[Ipopt] IPOPT, MUMPS and MPI

[Stefan Vigerske]

Hi,

my personal opinion is to keep the main code running in parallel and to 
let Ipopt run on a single processor. If you have multiple cores on a 
single processor, you can use multithreaded linear algebra packages 
(MKL, GOTO) to get a good speedup.
Using optimized versions of BLAS and LAPACK is a good suggestion anyway.

Replacing MUMPS by a higher performance linear solver like Pardiso or 
one of the HSL routines (free for academics) could also be beneficial. 
These may also allow to be run in parallel on a multicore system.

Using MUMPS with MPI in Ipopt is not trivial, you would have to change 
the build system and code at some places (as you have seen). Even if you 
run Ipopt/Mumps without MPI but link to an application that uses MPI, 
you may have to do some changes to avoid duplicate symbols.
It may be more beneficial to change to a different linear solver, as 
mentioned above.

Stefan

On 10/17/2012 06:52 PM, Karla Telidetzki wrote:
> Hello all,
>
> I have an aerodynamic shape optimization code, written in Fortran, which is
> currently using SNOPT. I am trying to implement IPOPT as well to see if we
> can get improved performance for some of our more difficult problems. The
> code uses MPI and runs in parallel, however the optimization algorithm only
> runs in a single processor. I have gone through the mailing list archives
> and have seen a variety of suggestions on how to implement IPOPT with MPI,
> such as using a different linear solver, installing the parallel version of
> MUMPS and even suggesting changing the MUMPS source code (
> http://list.coin-or.org/pipermail/ipopt/2012-February/002784.html). I was
> wondering if anybody on this list has any suggestions as to the best plan
> of action for my problem (i.e. Running the main code in parallel, but
> running the optimization algorithm in serial in a single processor either
> using MUMPS or a different linear solver).
>
> Thanks for your help.
>
> Karla Telidetzki
>
>
>
> _______________________________________________
> Ipopt mailing list
> Ipopt at list.coin-or.org
> http://list.coin-or.org/mailman/listinfo/ipopt
>