Performance regression

[mrdoob]

We are considering using cannon-es in three.js but seems like there's a performance regression.

@Mugen87 converted cannon.js to modules and here's the running example:
https://raw.githack.com/mrdoob/three.js/dev/examples/physics_cannon_instancing.html

This is the same example using cannon-es:
https://raw.githack.com/Mugen87/three.js/dev45/examples/physics_cannon_instancing.html

mrdoob/three.js#19709

[marcofugaro]

Thanks for the issue and the usage example.
This also looks related to #16. I'll look into it soon!

[codynova]

I only had a moment to look at this as I'm headed to bed but I definitely agree it looks related to #16 and #18

That githack service is awesome, I've never seen that before.

[arpu]

maybe i am wrong here but the last para should be not the frameRate
https://github.com/mrdoob/three.js/blob/dev/examples/jsm/physics/CannonPhysics.js#L193

more a value of 10 for maxSubSteps

[marcofugaro]

Hello friends, I've discovered the issue.

Basically there are tons or commit which are in the cannon.js repo but haven't been published to npm or built into the cannon.js bundle. Here is the full list.
When we built the source files again in cannon-es, we inherited automatically all those commits.

Between these commits there is 86b0444, where schteppe implemented the fixed timestepping from the last snippet of this famouse article, however it was left unfinished or was not tested properly, leading to the framerate issues.

I've tried to fix those issues in #40, but of course I'm not schteppe. If you guys @mrdoob and @Mugen87 could test with the cannon-es bundle from that branch that would be grand 🙏

Also @codynova issue #16 should be fixed as well, could you do a test? 🙏

@arpu is right, the third parameter is used to tell how many maximum internal steps the step() function can take if there is too much lag. If you set it to 1, it doesn't take any internal steps, but the simulation is technically slower. If you set it to an high number, it takes as many steps as needed to make the interaction have a fixed timestep, however if the step is expensive (a lot of boxes), taking too many steps can cause lag by itself.

[Mugen87]

@marcofugaro The performance seems better with your changes. However if I have more than 400 cubes in the scene, the performance drops significantly to an unusable framerate (approx. 8 FPS). This does not happen with the original cannon.js version.

[marcofugaro]

@Mugen87 Oddly enough, that's a feature.
Please read the description of #40, it's the whole idea behind this article

If you want the behaviour of the previous cannon.js, you can call world.step(frameTime) without any other arguments or just world.step(frameTime, delta, 1). The simulation with 400 boxes will run "in slow motion" but at a stable framerate.

However now that you mention it, it makes sense to have the threshold at at least 30fps, not 10.

EDIT: I've updated the PR, could you test now?

[marcofugaro]

@Mugen87 do you have time to test again? 🙏
It shouldn't go below 30fps, rather it will slow down a bit.

[Mugen87]

Performance seems better now. Mostly around 25-27 FPS when having 400 cubes in physics_cannon_instancing.

However, the performance of cannon-es is still worse than the original cannon.js. With the same settings (400 cubes in physics_cannon_instancing), I get around 52 FPS with cannon.js.

[marcofugaro]

Alright, that is working as intended! 🎉

The simulation now has more time to catch up than before, this works in favor of people using cannon.js in a web worker.

Users can still get the previous cannon.js behaviour (~60fps) by passing 1 to the third argument of the step() function (suggested when there are a lot of bodies to simulate).

world.step(frameTime, delta, 1)

Thanks for testing again!

[Mugen87]

So THREE.CannonPhysics needs an change in the following place to restore the "old" behavior?

https://github.com/mrdoob/three.js/blob/3ca144cac1c81ada7496e2f4790877848c6b4aa9/examples/jsm/physics/CannonPhysics.js#L193

[marcofugaro]

Yup, currently in that line, the third parameter frameRate is wrong, as it's not related to the framerate at all.

It should be called maxSubSteps and should be set to a value of 1 or a really small integer in your case.
It was like this also in the original cannon.js.

We should really write this stuff in the docs once they are up.

[codynova]

A new version of cannon-es has been released with Marco's fixes for variable framerate simulation [email protected]