Performance Regression

[alfonsocj]

Our team has found that the performance on the new architecture is not on par with the old one when a text heavy screen is mounted on the navigation stack.

I have a reproducible example where I ran the tests. Here's a spreadsheet with the results. The raw data for the results can be found here.

The following results are from a release version of the app using an iPhone 12 Pro Max.

Results (times in ms)

Event	Old Arch Non Virtualized	Old Arch Virtualized	New Arch Non Virtualized	New Arch Virtualized
js_navigation_1	42.195	35.59	168.415	28.088
js_navigation_2	27.866	27.735	176.643	25.733
native_navigation_1	29.586	23.3	126.065	17.152
native_navigation_2	31.795	24.164	178.757	20.418
500-views	126.551	135.619	307.222	171.328
1000-views	237.216	235.402	452.423	313.391
1500-views	342.467	352.763	613.525	463.766

Non Virtualized (times in ms)

event	Old Arch	New Arch	Percentage Increase
js_navigation_1	42.195	168.415	-299.13%
js_navigation_2	27.866	176.643	-533.90%
native_navigation_1	29.586	126.065	-326.10%
native_navigation_2	31.795	178.757	-462.22%
500-views	126.551	307.222	-142.77%
1000-views	237.216	452.423	-90.72%
1500-views	342.467	613.525	-79.15%

Virtualized (times in ms)

event	Old Arch	New Arch	Percentage Increase
js_navigation_1	35.59	28.088	21.08%
js_navigation_2	27.735	25.733	7.22%
native_navigation_1	23.3	17.152	26.39%
native_navigation_2	24.164	20.418	15.50%
500-views	135.619	171.328	-26.33%
1000-views	235.402	313.391	-33.13%
1500-views	352.763	463.766	-31.47%

Here's a video of the test performed:
https://user-images.githubusercontent.com/3791120/204658443-30e486cd-f7f1-4ffe-be58-c81ab2159025.MP4

Each test was performed 10 times and the results in the spreadsheet are the average of the events.

Reproduction

https://github.com/alfonsocj/new-arch-perf-example

The results seem to be similar to previous discussions #88, #85 where heavy text screens seem to perform better on the old architecture, even though those tests were performed on Android builds.

I created an issue on React Navigation since it could be related react-navigation/react-navigation#11054.

[cortinico]

Thanks for sharing this @alfonsocj. Have you engaged with folks on React Navigation by any chance?

[alfonsocj]

Hello, nobody from React Navigation has replied to the issue I created. I'll try to ask them on the discord channel. Hopefully someone can reproduce with the example I provided.

[alfonsocj]

An update, I tested on my playground app without React Navigation and I still see a considerable performance regression when there are a lot of nodes, so we can discard the idea of React Navigation causing the performance issue.

[cortinico]

Thanks for flagging. I'd like to pass this to one of my colleague which has more context than me, but is sadly currently on vacation. We'll have to wait a bit till he gets back 👍

[alfonsocj]

Perfect. Let me know if there's something else I can help with.

[alfonsocj]

Happy New Year! @cortinico, could you please pass this to your colleague if he's back from vacation? Thanks!

[alfonsocj]

Hello! @cortinico any progress/updates on this issue?

[cortinico]

cc @sammy-SC if you can provide any insights here

[alfonsocj]

@sammy-SC let me know if there's something I can help you with.

[sammy-SC]

Hello @alfonsocj

thank you for putting this together.

At what point are markers started and ended? I noticed in one of your examples it says "endPerformanceMarker" and is called from onLayout event. Event onLayout does not reflect when a view is painted on the screen. It could be that the old and new infrastructure dispatches the events at different points in the pipeline.
One way to know that a view was painted on the screen is to have a custom component that overrides didMoveToWindow method on UIView and check that window property is not nil.

[sammy-SC]

Hello @alfonsocj,

we spent some time analysing performance and summarised our findings in #123.

Could you please try to run your benchmarks on the 0.72.0-RC.1 and see what the performance is like? We've made some improvements.