[Ipopt] IPOPT, Pardiso, and GPUs

[Olaf.Schenk at unibas.ch]

Hi,

We have published a paper two years ago on the use of GPU BLAS within  
IPOPT/PARDISO:

"O. Schenk, M. Christen, and H. Burkhart, Algorithmic Performance  
Studies on Graphics Processing Units, J. Parallel Distrib. Comput. 68  
(2008) 1360-1369"

Replacing the BLAS with CUDA BLAS is not enough in order to get a  
sufficient performance boost. It also needs a special configuration of  
PARDISO and this version not yet available at www.pardiso-project.org.

As an alternative you can use the new iterative solver in  
IPOPT/PARDISO. The only option that you have to add to the file  
ipopt.opt is

pardiso_iterative yes

Let me know if it is working for you.

Regards,
Olaf



Quoting btoth at physics.ucsd.edu:

> Has anyone had experience with running IPOPT with Pardiso on GPUs?  Since
> I assume that only the linear solver aspects of the IPOPT algorithm would
> be running on the GPU, does this solely become a problem of configuring
> IPOPT with CUDA BLAS and LaPack libraries?  Would Pardiso itself need to
> be configured differently (A question for the Pardiso board)?  Or would
> there be fundamental changes necessary to underlying IPOPT code?
>
> I assume that this is the direction that optimization, as with other forms
> of scientific computing, is headed, but I have not found much literature
> or direction specifically for IPOPT.
>
> Thanks.
>
> _______________________________________________
> Ipopt mailing list
> Ipopt at list.coin-or.org
> http://list.coin-or.org/mailman/listinfo/ipopt
>



----------------------------------------------------------------
This message was sent using IMP, the Internet Messaging Program.