MGLRU results on Android
summary
Evaluation of MGLRU on Android uses a Pixel 6 Pro (12 GB RAM, 3 GB zram) and a 30‑app cycle test run for ~200 iterations. Results show 6.6% overall app‑launch reduction, hot starts improve (hot:warm 3.3→3.8, hot latency −~3%), and total kills drop ~8% despite Activity Manager kills rising ~10% and >800 deliberate home kills to avoid generation wrap. Working‑set refills fall sharply (anon −65%, file −81%), reclaim CPU for kswapd drops ~54% and reclaim efficiencies rise, but directory‑claim stalls increase ~90% with long‑tail latencies; larger swap and generation limits mitigate this.
AI-written summary · about
chapters (11)
- 0:00Introduction to MGLRU
- 1:07Evaluation setup on Android
- 2:12App launch time improvements
- 4:03App kill metrics analysis
- 6:34Working set refill results
- 7:17Reclaimed CPU usage efficiency
- 8:24Directory reclaim latency issues
- 10:54Swap size impact on performance
- 11:59No swap device comparison
- 12:45Hot cold generation inversion fix
- 13:53Q&A and future plans
related talks
38mOptimize your app performance for the next generation of Android
38mLow effort, High impact Android performance wins- 38mOptimize your app performance for the next generation of Android
40m640K ought to be enough for anyone
39mUnderstanding Android memory usage
36mAndroid memory and games
46mDeveloping for Android: the Naughty Bits
37mUnderstanding Android Runtime (ART) for faster apps