Performance Regression When Migrating from Box2D 2.2 to 3.1 in the Defold game engine

[deurell]

Hi Erin and friends of Box2D,

First off, thanks for your continued work on Box2D. We really appreciate the ongoing improvements and the cleaner architecture in 3.x.

We've recently completed a migration from Box2D 2.2 to 3.1 as part of our work on the Defold game engine, and wanted to share some performance results with the hope of getting some insight from the community.

Goal

Evaluate Box2D 3.1 under realistic Defold physics workloads and compare it with our current Box2D 2.2-based integration.

Performance Test Setup

• Platform: macOS (arm64, M1)
• Workload: Step simulation of dynamic bodies in Defold
• Object counts: 10, 100, 1000, 2000, 10,000
• Variants tested:
  • Box2D 2.2 (Optimized Defold fork, single-threaded)
  • Box2D 3.1 (single-threaded)
  • Box2D 3.1 (multi-threaded using threadpool with 1, 2, 4 threads)
• Metrics: Time spent in StepWorld2D, averaged across multiple runs

Summary of Results (ms)

Objects	2.2	3.1 (ST)	3.1 (~1)	3.1 (~2)	3.1 (~4)
10	0.035	0.091	0.709	0.825	1.406
100	0.301	0.427	1.409	1.734	1.706
1000	1.983	3.625	3.902	3.634	3.576
2000	4.047	5.852	5.685	5.471	5.129
10000	8.966	29.352	30.342	26.349	25.084

Observations

• Performance in 3.1 is consistently slower than 2.2, even with multithreading.
• Threading overhead dominates at low body counts. Only at 2000+ bodies does it become marginally competitive with 2.2.
• Our 2.2 fork includes some optimizations (e.g., early-outs in broad/narrow-phase etc.) that may contribute to the delta.
• We’ve also increased maxPolygonVertices in 3.1 and also enable dynamic/static/kinematic interactions, which have a small impact.

Goal / Question

We’d love to know:

• Has anyone else observed similar performance regressions when moving from 2.2 to 3.x?
• Are there recommended tuning options, build flags, or solver settings we may have missed?
• Are there any planned optimizations for 3.1+ that could improve performance for low/medium-body-count games like those typical in Defold?

We’d be happy to collaborate, test patches, or provide more data if needed.

Thanks again for the great work!

– Mikael, Defold team

[aganm]

it would be helpful if your benchmark also contained the box2d 2.2 original (without your optimizations), to see if this performance difference solely comes from your own optimizations, or if we're truly dealing with a performance regression.

[aganm]

also.. was this a debug or release build? with optimization enabled or without?

[erincatto]

Your results table shows 3.1 slower even at higher body counts and poor gains from threading. I ran the barrel benchmark just now comparing 2.4.2 to main (3.1+).


This shows main being about 3x faster than 2.4.2 single threaded.
Version 2.4 is dominated by continuous collision turning that off I get:

Leaving main about 1.8x faster.
Now, using 4 threads on main I get main being over 6x faster than 2.4.2 with continuous off or 10x faster with continuous on.

[erincatto]

I ran lower body counts of around ~100 and they have similar performance. This is not a case I have optimized for in 3.1.