Unblocking The Main Thread - Solving ANRs and Frozen Frames
- 1. @snnkzk Unblocking The Main Thread: Solving ANRs and Frozen Frames Sinan Kozak Staff Android Developer Google Developer Expert
- 2. @snnkzk 2 Sinan Kozak 12 years Android experience Google Developer Expert Staff Android Developer @snnkzk I wrote a chapter about performance in 2014 book with GDG community Ankara, Turkey.
- 3. @snnkzk 3 We deliver food over 70 countries with a single rider application
- 4. @snnkzk 4 Production is wild, really wild for Android 10s of thousands of different devices
- 5. @snnkzk 5 5
- 6. @snnkzk 6 Different stages of Jank What to measure?
- 7. @snnkzk 7 Rendering a frame between 16ms and 700ms Scroll Animations Frame render Slow Rendering Rendering a frame between 700ms and 5s App startup Navigation Frozen Frame Blocking main thread over 5 seconds Not being able to handle user inputs Application Not Responding Different stages of Jank
- 8. @snnkzk 8 Apps have limited resources Focus on things we can control
- 9. @snnkzk 9 What bad rendering looks like
- 10. @snnkzk 10 What acceptable render looks like
- 11. @snnkzk 11 Can slow runtime cause app crashes? YES ● User will continue to click the screen. New events without being able to handle to current event will cause unexpected app states. ● Foreground services crashes when their start is delayed
- 12. @snnkzk 12 Is there a magic solution? Maybe…
- 13. @snnkzk 13 Reuse thread pool Reducing allocation Fix memory leaks Release Memory Pressure R8/Proguard optimization Fast dependency Injection WebP and vector images Use reportFullyDrawn Baseline profile Render First Frame Execute in background Reduce recomposition Use correct Compose APIs Measure and optimize Drain Main Thread Different categories of solutions
- 14. @snnkzk 14 Quick wins 01 Execute in background RxJava, Coroutine and other thread solutions 02 Thread pool reuse Utilize thread pools 03 R8/Proguard optimization Auto apply proven optimizations 04 Reducing allocation Only store necessary data in memory 05 No memory leak Fix all leaks 06 Dependency Injection Time consuming injection should be in background 07 WebP and vector images Simpler the image faster to draw 08 Less recomposition Avoid recurring recomposition 09 Correct Compose API Use lambda based api for fast changing values 10 reportFullyDrawn Talk to OS about what needs to prioritize 11 Baseline profile Make sure OS optimize app startup 12 Measure and optimize We can have optimized apps
- 15. @snnkzk 15 15 Reducing allocation Garbage collector pauses runtime to clean up Spare memory for all features Photo by Andreas Gücklhorn on Unsplash
- 17. @snnkzk 17 In your heap dump, look for memory leaks caused by any of the following: ● Long-lived references to Activity, Context, View, Drawable, and other objects that might hold a reference to the Activity or Context container. ● Non-static inner classes, such as a Runnable, that can hold an Activity instance. ● Caches that hold objects longer than necessary. Memory leak In your heap dump, look for memory leaks caused by any of the following: ● Long-lived references to Activity, Context, View, Drawable, and other objects that might hold a reference to the Activity or Context container. ● Non-static inner classes, such as a Runnable, that can hold an Activity instance. ● Caches that hold objects longer than necessary.
- 18. @snnkzk 18 Running with memory leaks
- 19. @snnkzk 19 Memory leak Detection: ● Leak Canary ● Android Studio ● Strict Mode ● Third party memory analyzer Leak Canary Plumber fixes known common issues. https://github.com/square/leakcanary/tree/main/plumber
- 20. @snnkzk 20 Share Thread Pools A single thread can use 1-3 MB memory
- 21. @snnkzk 21 Share Thread Pools Most of features have their own thread pool ● RxJava ● Coroutine ● OkHttp ● Image loader ● AsyncTask ● AndroidX arch components ● RecyclerView + DiffUtil ● WorkManager ● … many more And they have APIs to change thread pool
- 22. @snnkzk 22 Share Thread Pools We shared Coroutine Dispatchers with ● RxJava ● OkHttp ● Coil ● WorkManager Thread count reduced 180+ to 130~ after refresh action “asOkHttpExecutorService” is a custom implementation that only handle execute function
- 23. @snnkzk 23 Runtime optimization Code we write is not same as core runs on device
- 24. @snnkzk 24 R8 and proguard are mainly obfuscation tools, but… They also shrink and optimize codes. In order to optimize your app even further, R8 inspects your code at a deeper level to remove more unused code or, where possible, rewrite your code to make it less verbose. The following are a few examples of such optimizations: ● If your code never takes the else {} branch for a given if/else statement, R8 might remove the code for the else {} branch. ● If your code calls a method in only a few places, R8 might remove the method and inline it at the few call sites. Prefer "proguard-android-optimize.txt" over default Improve app performance by up to 30% R8/Proguard optimization
- 25. @snnkzk 25 More resources optimizations High level documentation https://developer.android.com/build/shrink-code https://www.guardsquare.com/manual/configuration/optimizations Detailed posts https://jakewharton.com/blog/
- 26. @snnkzk 26 Use multi thread solutions ● IO operations ● Data parsing ● Encryption ● Any time consuming operation Move work to the background Photo by Campbell on Unsplash
- 27. @snnkzk 27 Execute in background with Coroutine https://kotlinlang.org/docs/coroutines-overview.html https://developer.android.com/kotlin/coroutines#use-coroutines-for-main-safety
- 29. @snnkzk 29 Recomposition Recompose count shouldn’t affect correctness. The compose runtime might end up recomposing a lot more than you’d expect. Just looking at the code does not reveal reasons. Functions can be inline Classes can be mutable
- 30. @snnkzk 30 https://dev.to/zachklipp/scoped-recomposition-jetpack-compose-what-happens-when-state-changes-l78 Donut hole skipping Compose API design enforces better performance with lambda and scoping
- 31. @snnkzk 31 Donut hole skipping ● Lambdas are special in Compose ● The runtime keeps track of values used in lambdas ● Lambdas get reevaluated if values change ● Composable can be skipped if they are not affected by change
- 33. @snnkzk 33 Rendering first pixel Costly operations during start up
- 34. @snnkzk 34 WebP WebP is an image file format from Google that provides lossy compression (like JPEG) as well as transparency (like PNG) but can provide better compression than either JPEG or PNG Average %26 less space More efficient Android Studio can convert images https://developer.android.com/studio/write/convert-webp Resource optimizations
- 35. @snnkzk 35 Resource optimizations Vector Drawables is a vector graphic defined in an XML file as a set of points, lines, and curves along with its associated color information. Scalable svg can be converted to Vector drawables Simpler is better The initial loading of a vector drawable can cost more CPU cycles than the corresponding raster image. Size save should be justified with performance Use Avocado tool to simplify complex vectors https://github.com/alexjlockwood/avocado
- 37. @snnkzk 37 Dependency Injection DI slowness effects first frame rendering
- 38. @snnkzk 38 Dependency Injection P50 P90 One of the our major improvements is creating OkHttp and Encrypted SharedPreference in the background thread.
- 39. @snnkzk 39 Dependency Injection Delay creation of expensive instances Create in background thread If you use Dagger, get instances from Provider or dagger.Lazy
- 40. @snnkzk 40 Play Store helps you Use reportFullyDrawn for better cloud profiles
- 41. @snnkzk 41 Baseline Profiles ● Helps app startup after new install or update ● Android run ahead of time compile to prepare app startup ● We can modify and create aggregated profiles ● Libraries can have their own baseline profiles ● Compose benefit from baseline profiles https://developer.android.com/topic/performance/baselineprofiles/create-baselineprofile
- 42. @snnkzk 42 Show me real example Dependency injection issue Inefficient vector drawable usage
- 43. @snnkzk 43 Effect of Vector Drawable
- 44. @snnkzk 44 Creation of Vector Images can be consuming
- 45. @snnkzk 45 Effect of Vector Drawable
- 46. @snnkzk 46 46 Talk to designers for simpler svgs You can use webp images or image loading libraries
- 47. @snnkzk 47 Random ANR points during app start up ANR stacktrace is delayed Android 14 has improved ANR capturing Filter ANR for Android 14 to get better picture ANR report only contain last executed things There could be other slowness on main thread
- 48. @snnkzk 48 Measure as much as you can ● Profile a variant close to release ● Measure high level first ● Add more granular metrics ● Use benchmark tests ● Profile before and after changes ● Collect data from production
- 49. @snnkzk 49 49 How to measure? Real data vs profiling
- 50. @snnkzk 50 How to measure? Data from profiling ● perfetto.dev ● Systrace ● CPU and memory allocation ● Layout inspector and debugger Data from users ● Firebase Performance Monitoring or similar solutions ● JankStat library ● Memory leaks ● Strict mode
- 51. @snnkzk 51
- 52. @snnkzk 52 52 Collect StrictMode violations Easiest way to collect data
- 55. @snnkzk 55 Visible performance is one side of the coin App needs to be performant in all operations.
- 56. @snnkzk 56 10 years retro of performance suggestions Comments are welcome in QA
- 57. @snnkzk 57
- 58. @snnkzk 58 10 years retro of performance suggestions View hierarchy - ViewStub > Composables with if check ListView - ViewHolder > RecyclerView + LazyColumn Memory leaks > Memory leaks still exists Running in background > Coroutine - RxJava We still have things to optimized. They are just different.
- 59. @snnkzk 59 I hope you have this chart in your next release
- 60. @snnkzk Thank you github.com/kozaxinan linkedin.com/in/sinankozak strava.com/athletes/sinankozak @snnkzk Do you have any questions?