91d21b6 chore(bench): add idle page, per-thread wakes and A/B root · neferwl

M go.mod +3 -0


@@ -24,3 +24,6 @@ require (
24	24		go.yaml.in/yaml/v3 v3.0.5 // indirect
25	25		gopkg.in/natefinch/lumberjack.v2 v2.2.1 // indirect
26	26		)
	27	+
	28	+	// Prototype requires the unreleased netpoller display loop.
	29	+	replace github.com/bnema/purego-libwayland => ../../../purego-libwayland/.worktrees/proto-netpoll


@@ -24,3 +24,6 @@ require (
24	go.yaml.in/yaml/v3 v3.0.5 // indirect	24		go.yaml.in/yaml/v3 v3.0.5 // indirect
25	gopkg.in/natefinch/lumberjack.v2 v2.2.1 // indirect	25		gopkg.in/natefinch/lumberjack.v2 v2.2.1 // indirect
26	)	26		)
		27	+
		28	+	// Prototype requires the unreleased netpoller display loop.
		29	+	replace github.com/bnema/purego-libwayland => ../../../purego-libwayland/.worktrees/proto-netpoll

M tools/bench/README.md +2 -0


@@ -6,3 +6,5 @@
6	6		tools/bench/bench.sh gif.html 15 # 32x32 GIF, 20 fps
7	7
8	8		Environment: `BENCH_X11=1` runs chromium through XWayland, `BENCH_DEBUG=<categories>` adds `--debug` categories, `BENCH_KEEP=1` leaves the processes running, `BENCH_PPROF=<addr>` sets the pprof address (default `localhost:6061`).
	9	+
	10	+	`idle.html` is a static page: run it with `BENCH_IDLE=1 tools/bench/bench.sh idle.html 15` (no frame is required). The output also reports `cs/s` and, for the thread with the most voluntary switches (usually the Go runtime's sysmon), its scheduler wakes per second (`top_wakes/s`, schedstat field 3). `BENCH_ROOT=<worktree>` builds neferwl from another worktree for A/B runs.


@@ -6,3 +6,5 @@
6	tools/bench/bench.sh gif.html 15 # 32x32 GIF, 20 fps	6		tools/bench/bench.sh gif.html 15 # 32x32 GIF, 20 fps
7		7
8	Environment: `BENCH_X11=1` runs chromium through XWayland, `BENCH_DEBUG=<categories>` adds `--debug` categories, `BENCH_KEEP=1` leaves the processes running, `BENCH_PPROF=<addr>` sets the pprof address (default `localhost:6061`).	8		Environment: `BENCH_X11=1` runs chromium through XWayland, `BENCH_DEBUG=<categories>` adds `--debug` categories, `BENCH_KEEP=1` leaves the processes running, `BENCH_PPROF=<addr>` sets the pprof address (default `localhost:6061`).
		9	+
		10	+	`idle.html` is a static page: run it with `BENCH_IDLE=1 tools/bench/bench.sh idle.html 15` (no frame is required). The output also reports `cs/s` and, for the thread with the most voluntary switches (usually the Go runtime's sysmon), its scheduler wakes per second (`top_wakes/s`, schedstat field 3). `BENCH_ROOT=<worktree>` builds neferwl from another worktree for A/B runs.

M tools/bench/bench.sh +18 -6


@@ -7,12 +7,14 @@
7	7		# BENCH_DEBUG=cat extra --debug categories (comma separated)
8	8		# BENCH_KEEP=1 leave neferwl and chromium running (prints PIDs)
9	9		# BENCH_PPROF=addr pprof listen address (default localhost:6061)
	10	+	# BENCH_ROOT=dir build neferwl from another worktree (A/B runs)
	11	+	# BENCH_IDLE=1 static page: no frames expected, per-second numbers only
10	12		set -euo pipefail
11	13
12	14		page=${1:?usage: bench.sh <page.html> <seconds>}
13	15		secs=${2:?usage: bench.sh <page.html> <seconds>}
14	16		here=$(cd "$(dirname "$0")" && pwd)
15		-	root=$(cd "$here/../.." && pwd)
	17	+	root=${BENCH_ROOT:-$(cd "$here/../.." && pwd)}
16	18		tmp=$(mktemp -d)
17	19		pids=()
18	20

@@ -67,22 +69,32 @@ cs() { cat /proc/"$nwl"/task/*/status \| awk '/^voluntary_ctxt/{v+=$2} END{print
67	69		cpu_ns() { cat /proc/"$nwl"/task/*/schedstat \| awk '{v+=$1} END{printf "%d\n", v}'; }
68	70		frames() { grep -c '"message":"frame"' "$log" \|\| true; }
69	71
70		-	t0=$(cpu_ns); f0=$(frames); c0=$(cs)
	72	+	# Per-thread "tid sched_in_count voluntary_switches" (schedstat field 3 = times scheduled in).
	73	+	threads() { for d in /proc/"$nwl"/task/*; do
	74	+	echo "${d##*/} $(awk '{print $3}' "$d/schedstat" 2>/dev/null) $(awk '/^voluntary_ctxt/{print $2}' "$d/status" 2>/dev/null)"
	75	+	done; }
	76	+
	77	+	th0=$(threads); t0=$(cpu_ns); f0=$(frames); c0=$(cs)
71	78		sleep "$secs"
72		-	t1=$(cpu_ns); f1=$(frames); c1=$(cs)
	79	+	th1=$(threads); t1=$(cpu_ns); f1=$(frames); c1=$(cs)
73	80
74		-	if [ "$((f1 - f0))" -le 0 ]; then
	81	+	if [ -z "${BENCH_IDLE:-}" ] && [ "$((f1 - f0))" -le 0 ]; then
75	82		echo "no frame presented in ${secs}s: the page did not animate or neferwl stalled" >&2
76	83		tail -n 20 "$tmp/nwl.out" >&2
77	84		exit 1
78	85		fi
79	86
80		-	python3 - "$((t1 - t0))" "$((f1 - f0))" "$((c1 - c0))" "$secs" "$page" <<'EOF'
	87	+	python3 - "$((t1 - t0))" "$((f1 - f0))" "$((c1 - c0))" "$secs" "$page" "$th0" "$th1" <<'EOF'
81	88		import sys
82	89		t, f, c, s, page = int(sys.argv[1]), int(sys.argv[2]), int(sys.argv[3]), float(sys.argv[4]), sys.argv[5]
	90	+	def parse(x):
	91	+	return {l.split()[0]: tuple(int(v) for v in l.split()[1:]) for l in x.splitlines() if len(l.split()) == 3}
	92	+	a, b = parse(sys.argv[6]), parse(sys.argv[7])
	93	+	d = {k: (b[k][0] - a.get(k, (0, 0))[0], b[k][1] - a.get(k, (0, 0))[1]) for k in b}
	94	+	top = max(d, key=lambda k: d[k][1])
83	95		cpu = t / 1e9 / s * 100
84	96		fr = max(f, 1)
85		-	print(f"{page} cpu%={cpu:.2f} fps={f / s:.1f} ms/frame={t / 1e6 / fr:.3f} cs/frame={c / fr:.0f}")
	97	+	print(f"{page} cpu%={cpu:.2f} fps={f / s:.1f} ms/frame={t / 1e6 / fr:.3f} cs/frame={c / fr:.0f} cs/s={c / s:.0f} top_tid={top} top_wakes/s={d[top][0] / s:.0f} top_vol/s={d[top][1] / s:.0f}")
86	98		EOF
87	99
88	100		if [ -n "${BENCH_KEEP:-}" ]; then

 #        BENCH_DEBUG=cat  extra --debug categories (comma separated)
 #        BENCH_KEEP=1     leave neferwl and chromium running (prints PIDs)
 #        BENCH_PPROF=addr pprof listen address (default localhost:6061)
+#        BENCH_ROOT=dir   build neferwl from another worktree (A/B runs)
+#        BENCH_IDLE=1     static page: no frames expected, per-second numbers only
 set -euo pipefail
 page=${1:?usage: bench.sh <page.html> <seconds>}
 secs=${2:?usage: bench.sh <page.html> <seconds>}
 here=$(cd "$(dirname "$0")" && pwd)
-root=$(cd "$here/../.." && pwd)
+root=${BENCH_ROOT:-$(cd "$here/../.." && pwd)}
 tmp=$(mktemp -d)
 pids=()
 cpu_ns() { cat /proc/"$nwl"/task/*/schedstat | awk '{v+=$1} END{printf "%d\n", v}'; }
 frames() { grep -c '"message":"frame"' "$log" || true; }
-t0=$(cpu_ns); f0=$(frames); c0=$(cs)
+# Per-thread "tid sched_in_count voluntary_switches" (schedstat field 3 = times scheduled in).
+threads() { for d in /proc/"$nwl"/task/*; do
+	echo "${d##*/} $(awk '{print $3}' "$d/schedstat" 2>/dev/null) $(awk '/^voluntary_ctxt/{print $2}' "$d/status" 2>/dev/null)"
+done; }
+th0=$(threads); t0=$(cpu_ns); f0=$(frames); c0=$(cs)
 sleep "$secs"
-t1=$(cpu_ns); f1=$(frames); c1=$(cs)
+th1=$(threads); t1=$(cpu_ns); f1=$(frames); c1=$(cs)
-if [ "$((f1 - f0))" -le 0 ]; then
+if [ -z "${BENCH_IDLE:-}" ] && [ "$((f1 - f0))" -le 0 ]; then
 	echo "no frame presented in ${secs}s: the page did not animate or neferwl stalled" >&2
 	tail -n 20 "$tmp/nwl.out" >&2
 	exit 1
 fi
-python3 - "$((t1 - t0))" "$((f1 - f0))" "$((c1 - c0))" "$secs" "$page" <<'EOF'
+python3 - "$((t1 - t0))" "$((f1 - f0))" "$((c1 - c0))" "$secs" "$page" "$th0" "$th1" <<'EOF'
 import sys
 t, f, c, s, page = int(sys.argv[1]), int(sys.argv[2]), int(sys.argv[3]), float(sys.argv[4]), sys.argv[5]
+def parse(x):
+    return {l.split()[0]: tuple(int(v) for v in l.split()[1:]) for l in x.splitlines() if len(l.split()) == 3}
+a, b = parse(sys.argv[6]), parse(sys.argv[7])
+d = {k: (b[k][0] - a.get(k, (0, 0))[0], b[k][1] - a.get(k, (0, 0))[1]) for k in b}
+top = max(d, key=lambda k: d[k][1])
 cpu = t / 1e9 / s * 100
 fr = max(f, 1)
-print(f"{page} cpu%={cpu:.2f} fps={f / s:.1f} ms/frame={t / 1e6 / fr:.3f} cs/frame={c / fr:.0f}")
+print(f"{page} cpu%={cpu:.2f} fps={f / s:.1f} ms/frame={t / 1e6 / fr:.3f} cs/frame={c / fr:.0f} cs/s={c / s:.0f} top_tid={top} top_wakes/s={d[top][0] / s:.0f} top_vol/s={d[top][1] / s:.0f}")
 EOF
 if [ -n "${BENCH_KEEP:-}" ]; then

A tools/bench/idle.html +3 -0

@@ -0,0 +1,3 @@
	1	+	<!doctype html>
	2	+	<html><head><meta charset="utf-8"><title>idle</title></head>
	3	+	<body style="margin:0;background:#202030;color:#ddd;font:24px sans-serif"><p>NeferWL idle benchmark: static page, nothing animates.</p></body></html>

@@ -0,0 +1,3 @@
	1	+	<!doctype html>
	2	+	<html><head><meta charset="utf-8"><title>idle</title></head>
	3	+	<body style="margin:0;background:#202030;color:#ddd;font:24px sans-serif"><p>NeferWL idle benchmark: static page, nothing animates.</p></body></html>