Tools to Help You Improve Android App Performance

By : Janice Kartika, Product Engineer - Gojek
Live : https://youtu.be/qhkNL0o7x3o?t=4429

Transcript

1. Tools to Help You Improve App Performance Janice Kartika Product Engineer - Gojek Android Dev Talks 2021
2. Contents 1 Bad and good app performance 2 Basic performance test 3 Profile GPU Rendering tool 4 Debug GPU Overdraw and Layout Inspector tool 5 Android (Memory) Profiler tool
3. Bad and Good App Performance Signs we can see and/or feel to determine bad/good app performance
4. Bad • Takes a long time to load or does not load • Stutter animations • App crashes randomly • Slow response to user actions • Drains battery • Use lots of data plan Good • Loads fast • Smooth animations • No crashes • Fast response to user actions • Respects limited battery resources • Optimizes data transmission
5. Frame Rate • 60 FPS (frames per second) • Appears as smooth motion to our eyes • 1000 ms / 60 frames = 16-ish ms for 1 frame • This 16 ms is not only for our app, but Android system also
6. Basic Performance Test No special tool required
7. Steps • (If possible) Install your app on low-end physical device. • Try yourself to crash the app + observing performance issue you feel + take notes. • Once done, ask at least 5 other people to do the same. This is a usability testing. • Why 5 people? [link to Jakob Nielsen’s research]
8. Profile GPU Rendering Tool To know which rendering pipeline is slowing down your app’s frame time
9. Enable Profile GPU Rendering Developer mode has to be enabled first, then:
10. Read the Graph Horizontal green line: 16 ms / frame Vertical bar: • Shows time needed to render a frame • Higher = slower • Color codes represent rendering stages each frame • Below green line: frame is rendered less than 16 ms • Above green line: frame is rendered more than 16 ms
11. Color Codes Draw High blue bar: • Views are invalidated • Complex logic inside onDraw method
12. Color Codes Command Issue High red bar: • Complex custom views • Requesting views to be redrawn (not necessarily invalidated) • Calling too many draw operations:
13. Color Codes Swap Buffers Represents the time the CPU is waiting for the GPU to finish its work. High orange bar: • App is doing too much work on GPU • Same optimization as Command Issue phase
14. Color Codes Sync & Upload The time to upload bitmap information from CPU to the GPU. Avoid high blue bar: • Reduce large image size (resizing). • Use prepareToDraw() to pre-upload beforehand.
15. Color Codes Measure / Layout Measure = calculate the size of objects on the screen. Layout = sizing and positioning objects on the screen. High green bar: • Nested layouts (try to optimize view hierarchy) • Check the code inside onMeasure() and onLayout()
16. Debug GPU Overdraw and Layout Inspector Tool Overdraw = where a pixel is redrawn multiple times in a single frame of rendering
17. Enable Debug GPU Overdraw Developer mode has to be enabled first, then:
18. Overdraw Color Codes True color → no overdraw → 1x → 2x → 3x → 4x or more Normal GPU Overdraw Enabled
19. Reduce Overdraw • Remove unnecessary, overlapping backgrounds • Flatten view hierarchy • Use Layout Inspector tool to help further (debug mode only)
20. Layout Inspector View hierarchy View properties Display
21. 3D View Layout Inspector • Display UI hierarchy in 3D view: rotation mode • Find this button (hope it’s enabled for you): • Available from API 29 and above • GIF source
22. Memory Profiler Tool As we have talked about optimizations in CPU and GPU, how about memory?
23. Android Memory 101 • Memory allocation → reserve memory for our app’s objects, processes, services in the assigned memory heap. • The used memory can be reclaimed when objects are no longer accessible (no reference to the objects). Garbage collector will do the work. • Too many temporary objects created by our code can pollute the memory heap → indirectly trigger multiple garbage collection events → degrade overall performance. • When the memory can’t be reclaimed because the reference to the objects are never released → memory leak. • Less and less memory slows down the app, eventually app would crash when no memory left available (last resort to free up memory).
24. Be Friendly with Memory • NEVER reference view objects outside UI thread or from static objects, avoid leaks. • Avoid allocating objects in inner loops to prevent too many garbage collection events. • Avoid allocating objects in onDraw() methods as this method is called 60 times per second (to achieve the smoothness 60 FPS). • Example:
25. Android Profiler Tool User inputs CPU Memory Network Energy
26. Memory Profiler Tool GC Force GC Capture heap dump
27. Example Memory-Inefficient App Repo link. Highlighted code:
28. Example Memory-Inefficient App • Image Source link • Increased memory over time • GC is performed but the decreased is less than the increased • Eventually app crashed
29. Dump and Inspect Heap • Snapshot of allocated memory at specific point in time. • Allocations: number of instance of that class in memory. • Shallow Size: total size of all instances of that class. • Retained Size: its shallow size plus the shallow sizes of the objects that are accessible from that class.
30. Dump and Inspect Heap • Depth: shortest number of hops from any GC root to the selected instance. • Shallow Size: size of the instance. • Retained Size: total size of memory being retained due to all instances in that class.
31. Case Study: Gojek • Not all views are required to be initialised because user may not need it at runtime. • Lazy initialization can be a solution. • Only create instance of views that user really needs to see.
32. Bonus: System Tracing • Enable from developer mode. • Record trace of your app. • The .perfetto-trace or .ctrace file will be generated. • Use Perfetto UI to open the file.
33. Bonus: System Tracing
34. Know What You Want to Know • I know from Profile GPU Rendering that a custom view is causing slow frame rendering. • Looks like the issue also exists in CPU rendering pipeline. • Using systrace, I want to know the time needed to execute draw method of that custom view.
35. We Can Use CPU Profiler, Right??? • Yes we can. • Sometimes profiler causes the app to be much more laggy than usual → affects the profiler result. • Link to related article by Paulina Sadowska.
36. Thank You!