Skip to content

performance.md

Latest commit

 

History

History
94 lines (67 loc) · 7.58 KB

performance.md

File metadata and controls

94 lines (67 loc) · 7.58 KB
Raw

성능

숫자는 재현 가능한 demo 경로의 물리 기기 결과다. prod는 실서버·네트워크·위치가 섞여서 커밋 숫자로 쓰지 않는다.

핵심 흐름

Journey What It Measures Metric
Startup to first content Cold app launch until the first usable station-list content is visible and reportFullyDrawn() is reached StartupTimingMetric
List scroll Frame stability while scrolling the price-first station list FrameTimingMetric
Refresh Frame stability while refreshing seeded nearby station data FrameTimingMetric
Open watchlist Frame stability while saving a station and opening the watchlist comparison screen FrameTimingMetric

최근 물리 기기 측정

  • Device: Samsung Galaxy S20+ 5G (SM-G986N)
  • Android version: 13 (API 33, build TP1A.220624.014)
  • Hardware: 8 cores, 2.84 GHz max, 11.1 GB RAM, CPU not locked
  • Build variant: demoBenchmark (isDebuggable=false, isProfileable=true, isMinifyEnabled=true via R8)
  • Compilation mode: verify (no baseline profile installed for this run — see Known Limitations)
  • Measurement date: 2026-05-18 (KST)
  • Source data: benchmark/build/outputs/connected_android_test_additional_output/benchmark/connected/SM-G986N - 13/com.gasstation.benchmark-benchmarkData.json

Reproduce with:

find benchmark/build/outputs/connected_android_test_additional_output -name '*benchmarkData.json' -print

결과

Hero journey Primary metric p50 p95 Iterations Samples
Startup to first content timeToInitialDisplayMs 347 ms 393 ms 10 10
Startup to first content timeToFullDisplayMs 546 ms 622 ms 10 10
List scroll frameDurationCpuMs 3.84 ms/frame 6.83 ms/frame 5 225
List scroll frameOverrunMs -3.50 ms/frame -0.48 ms/frame 5 225
Refresh frameDurationCpuMs 3.83 ms/frame 6.05 ms/frame 5 185
Refresh frameOverrunMs -3.42 ms/frame -1.15 ms/frame 5 185

Negative frameOverrunMs values mean the frame finished its work that far ahead of its display deadline; positive values would indicate jank. p95 ≤ 0 across both scroll and refresh on this device means no dropped frames at the 95th percentile of observed samples.

Baseline profile 흐름

The baseline profile generator covers:

  • App startup
  • First station-list content
  • Seeded refresh
  • Station-list scroll
  • Watchlist entry after saving a station

The generator and its companion openWatchlistFrameTiming benchmark depend on station-list-watch-toggle, the persistent bottom-nav-watchlist tab, and watchlist-card appearing within the benchmark helper timeout. These ASCII test tags are exposed as resource IDs so benchmark selectors stay separate from Korean accessibility copy. See Known Limitations for the current status of those two scenarios.

명령

./gradlew :app:assembleDemoBenchmark :benchmark:assembleBenchmark
ANDROID_SERIAL=<device serial> ./gradlew :app:installDemoBenchmark :benchmark:connectedBenchmarkAndroidTest

The :app benchmark build type forks release with isDebuggable=false, isProfileable=true, and the debug signing config so the same minified APK macrobenchmark expects can be installed and traced without a release keystore. The connected command installs demoBenchmark before running the test APK because the benchmark module launches the target activity explicitly as com.gasstation.demo/com.gasstation.MainActivity.

결과 읽기

  • Startup numbers are used to replace the README startup metric table. timeToInitialDisplayMs is the moment the first frame after Activity launch lands; timeToFullDisplayMs is the moment reportFullyDrawn() is called (i.e., the first real station content is laid out).
  • Frame timing numbers (frameDurationCpuMs) describe how long the UI thread spent producing each frame during the journey. p95 below the device's frame budget (~16.6 ms at 60 Hz, ~8.3 ms at 120 Hz) means scroll/refresh are not the bottleneck on this device.
  • Perfetto traces are local diagnostic artifacts and are not committed unless a future investigation needs a small excerpt or screenshot.

한계

  • Engineering-quality gate timing is not app performance evidence. The 52-suite/90-test convention run and its later narrow marker fix measure build verification, not a hero journey. The governed Linux invocation stopped at infrastructure preflight before attempt allocation, so Linux 90/90, package seal, recovery, cleanup, and default-Colima noninterference remain NOT_MEASURED; no new physical-device, emulator, hosted, or reproducible-APK run was performed. Therefore none of that evidence changes the Macrobenchmark table or APK-size snapshot above.
  • API 24/28/36 에뮬레이터 결과는 성능 측정이 아니다. 기기 검증 워크플로는 demo UI, Room migration, API 36 Geocoder callback의 정확성을 본다. 에뮬레이터 결과가 나중에 PASS여도 물리 기기 Macrobenchmark 표, trace, README 성능 주장을 바꾸지 않는다. 지금 런타임 상태는 NOT RUN이다. 기기 검증을 본다.
  • Baseline profile not installed. BaselineProfileGenerator.collectHeroJourney did not produce a committed physical-device profile in the latest measured run, so compilation mode stays at verify. Startup numbers above are realistic for first-install / post-update users and represent a lower-bound improvement target once a baseline profile is generated.
  • openWatchlistFrameTiming은 아직 물리 기기에서 다시 측정하지 않았다. 벤치마크 헬퍼는 com.gasstation.demo/com.gasstation.MainActivity를 직접 열고 station-list-watch-toggle, bottom-nav-watchlist, watchlist-card를 기다린다. 기기 검증은 같은 selector 흐름을 Pixel 8 API 37 에뮬레이터에서 확인했고, :benchmark:assemble은 제품 벤치마크 계약을 컴파일된 채로 유지한다. 다만 새 물리 기기 JSON/trace는 없다. ANDROID_SERIAL=<physical device> ./gradlew :app:installDemoBenchmark :benchmark:connectedBenchmarkAndroidTest가 통과하고 find benchmark/build/outputs/connected_android_test_additional_output -name '*benchmarkData.json' -print에 새 JSON이 보이기 전에는 README 성능 숫자를 바꾸지 않는다.
  • Cooling and thermal state not enforced. macrobenchmark warned about SUSTAINED_PERFORMANCE_MODE being unavailable; results above are the median over 10 startup iterations and 5 frame iterations, which mitigates but does not eliminate device-side thermal variance. Re-run on a cooled device before committing future numbers if comparisons span multiple firmware revisions.

APK 크기 (demo)

R8 minification on the benchmark build type produces a usable size baseline for what production-shaped users would download. Measured from the same build that produced the numbers above:

Variant APK Size
demoBenchmark (R8 minify=true, profileable, debug-signed) app/build/outputs/apk/demo/benchmark/app-demo-benchmark.apk 2.51 MB
demoDebug (minify=false, debuggable) app/build/outputs/apk/demo/debug/app-demo-debug.apk 22.70 MB

The roughly 9× difference confirms R8 + resource shrinking is doing the work expected of it for the production-shaped APK; it does not represent a separate optimization opportunity, just the build-type cost on this codebase. Reproduce with:

./gradlew :app:assembleDemoBenchmark :app:assembleDemoDebug
ls -l app/build/outputs/apk/demo/{benchmark,debug}/*.apk