[Ipopt] IPOPT and PARDISO

[Olaf.Schenk at unibas.ch]

Hi Johan,

A reliable computation of the inertia is very important in IPOPT and  
we worked extensively with Andreas in order to make PARDISO a reliable  
and efficient linear solver option in IPOPT.

The paper has been published in

O. Schenk, A. Wächter, and M. Hagemann, Matching-based Preprocessing  
Algorithms to the Solution of Saddle-Point Problems in Large-Scale  
Nonconvex Interior-Point Optimization. Journal of Computational  
Optimization and Applications, pp. 321-341, Volume 36, Numbers 2-3 /  
April, 2007.

http://fgb.informatik.unibas.ch/people/oschenk/publications/references/schenk-coap-07.pdf

This version is based on advanced weighted graph matchings and some  
additional block scalings in the Jacobian. It is not included in the  
MKL, but you can download this new version at  
http://www.pardiso-project.org/.

Olaf


> Hi Johan,
>
> Thanks you for your email.  In the future, for Ipopt related issues
> please write to the Ipopt mailing list; others might be able to answer
> more quickly or better than me, and the answer might also be of
> interest to someone else.  (I'm copying my answer to the mailing list
> now.)
>
>> [...] My question is about the status of IPOPT and PARDISO, and in   
>> particular the version of PARDISO that is included in MKL. I have   
>> read in the Ipopt documentation that the MKL version is not   
>> suitable for use with Ipopt, but that statement is contradicted by   
>> the the ticket #88 in the Ipopt trac. So, I am not quite sure about  
>>  the situation. Part of the reason why I ask is that Modelon AB  
>> runs  JModelica.org as a dual licensing project similar to MySQL or  
>>  TrollTech (Qt). If I understand correctly, the MKL license allows   
>> for distribution of binaries under commercial licenses, and this   
>> would then be a possible match to bundle PARDISO and IPOPT with the  
>>  JModelica.org platform. If the version of PARDISO in MKL is ready   
>> to go with IPOPT then I guess i would be straight forward, but more  
>>  tricky if this is not the case.
>
> The version of Pardiso that is in the MKL does not have the features
> that are required by Ipopt for nonconvex optimization (it does not
> compute the inertia of the linear system).  The ticket you are
> referring to has the changes that you can make if you KNOW that your
> problem is convex (or if you want to just hope that you are lucky,
> ignoring possible nonconvexity of your problem).  So, unless you are
> solving convex problems, I guess this is not the answer you had hoped
> for.
>
> One thing that might be helpful is that Stefan Vigerske wrote a
> "library loader" that allows Ipopt to load Pardiso at runtime, even if
> Ipopt has not been configured and compiled with Pardiso.  If you choose
> "linear_solver=pardiso" and Ipopt wasn't linked with Pardiso, it will
> attempt to find "libpardiso.so" and link to it.
>
>> As a side note, we will be shipping binaries of IPOPT compiled with  
>>  MUMPS för Windows (mingw) which may be of some general interest. I  
>>  had a look at the binaries distributions of IPOPT available, but I  
>>  did not find MinGW amongst the provided alternatives.
>
> Sure it would be great if there are binaries of Ipopt around!  If you make
> binaries available for download on a website, we can add a link to that
> from the Ipopt Wiki page.
>
> Regards,
>
> Andreas



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