How many workers for fmincon with UseParallel?

I'm solving a biggish problem where the function takes a while to compute and want to throw as many workers as possible on the problem.

1. Is the reason to set UseParallel to true when solving a problem with fmincon that it can potentially speed up the evaluation of the problem's gradient?
2. Say my problem has 100 variables. Could I then theoretically improve performance with up to 100 workers? Or does fmincon never use that many parallel processes?

[Bruno Luong]

"Is the reason to set UseParallel to true when solving a problem with fmincon that it can potentially speed up the evaluation of the problem's gradient?"

Only if you don't provide the gradient then fmincon estimates the gradient using finite differnce, each component of the gradient can be estimated in parallel and UseParallel is set to TRUE.

Answer 1

[Matt J]

Is the reason to set UseParallel to true when solving a problem with fmincon that it can potentially speed up the evaluation of the problem's gradient?

Yes.

Say my problem has 100 variables. Could I then theoretically improve performance with up to 100 workers? Or does fmincon never use that many parallel processes?

Theoretically. The number of workers can be set by you using parpool, if you have the Parallel Computing Toolbox, or fmincon will call parpool with default settings. How many workers you can create in a pool depends on your Matlab version and the resources available to your computer.

[Fredrik P]

I'm trying to run the problem on a cluster. There I could have lots and lots and lots of workers :-)

[Matt J]

If it's a cluster, you'll also have to consider the time for transmitting data to the cluster nodes. If the work done by each node is much less than the communication time, you may not benefit much.

[Fredrik P]

Good point, but I don’t think that should be a problem here…