Knitro on computing clusters, with many GPUs

[László Sándor]

Hi,

I consider using Knitro (most likely calling it from Maltab), but I wonder about the most efficient way of doing so. What kind of resources can I give Knitro if I have access to a computing cluster with many cores, GPUs, and memory? What *should* I give it? Is Knitro intelligent about parallelizing itself by default, or do I need to change anything? The documentation is silent on GPUs or clusters. Does it mean that they don't help Knitro?

Sorry if I missed the obvious (though I tried the FAQ and the documentation).

Thanks a lot,

Laszlo

PhD candidate in Economics

Harvard University

[KNITRO support team]

Hi Laszlo,

- The current version of KNITRO (7.0) uses a BLAS library like Intel
MKL which takes advantage of multiple cores.
- If you use ktrlink with the parallel toolbox and that you compute
derivatives through finite differences, the evaluation of the function
in each direction will be parallelized.
- KNITRO 7.0 does not run yet on GPUs, however, we could imagine that
the user code needed to evaluate function, constraints, gradients and
hessian is implemented to run on GPUs or on a cluster.

Best regards.

Nicolas Omont