Improving simulation speed in Python

[tsreyas]

I'd like to improve the speed of my simulation, and thought of implementing the multiprocessing module in my simulation to distribute the load among CPUs, but I don't know how to split the simulation into different processes. I've tried including the pool function in the

if __name__ == '__main__':
	main()

part, but it doesn't seem to be working, either.

If it's of any help, the AdvanceTimer profiler tells me that most of the computation load is from the constraint solver. I've tried changing the solvers and solver parameters, but to no avail.

How can I speed up my simulation?

I've attached my current code to this post
code.txt
.

[alxbilger]

Hello @Hellscythe789

From what I can read in your scene, you tried using multithreading by:

• using parallelODESolving and parallelCollisionDetectionAndFreeMotion in the FreeMotionAnimationLoop.
• using ParallelBruteForceBroadPhase and ParallelBVHNarrowPhase to speed up the collision detection.

I warn you that you maybe created a multithreading bottleneck here. ParallelBruteForceBroadPhase is interesting if you have many objects, otherwise, use BruteForceBroadPhase. ParallelBVHNarrowPhase is interesting if you have many elements in your collision models. However, due to parallelCollisionDetectionAndFreeMotion and parallelODESolving, you are already doing jobs on all your cores. So ParallelBVHNarrowPhase will not be efficient in that case. I advise to try different combinations of parameters and always benchmark your simulation.

In any case, the mentioned techniques of multithreading do not have any impact on the constraint solving step, which appears to be the bottleneck of your simulation. There is not much you can do to speed up this process using multithreading. However you can try to:

• Reduce the number of constraints: decrease the resolution of your collision models, or play with the intersection method to filter contacts.
• Usually constraints are solved when a direct linear solver is used. I see you use a CGLinearSolver. Try using a SparseLDLSolver, or a EigenSimplicialLDLT instead.

Note that there is a page dedicated to improving the performances on the documentation website: https://www.sofa-framework.org/community/doc/using-sofa/performances/improve-performances/

Alex

[hugtalbot]

has @alxbilger's advise helped you @Hellscythe789 ?

[tsreyas]

I'm sorry, I was busy and forgot to test those suggestions out. I'll let you know how it goes when I do. @alxbilger @hugtalbot

[tsreyas]

Sorry for the late update.

I replaced the ParallelBruteForceBroadPhase and ParallelBVHNarrowPhase components with the regular versions, replaced CGLinearSolver with SparseLDLSolver, and made my meshes a lot coarser overall.

I'd say 7 fps on contact is a much better result than 1-2 fps. Thank you @alxbilger !

[alxbilger]

Great!
If your simulation involves only 3d dofs, I advise to use the template CompressedRowSparseMatrixMat3x3d with SparseLDLSolver in order to speed up the matrix assembly.