simtk.org

Hello,
I have a MATLAB code that takes 12 mins to execute. I have to reduce it to less than a minute. However, I am not sure which parallel computing technique is compatible with OpenSim.
Since the integration in the manager takes 80% of the execution time, I was thinking if I can pass this part of code to execute on multiple threads/workers. The part of the code is mentioned below

manager = Manager(main_model); % Manager will be the forward dynamics integrator
state.setTime(0);

manager.initialize(state); % State of the model is given to the manager for update
state = manager.integrate(timed); % Integrate for the time duration of 'timed'
main_model.realizeAcceleration(state);

If anyone has tried a similar problem, kindly help.

Regards,
Shivangi

Hi Shivangi,

I had looked to do something similar. From my limited understanding (maybe another community member with a better understand can clarify) the forward dynamic calculations are performed on a single core, and there does not seem to be a way of accessing how the calculations are distributed across cores. So I don't think what you want to do is possible, at least not on the official release of OpenSim.

Maybe this post could help: viewtopicPhpbb.php?f=91&t=15042&p=43684&start=0&view=

However, if you are running many forward dynamic simulations, you can instead parallise the simulations. Each forward dynamic simulation is computed on a different core, this doesn't speed up an individual simulation, but will reduce the total amount of time you spend waiting for results.

Best,
Samuele

Thanks Samuele, well said.

I'd just add that forward integration is naturally sequential since it evolves over time so it'd be hard to make a general scheme that integrates in parallel. However in cases of algorithms that run multiple forward integrations (e.g. in an optimization) this becomes an easier problem to tackle.

Best regards,
-Ayman

Thank you, Samuele and Ayman for your replies. Much appreciate your help.

Regards,
Shivangi