android-performance

作者： openai

使用Simpleperf CPU性能分析、Perfetto或Compose跟踪、gfxinfo帧数据、dumpsys等在adb目标上收集并解读Android性能证据。

npx skills add https://github.com/openai/plugins --skill android-performance

Android Performance

Use this skill to capture Android performance evidence for adb-installable apps. CPU sampling usually requires a debuggable or profileable build; frame stats, Perfetto, and logcat can still help when an app cannot be sampled. Compose with ../android-emulator-qa/SKILL.md for device selection, build/install/launch, UI driving, screenshots, UI trees, and logcat capture.

Core Workflow

Pick one focused user-visible flow.
Choose the trace type that matches the question.
Record the flow with clear start and stop boundaries.
Pull or copy the trace produced by that run, then generate reports from that file.
Interpret reports with caveats about device, build type, sample count, and profiler limits.

Avoid broad "use the app for a while" captures. They make traces hard to attribute and usually hide the functions that matter.

Use a local adb target for meaningful timing. Store outputs in a run-specific artifact folder outside the skill directory:

if [ -z "${ARTIFACT_DIR:-}" ]; then
  ARTIFACT_DIR="$(mktemp -d "${TMPDIR:-/tmp}/codex-android-perf.XXXXXX")"
fi
mkdir -p "$ARTIFACT_DIR"

Do not put ARTIFACT_DIR under SKILL_DIR; the skill folder is for bundled instructions and scripts, not run artifacts.

Choosing A Trace

Use Simpleperf when the question is "what functions are taking CPU time?" or when you need a sampled profile of Kotlin, Java, native, or framework execution.
Use Perfetto when the question is frame timing, startup timeline, scheduler gaps, binder work, lock contention, main-thread stalls, Compose recomposition, or why a flow felt janky.
Use gfxinfo framestats for a quick manual frame/jank snapshot. Pair it with Perfetto when you need root cause.
Use meminfo / heap dumps when the question is retained Java/Kotlin objects, PSS, native heap, or object counts after a focused flow.

Simpleperf CPU Profiles

Simpleperf --app works best when the installed package is debuggable or profileable from shell. Preflight before recording:

SERIAL="<adb-serial>"
PACKAGE="<app package>"

adb -s "$SERIAL" shell dumpsys package "$PACKAGE" | grep -Ei 'DEBUGGABLE|profileable|isProfileable' || true

If the package is not debuggable/profileable and simpleperf record --app fails, install a debug/profileable build when possible. If that is not possible, use Perfetto or gfxinfo instead of treating missing CPU samples as evidence.

Start recording in one terminal or as a long-running Codex command session:

SERIAL="<adb-serial>"
PACKAGE="<app package>"
MAX_DURATION_SECONDS=60

adb -s "$SERIAL" shell rm -f /data/local/tmp/perf.data
adb -s "$SERIAL" logcat -c

adb -s "$SERIAL" shell simpleperf record \
  --app "$PACKAGE" \
  -o /data/local/tmp/perf.data \
  -e cpu-clock -f 4000 -g \
  --duration "$MAX_DURATION_SECONDS"

While that command is running, perform exactly one focused flow with adb input, UI automation, or android-emulator-qa.

Stop Simpleperf from another command and wait for the recording command to exit:

adb -s "$SERIAL" shell 'pid="$(pidof simpleperf 2>/dev/null || true)"; [ -n "$pid" ] && kill -INT $pid'

If that returns Operation not permitted, send Ctrl-C to the original adb shell simpleperf record command session and wait for it to exit.

Pull and report the capture:

adb -s "$SERIAL" pull /data/local/tmp/perf.data "$ARTIFACT_DIR/perf.data"
adb -s "$SERIAL" logcat -d > "$ARTIFACT_DIR/logcat.txt"

SKILL_DIR="<absolute path to this loaded skill folder>"
FIRST_PARTY_REGEX="$(printf '%s' "$PACKAGE" | sed 's/\./\\./g')"
"$SKILL_DIR/scripts/simpleperf_hotspots.sh" \
  "$ARTIFACT_DIR/perf.data" \
  "$ARTIFACT_DIR" \
  --serial "$SERIAL" \
  --first-party-regex "$FIRST_PARTY_REGEX"

Do not derive SKILL_DIR from the target app repo's pwd; installed plugins usually live outside the app being profiled. Keep FIRST_PARTY_REGEX scoped to the app's package or app-owned module prefixes; avoid broad framework patterns such as kotlin, Compose, or androidx.compose when reporting app-owned rows.

The helper writes:

$ARTIFACT_DIR/simpleperf-self.txt
$ARTIFACT_DIR/simpleperf-children.txt
$ARTIFACT_DIR/simpleperf.csv when supported by the installed Simpleperf

If host Simpleperf is not installed, the helper searches Android Studio and Android SDK/NDK locations. If unavailable, it falls back to device-side adb shell simpleperf report when the device still has /data/local/tmp/perf.data.

Reading Simpleperf

Simpleperf reports sampled CPU execution. It does not directly measure suspended coroutines, network latency, lock wait time, or other wall-clock waits. If a flow feels slow but Simpleperf shows little app CPU, capture Perfetto to inspect scheduler gaps, binder work, locks, frame timing, and app trace sections.

Read reports this way:

Self/Overhead: samples where the function itself was executing. Use this for hot leaf work such as parsing, formatting, diffing, sorting, allocation-heavy iteration, or JSON/protobuf processing.
Children/inclusive: samples in the function and its callees. Use this for expensive entry points such as repositories, use cases, view models, Composables, startup initializers, or feature coordinators.
Shared Object / Symbol: prefer app-owned package frames, feature modules, domain/data/UI modules, and generated app code. Treat Android framework, Kotlin runtime, Compose, and native/runtime frames as context unless the app-owned caller is visible.
Percentages: useful for ranking functions inside one capture. For user-facing timing claims, pair with Perfetto, gfxinfo, or repeated wall-clock measurements.

When interpreting a hotspot, note symbol/function name, self or inclusive percentage, approximate sampled CPU time when available, caller stack or owning source file, flow steps, artifact paths, and whether the capture is single-run or repeated.

Perfetto / Compose Trace

If the app repo already documents a Perfetto/System Trace command for that project, use it. Otherwise use Perfetto directly. The light command below captures scheduler/frequency/Android atrace categories and app Trace sections for PACKAGE; it is not a substitute for a full project-specific Perfetto config when you need detailed frame timeline or Compose runtime internals.

SERIAL="<adb-serial>"
PACKAGE="<app package>"
TRACE_DURATION_SECONDS=30
TRACE_BASENAME="app-flow-$(date +%Y%m%d-%H%M%S).pftrace"
TRACE_DEVICE="/data/misc/perfetto-traces/$TRACE_BASENAME"

PERFETTO_PID="$(adb -s "$SERIAL" shell perfetto \
  --background-wait \
  -o "$TRACE_DEVICE" \
  -t "${TRACE_DURATION_SECONDS}s" \
  --app "$PACKAGE" \
  sched freq idle am wm gfx view binder_driver hal dalvik | tr -d '\r' | tail -n 1)"
printf 'Perfetto PID: %s\n' "$PERFETTO_PID"

Run exactly one focused flow before TRACE_DURATION_SECONDS expires. To stop early, gracefully terminate the background Perfetto process and give it a moment to flush:

adb -s "$SERIAL" shell kill -TERM "$PERFETTO_PID" 2>/dev/null || true
adb -s "$SERIAL" shell "
  last_size=-1
  stable_count=0
  i=0
  while [ \$i -lt 30 ]; do
    size=\$(ls -l '$TRACE_DEVICE' 2>/dev/null | awk '{ print \$5 }')
    if [ -n \"\$size\" ] && [ \"\$size\" -gt 0 ] && [ \"\$size\" = \"\$last_size\" ]; then
      stable_count=\$((stable_count + 1))
      [ \$stable_count -ge 2 ] && exit 0
    else
      stable_count=0
    fi
    last_size=\"\${size:-0}\"
    i=\$((i + 1))
    sleep 1
  done
  exit 1
"

Prefer letting TRACE_DURATION_SECONDS expire instead of stopping early. If the stop command fails because the trace already ended, still wait until the output file exists and its size is stable before pulling. If the direct command is too coarse, use Android Studio System Trace or a project-specific Perfetto config. Only report frame timeline or Compose recomposition details when those tracks/events are actually present in the captured trace; the light command above does not guarantee them.

Pull the exact on-device trace from this run:

adb -s "$SERIAL" pull "$TRACE_DEVICE" "$ARTIFACT_DIR/$TRACE_BASENAME"

In Perfetto, inspect:

main-thread slices around missed frames or long startup sections
frame scheduling, frame timeline, and render thread lanes
Compose runtime tracing sections for recomposition work when enabled
binder transactions, monitor contention, scheduler gaps, and app log markers

gfxinfo Framestats

Use this for a quick manual frame snapshot:

SERIAL="<adb-serial>"
PACKAGE="<app package>"

adb -s "$SERIAL" shell pidof "$PACKAGE"
adb -s "$SERIAL" shell dumpsys window | grep -F "$PACKAGE"
adb -s "$SERIAL" shell dumpsys gfxinfo "$PACKAGE" reset
# Perform the focused flow.
adb -s "$SERIAL" shell dumpsys gfxinfo "$PACKAGE" > "$ARTIFACT_DIR/gfxinfo.txt"
adb -s "$SERIAL" shell dumpsys gfxinfo "$PACKAGE" framestats > "$ARTIFACT_DIR/gfxinfo-framestats.txt"

Capture from a stable, responsive screen. If dumpsys gfxinfo fails to dump the process, or the device shows an ANR/dialog/splash screen instead of the flow, discard that capture and use Perfetto for root cause.

Read the headline summary first: total frames, janky frames, frame percentiles, slow UI thread, slow draw commands, and frame deadline misses. On emulators, absolute smoothness numbers are noisy; percentile spikes and slow draw/UI counters are still useful for deciding whether to take a Perfetto trace.

Memory / Leak Artifacts

Use this on an adb target after narrowing the investigation to one flow. Exercise the flow, return to a stable screen, then capture memory artifacts from that state.

For quick Java/native/PSS/object-count snapshots:

SERIAL="<adb-serial>"
PACKAGE="<app package>"

adb -s "$SERIAL" shell am force-stop "$PACKAGE"
adb -s "$SERIAL" shell monkey -p "$PACKAGE" 1
# Exercise the focused flow, then navigate back to a stable idle screen.
adb -s "$SERIAL" shell dumpsys meminfo "$PACKAGE" > "$ARTIFACT_DIR/meminfo-flow.txt"

Read TOTAL PSS, Java heap, native heap, graphics, Views, Activities, binder counts, and object counts. Treat one noisy sample as a lead, not a conclusion.

For retained Kotlin/Java objects, prefer Shark CLI when it is available. It works with Android heap dumps and produces text output the agent can inspect and cite.

HEAP="/data/local/tmp/app-flow.hprof"
HPROF="$ARTIFACT_DIR/app-flow.hprof"

if ! command -v shark-cli >/dev/null; then
  echo "Install Shark CLI, or analyze the HPROF with Android Studio Profiler / MAT." >&2
fi

adb -s "$SERIAL" shell am dumpheap -g "$PACKAGE" "$HEAP"
adb -s "$SERIAL" pull "$HEAP" "$HPROF"
adb -s "$SERIAL" shell rm -f "$HEAP"

if command -v shark-cli >/dev/null; then
  shark-cli --hprof "$HPROF" analyze | tee "$ARTIFACT_DIR/shark-analysis.txt"
fi

Read shark-analysis.txt first when it exists. Report suspected leaking objects, retained sizes, and reference chains. Look for retained feature objects, activities, fragments, view models, Compose state holders, repositories, listeners, callbacks, and caches that should have been released after leaving the flow. If Shark CLI is unavailable, still preserve the HPROF path and inspect it with the best available heap analyzer; do not claim leak roots from meminfo alone.

For native allocation growth, capture a Perfetto trace with heapprofd enabled. Keep the duration in the config; current Android perfetto rejects -t together with --config.

TRACE_DEVICE="/data/misc/perfetto-traces/native-alloc.pftrace"

adb -s "$SERIAL" shell perfetto -o "$TRACE_DEVICE" \
  --txt -c - <<EOF
duration_ms: 60000
buffers { size_kb: 262144 fill_policy: RING_BUFFER }
data_sources {
  config {
    name: "android.heapprofd"
    heapprofd_config {
      sampling_interval_bytes: 65536
      shmem_size_bytes: 8388608
      block_client: true
      process_cmdline: "$PACKAGE"
    }
  }
}
EOF

adb -s "$SERIAL" pull "$TRACE_DEVICE" "$ARTIFACT_DIR/native-alloc.pftrace"

Analyze the trace with trace_processor_shell and save the outputs:

SKILL_DIR="<absolute path to this loaded skill folder>"
"$SKILL_DIR/scripts/heapprofd_reports.sh" \
  "$ARTIFACT_DIR/native-alloc.pftrace" \
  "$ARTIFACT_DIR"

Read heapprofd-summary.txt, heapprofd-top-allocations.txt, heapprofd-top-stack.txt, heapprofd-health.txt, and meminfo together. Report net native allocation size, top allocating frames/mappings, the expanded stack for the largest callsite, and whether trace stats show heapprofd health issues such as client errors, packet loss, or buffer overruns. Prefer Java heap dumps for retained app objects; heapprofd is for native allocation behavior.

Report

Report:

exact flow, device/emulator, Android version, build variant, and run count
artifact paths for every trace/report used
top hotspots or frame/jank evidence with percentages, durations, or counts
whether evidence is CPU samples, frame timeline, frame stats, or memory artifacts
caveats such as emulator noise, low sample count, cold-start compilation, or missing symbols
next smallest trace or code change when current evidence is insufficient