I'm not aware of any benchmarks, but generally agree with TMNCorp's assessment. dyn4j, while no slouch, made some design choices which reduces performance slightly, but yielded a more structured, clearly defined, codebase and API.
TMNCorp wrote:I've searched and the only related thing I've found is a test on the sandbox called Parallel which says does use multithreading, but I'm not sure exactly for what.
The Parallel test was actually an attempt to multi-thread some parts of the engine. The experiment failed - it always yielded worse performance. The number of bodies needed to be so high for a performance difference to be seen vs. single threaded that you'd be out of the realm of realtime simulation anyway. This is not a general statement about multi-threaded physics, rather, it's the experience I had with dyn4j and the constraints on API and maintenance that I set for myself.
The best attempt at "multi-threaded" physics I've seen so far is with Bullet and it's full GPU implementation. But there are limitations there as well.
However, I do think they have the right approach - the whole physics pipeline needs to be re-tooled with parallel algorithms to get any benefit. This is non-trivial. So from that, you have to ask the question, is it really necessary? Can we achieve the same thing another way? Hence TMNCorp's LOD reference and the ecosystem that is currently available.
A simple, but not always obvious, approach might be to have a physics "World" per thread, that simulates a specific region of a larger world. Naturally, this adds complexity (swapping bodies from one world to another across boundries, what happens when a body is in more than one, etc) but would scale as you say. However, it may prove easy depending on your application, thereby removing the need for the extra complexity of a multi-threaded physics engine.
The only thing I would disagree with is
TMNCorp wrote:and lacks some niche Box2D shapes
dyn4j offers all the primitives that Box2d offers, plus more. dyn4j is missing the GearJoint from Box2d though.