Bullet stress test with 5K dynamic cubes

Post Reply
hyyou
Posts: 73
Joined: Wed Mar 16, 2016 10:11 am

Bullet stress test with 5K dynamic cubes

Post by hyyou » Thu Sep 14, 2017 6:53 am

I am trying to encapsulate Bullet to my ECS game engine.
After some test, I can simulate 2000 cubes, dropped randomly on a 2D static plane.
Cubes rarely touch each other.

With rendering + my engine cost, in PC, I get :-
  • >80 fps in normal case
  • 50 fps in the worst case (a brief period that all cubes start to touch the ground)
I ran with :-
  • Windows 7, Intel Core i5-4460 CPU@3.20GHz
  • Visual C++, -O2, release mode
  • single thread
  • without continuous detection (i.e. no CCD)

In this test, I don't share shape, i.e. I use 1 btBoxShape per 1 btRigidBody.
According to the profile, Bullets take about 30% of overall execution time.

My conclusion

Without multi-threading, Bullet can support around 5,000 cubes (if I try very hard).

Question

Is it a correct estimation?
How many btRigidBody your favorite engine can support? (if any)
In 2017, is it practical to achieve https://www.youtube.com/watch?v=J9HaT23b-xc (tower 5,000 planks) in real time?

It is so hard to find some reliable Bullet stress test.
I would like to hear the rough number before designing more gameplay / optimizing further.

Please answer if you happen to know something. I am so lonely in this darkness.

Dirk Gregorius
Posts: 874
Joined: Sun Jul 03, 2005 4:06 pm
Location: Kirkland, WA

Re: Bullet stress test with 5K dynamic cubes

Post by Dirk Gregorius » Thu Sep 14, 2017 9:28 pm

Pierre has done excessive testing which you can read about here:
http://www.codercorner.com/blog/?p=748

User avatar
drleviathan
Posts: 386
Joined: Tue Sep 30, 2014 6:03 pm
Location: San Francisco

Re: Bullet stress test with 5K dynamic cubes

Post by drleviathan » Fri Sep 15, 2017 9:22 pm

Interesting. That was back in 2013 and development has continued on both sides. I wonder what the comparisons look like now.

The GDC-2017 talks by Nvidia about PhysX were impressive. I'm sure they're keeping the CPU side competitive but they're working hard to make the simulation run even faster on Nvidia GPU's.

Dirk Gregorius
Posts: 874
Joined: Sun Jul 03, 2005 4:06 pm
Location: Kirkland, WA

Re: Bullet stress test with 5K dynamic cubes

Post by Dirk Gregorius » Sat Sep 16, 2017 1:40 am

Pierre open sourced Peel. You can download it here:
https://github.com/Pierre-Terdiman/PEEL

I have integrated Rubikon into as well and use it constantly to compare performance. So you can run these tests yourself.

Post Reply