1

I am considering working on a rigid body parallel physics engine, with MPI, as my own project. Have some experience with serial engines. So far, I couldnt find any existing projects of that type, does enyone knows about such things?

I know that MPI is not the best choice for real time physics, lots of time will be lost on duplicating data send/rcv between nodes. I planning on running it on non shared memory machine though.

Does it sounds like something worth doing? Thanks

kirbo
  • 1,707
  • 5
  • 26
  • 32

1 Answers1

2

If you can fit the job into the available memory then a GPU (either CUDA or OpenCL) may be the way to go.

MPI isn't really made for responsiveness, even with exotic low latency interconnects the message passing is slow. Unless it;'s changed since my day a lot of the code is polling nodes and waiting for a reply which doesn't help.

Martin Beckett
  • 94,801
  • 28
  • 188
  • 263
  • ++ Right. That's my experience. MPI is best if you can farm out relatively long-running work to multiple CPUs and then collect the results. That's how we use it. – Mike Dunlavey Feb 12 '11 at 00:45