Discussion about LiquidCore performances about liquidcore

[ericwlange]

@capezzbr I found some low hanging fruit and picked it. It improves the speed by 4x.

Test on my emulator with 0.4.5:

02-17 10:07:12.581 26359-26359/? I/RESULT: LiquidCore run in 120ms
02-17 10:07:12.621 26359-26359/? I/RESULT: V8Runner run in 3ms

Test on my emulator with latest changes:

02-17 10:03:51.498 26239-26239/? I/RESULT: LiquidCore run in 38ms
02-17 10:03:51.529 26239-26239/? I/RESULT: V8Runner run in 3ms

Test with 0.4.5 on real hardware (Lenovo Tablet, API 23):

02-17 11:13:31.568 8094-8094/capezzbr.liquidcoreplayground I/RESULT: LiquidCore run in 786ms
02-17 11:13:31.699 8094-8094/capezzbr.liquidcoreplayground I/RESULT: V8Runner run in 11ms

Test with changes on tablet (much better than 4x):

02-17 11:47:54.168 10264-10264/? I/RESULT: LiquidCore run in 170ms
02-17 11:47:54.258 10264-10264/? I/RESULT: V8Runner run in 10ms

Pixel, 0.4.5:

02-17 11:39:31.324 5993-5993/capezzbr.liquidcoreplayground I/RESULT: LiquidCore run in 444ms
02-17 11:39:31.379 5993-5993/capezzbr.liquidcoreplayground I/RESULT: V8Runner run in 12ms

Pixel, latest:

02-17 11:37:39.979 5673-5673/capezzbr.liquidcoreplayground I/RESULT: LiquidCore run in 166ms
02-17 11:37:40.044 5673-5673/capezzbr.liquidcoreplayground I/RESULT: V8Runner run in 14ms

It is checked in if you'd like to try it. Obviously LQ is still an order of magnitude slower. I will have to dig into the Java side to see if there are ways to reduce the overhead.

[ericwlange]

I've been thinking about this, though I haven't been able to dedicate time to it yet. jv8 also uses reflection, albeit at a much simpler level. Thanks for putting the project together. I will try put some time into it this weekend.

[ericwlange]

I have released version 0.4.6. Made some more big improvements. I can see some more opportunities for improvement, but it would require a pretty invasive re-write of the C++ code, so I will just leave that as a project for later. Latest tests with release version:

Emulator:

03-11 09:28:43.255 3820-3820/? I/RESULT: V8Runner run in 5ms
03-11 09:28:43.277 3820-3820/? I/RESULT: LiquidCore run in 10ms

Ancient Samsung Galaxy Nexus (API 18):

03-11 10:30:43.918 5423-5423/capezzbr.liquidcoreplayground I/RESULT: V8Runner run in 101ms
03-11 10:30:44.215 5423-5423/capezzbr.liquidcoreplayground I/RESULT: LiquidCore run in 200ms

So, for this particular test, LiquidCore is still about half the speed as J2V8, but obviously much closer in performance.

Let me know if you have any issues with this version!

[capezzbr]

👋 @ericwlange, sorry for the bump, just wondering if you could think at a way I could improve the performances?

[capezzbr]

@ericwlange thanks for looking into that, let me know if I can help somehow, I'll be more than happy to test any future implementation 👍

[capezzbr]

@ericwlange wow, that looks promising. Just updated the app I'm working on to use com.github.LiquidPlayer:LiquidCore:master-SNAPSHOT, and the speed improvements are definitely noticeably, great work 👏.

I will have to dig into the Java side to see if there are ways to reduce the overhead.

That would be awesome, I feel that if LQ would be as performant as jv8 I'll be definitely able to switch to use LQ in production, since it is way more stable and I like the fact that I can use almost the same API I'm using on the twin iOS app 👍 let me know if there is anything I could help with 🙏

[ericwlange]

Just a quick update on this. I found many sources of the slowness, and most all if it is related to reflection. I have a version that is twice as fast as the one I sent you last (not checked in yet), but there is still more to do. I will spend some time on it this weekend and let you know when something is ready to test.

[capezzbr]

That sounds fantastic, thanks again for looking into that 👍

[ericwlange]

Ok, please give the latest a try ...

My emulator:

03-04 12:19:18.477 16198-16198/? I/RESULT: V8Runner run in 4ms
03-04 12:19:18.511 16198-16198/? I/RESULT: LiquidCore run in 17ms

There is obviously still some difference between the two, but J2V8 is an incredibly thin (and hence, less feature-rich) wrapper around V8 so it will be difficult to achieve quite that level. I do think there is some more that I can do, but it will all be on the margin at this point. Can you see if this meets your needs? I will test it some more and put it into a release if you're satisfied.

Thanks!

[capezzbr]

Hey @ericwlange, just tested your latest improvements on the app I'm working on, and the performance improvement are fantastic 👏

There is obviously still some difference between the two, but J2V8 is an incredibly thin (and hence, less feature-rich) wrapper around V8 so it will be difficult to achieve quite that level.

That makes totally sense 👍

[capezzbr]

Hey @ericwlange thanks a lot for the hard work on investigating on those performance improvements 👏 Will definitely check out the new version soon and let you know if I find anything off 👍