swung0x48
141227e8ba
run_matrix.sh drives interleaved monolith/inproc/spawn+shm/spawn+tcp benchmark rounds on an adb device with per-run arm-marker verification; profile_matrix.sh layers simpleperf, ftrace wakeup capture and a /proc sampler on top; make_report.py aggregates runs into REPORT.md. The wifi kit carries the cross-host control-plane hang repro and phone-side strace/fd-reconcile helpers. .trace-work/ (run output) is now ignored.
3.6 KiB
Disaggregated-mode device perf harness
A/B/C/D performance matrix and profiling harness for MobileGL's transport arms
(monolith / inproc / spawn+shm / spawn+tcp), built for the adb-attached
Android bench device. Written from Git Bash on a Windows host; requires
MSYS_NO_PATHCONV=1 (the scripts set it themselves).
Prerequisites: the disaggregated trace APK installed
(./gradlew -p android-plugin :app:assembleTraceRelease -Pmobilegl.buildDisaggregated=ON -Pmobilegl.buildDisaggregatedInproc=ON -Pmobilegl.pipePush=ON -Pmobilegl.debuggableRelease=true,
then sign with the debug keystore and adb install -r), and trace fixtures under
tools/trace_replay/fixtures/.
All scripts honor SER=<adb serial> (default 2f7cbe2e) and
PERF_WORK=<results root> (default <repo>/.trace-work/perf-compare, git-ignored).
run_matrix.sh — perf matrix driver
bash run_matrix.sh [smoke|prep|openra|rd12|all]
Interleaved A/B/C/D rounds (monolith -> inproc -> shm -> tcp), N reps per arm
(OPENRA_REPS / RD12_REPS, default 3). Interleaving is the substitute for
frequency pinning on unrooted devices. Per run it force-stops the app, clears
stale results, starts/stops the tcp server for the tcp arm, verifies the arm
markers in logcat (Config: MOBILEGL_TRANSPORT=..., spawn ARMED,
run-ahead ARMED — a run without them is invalid), and pulls
result.json / benchmark.json / PipeStats dumps into
$PERF_WORK/<workload>/<arm>-r<n>/.
benchmark mode never snapshots, so ssim stays -1; the matrix adds one non-benchmark ssim round per workload/arm for the golden gate.
make_report.py — aggregate into REPORT.md
python make_report.py [results_dir]
Renders raw per-rep tables, best-of-N aggregates with spread, the ssim gate,
per-run arm verification, and merged client+server PipeStats rollups.
Note the spawn-arm traps: client-side frames=0 is structural (use the server
dump), and logcat MGPipe stats: mixes non-counter lines — trust the JSON dumps.
profile_matrix.sh + run_profile_batch.sh — profiling lane
Same runner with PROF=1: adds simpleperf (perf_event_paranoid=-1 device)
for the client and server processes, ftrace sched capture for exact wakeup
latencies (shell-writable on this device), and the /proc sampler below.
bash run_profile_batch.sh # the canned batch; edit the run list inside
PROF=1 ONE=tcp WL=rd12 REP=1 bash profile_matrix.sh one # single arm
ONE / WL / REP / EXTRA_ENV / EXTRA_SERVER_ENV select and parameterize
a single run; e.g. EXTRA_SERVER_ENV=MOBILEGL_IPC_SPIN_US=2000 for the
deep-idle-exit A/B.
sampler.c — device-side /proc sampler
sampler <pid> <duration_s> <interval_ms> <out.csv>. Per-tid blocked-syscall
histogram (/proc/<tid>/syscall) + CPU/switch counters. Exists because strace
attach is SELinux-blocked for shell uid on the bench device, while same-uid
/proc reads via run-as work. Build for the device with the NDK
(aarch64-linux-android*-clang), push, run under run-as.
netbench.sh — Wi-Fi TCP throughput baseline
WSL -> phone bulk transfer over toybox nc sink, 3 reps. The nc listener
must keep stdin open or it exits on EOF and RSTs the connection.
phone_strace.sh / fdprobe.sh / wifi_hang_repro.sh — cross-host hang kit
Repro and evidence-capture for the cross-host control-plane hang (WSL client ->
phone server wedges at eglMakeCurrent; bytes ACKed, rx queue drained, session
child never sees POLLIN). phone_strace.sh attaches the device
/system/bin/strace as the app uid via run-as (shell-uid ptrace is blocked);
fdprobe.sh reconciles /proc/<pid>/fd socket inodes against /proc/net/tcp.