Skip to content

android_latency_benchmark.md

Latest commit

 

History

History
47 lines (34 loc) · 1.52 KB

android_latency_benchmark.md

File metadata and controls

47 lines (34 loc) · 1.52 KB
Raw

Android Latency Benchmark Recipe (ADB + TFLite)

Use repeatable ADB commands and collect enough runs to reduce variance.

Prerequisites

  • Android device with developer mode + USB debugging.
  • adb installed and device visible with adb devices.
  • benchmark_model binary and your .tflite model on device.

Minimal workflow

# 1) Push model and benchmark binary
adb push model_int8.tflite /data/local/tmp/model_int8.tflite
adb push benchmark_model /data/local/tmp/benchmark_model
adb shell chmod +x /data/local/tmp/benchmark_model

# 2) Optional: set fixed governor/profile if your lab policy allows it
# adb shell su -c 'echo performance > /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor'

# 3) Warm-up + timed runs
adb shell /data/local/tmp/benchmark_model \
  --graph=/data/local/tmp/model_int8.tflite \
  --num_threads=4 \
  --warmup_runs=20 \
  --num_runs=100

Recommended reporting

  • Device model, Android version, SoC, and thermal state.
  • Thread count and delegate used (CPU, NNAPI, GPU).
  • P50/P90/P99 latency and standard deviation.
  • Peak RSS/memory (if available).

Thermal and variance gotchas

  • Run after cooldown; avoid charging during measurement.
  • Discard first run block if app/process cold-start dominates.
  • Repeat across multiple sessions; single-run numbers are noisy.
  • Keep background services minimal (airplane mode helps consistency).

Optional tracing

  • Use Perfetto for scheduling/thermal traces around benchmark windows.
  • Correlate latency spikes with frequency throttling events.