Multithreaded rendering on Android | Hacker News Reader

Multithreaded rendering on Android

code.facebook.com

87 pointsdowhileone20 comments

This is really nice, but it would be good if they could release some metrics for their example such as the News Feed. How much did it improve the scroll performance? The main reason is because I'd like to know if the extra complexity is worth the effort. I'm sure that from Facebook scale any improvement is important but how good is it?

It is mentioned in the article that Android documentation does not recommend doing multithreading optimizations since the UI Toolkit itself is not thread-safe, does this mean that Android's own UI Toolkit cannot be used within this context? What is the level of integration that Litho and Android's UI Toolkit can have? Also, in regards to Android's Accessibility APIs, does Litho components handle/have that capability?

I guess I should have done a bit more research/experimentation on Litho, but these are some questions I have. I'd really love to use it though.

For the record, I prefer Litho regardless of performance characteristics. The immutable data model works really well with AutoValue, which in turn works well with lots of things in Android and I find the API a lot cleaner than the default RecyclerView API (there are alternative Adapter implementations for RecyclerView that bring it's API closer to Litho)

what alternative adapter implemntations are there ? could you cite anything ?

Airbnb's Epoxy is the best one I've used:

https://medium.com/airbnb-engineering/epoxy-airbnbs-view-arc...

https://github.com/airbnb/epoxy

> How much did it improve the scroll performance? We are still have some more work to do in News Feed but on other surfaces in the app we have seen a 42% improvement in skipped frames when converting from Views to Litho. This is measured by sampling skipped frames while scrolling in production.

> does this mean that Android's own UI Toolkit cannot be used within this context? Views should not be accessed on a background thread according to the Android documentation. Litho renders inside of a light weight wrapper view and interacts with this wrapper view on the UI thread. Nevertheless, Litho moves most of the heavy lifting from rendering to the background.

> What is the level of integration that Litho and Android's UI Toolkit can have? During the conversion to Litho, News Feed rendered with some standard views and some Litho

>in regards to Android's Accessibility APIs, does Litho components handle/have that capability? Yup! It's a full featured UI framework. Animations are under active development though.

Erm, android already does multi-threaded rendering via HWUI[1] at the Canvas level.

What they're talking about doing here is multi-threaded layout which isn't nearly that impressive. Heck, you can call measure()/layout() off-thread right now on views, there's just an invalidate when they get inserted into the hierarchy.

Also something not mentioned is that each CPU core you spin up is eating into XX% of your battery life. So you may get a faster layout but your users may see worse battery usage because of it.

[1] - https://android.googlesource.com/platform/frameworks/base/+/...

That is not correct. There is only one HWUI render thread per app. HWUI doesn't use multithreaded rendering. You can verify this easily running "ps -t".

Yup, but it's different from the UI thread which is what counts for me.

For what it's worth unless you're on Vulkan(which few Android devices support) you're going to be limited to a single dispatch thread by GLES' threading semantics anyway.

I still rest my point that they're doing layout, which is fine but doesn't consist of multi-threaded rendering. Using their approach they're still going to have to set some sort of absolute layout views which must go through a measure and layout pass on the UI thread.

HWUI does have 2 helper threads that it uses to off-load some work (hwuiTask1 & 2). But otherwise yes it's largely dominated by one thread.

Indeed, moving to multiple cores is about doing more work faster which has a direct impact on battery life. You'd have to measure whether this allows you to race-to-sleep faster or not which is a delicate balancing act.

It is telling that battery life is almost never mentioned or discussed in depth anywhere. What is the battery impact of React Native? Or of this solution?

Yup, completely.

It's a really nasty problem too because power usage isn't linear. Chips tend to have voltage steps[1] as they increase in frequency so if you can do something longer and at lower frequency it uses less power than the same workload at a higher frequency for shorter time.

Compound that with the fact that you really want to bring the whole SoC to a low power state and just keep the display lit since 90% of the time users aren't actually moving something.

Arm has tried to solve this with big.LITTLE[2] to mixed results. It turns out that it's hard to build a general purpose scheduler that's power aware much like it's hard to build a general purpose memory allocator that's fast in every case.

[1] https://en.wikipedia.org/wiki/Dynamic_voltage_scaling

[2] https://en.wikipedia.org/wiki/ARM_big.LITTLE

Battery life has been a major focus 2 or 3 iterations of Android ago.

Google released the tool they created in order to benchmark it (battery historian). btw, releasing the tool they created for a specific improvement is what they do each time they focus on improving one aspect of the platform, and that's an excellent practice.

Google probably cares to some extent about battery life. I doubt that the mobile team behind let's say Youtube cares, but the framework team does.

Third party devs tend to do not worry at all about battery life .. unless they use a really excessive amount of battery and are shamed for it

So FB mixes this and react native in their app?

We use both at Facebook! React Native works cross platform whereas Litho provides performance that beats the Android Toolkit - so they are each beneficial for different use cases. Litho is heavily inspired from RN so the programming model should be familiar.

If I remember right, this was created before React Native. And, they still do have a lot of Android code in the app, even with RN.

The Android Framework is a bit of an old framework, not using reactive elements. Google knows this, and is building flutter a reactive UI toolkit.

Also android developers are using Reactive components more and more, but the UI is still very much statefull. So i see a bright future for this approach (maybe not the facebook library, but the approach itself)

I'm not sure if I'm the only one, but since from the recent patent grant debate, the first thing I do on a Facebook git repository is to check whether there's a PATENTS file.

Are there any examples of how to convert an existing project to Litho ?

We've been doing a lot of that at Facebook and we do want to document this process a bit better. Stay tuned :-)

Yet another steaming pile of overengineering, great.

Reply on news.ycombinator.com