[1:]), 1) if len(ft) > 2 else None,
+        "steady_MB_per_frame": round(steady_mb / len(steady), 2) if steady else None,
+        "steady_MBps": round(steady_mb / steady_s, 1) if steady_s else None,
+        "wait_replies": summary.get("wait_replies"),
+        "rtt_mean_ms": round(float(summary["rtt_mean_us"]) / 1000, 1) if "rtt_mean_us" in summary else None,
+        "rtt_p50_ub_ms": round(int(summary["rtt_p50_upper_us"]) / 1000, 1) if "rtt_p50_upper_us" in summary else None,
+        "rtt_p99_ub_ms": round(int(summary["rtt_p99_upper_us"]) / 1000, 1) if "rtt_p99_upper_us" in summary else None,
+        "stage_MB": round(int(summary["stage_bytes"]) / 1e6) if "stage_bytes" in summary else None,
+        "client_cpu_s": round(int(summary["client_thread_cpu_ns"]) / 1e9, 1) if "client_thread_cpu_ns" in summary else None,
+        "wall_s": round(int(summary["wall_ns"]) / 1e9, 1) if "wall_ns" in summary else None,
+    })
+
+cols = ["run", "rc", "ssim", "frames", "fps", "steady_fps", "median_ms", "p95_ms", "first_frame_s",
+        "steady_MB_per_frame", "steady_MBps", "wait_replies", "rtt_mean_ms", "rtt_p50_ub_ms", "rtt_p99_ub_ms", "stage_MB", "client_cpu_s", "wall_s"]
+print("| " + " | ".join(cols) + " |")
+print("|" + "---|" * len(cols))
+for r in rows:
+    print("| " + " | ".join("" if r[c] is None else str(r[c]) for c in cols) + " |")