[Coin-ipopt] Can you make a completely open source build?

[kevin.c.furman at exxonmobil.com]

Unfortunately, the old Harwell Archive routines are only free for
non-commercial purposes.  My understanding is that industry and national
labs probably need to purchase them to use them, even if for research
purposes.

I think PARDISO has a similar licensing scheme and similarly good
performance compared to MA57.  Since PARDISO is offered as part of the
Intel MKL, it might be a good alternative choice since I think it would be
more commonly available.  MKL probably also has alternatives to most of the
other Harwell routines used by IPOPT as well.

I believe IBM Watson Research center actually had their own solver for
sparse symmetric indefinite linear systems as as part of the WSMP (Watson
Sparse Matrix Package).  Perhaps IBM would consider open sourcing that
package.

You might find the following recent reports comparing these kinds of code
interesting:
ftp://ftp.numerical.rl.ac.uk/pub/reports/ghsRAL200505.pdf
ftp://ftp.numerical.rl.ac.uk/pub/reports/ghsNAGIR20051.pdf

Regards,

Kevin



                                                                                                                                        
                      Carl Damon Laird                                                                                                  
                      <claird at andrew.cmu.ed          To:      kevin.c.furman at exxonmobil.com                                             
                      u>                             cc:      coin-ipopt at list.coin-or.org                                               
                                                     Subject: Re: [Coin-ipopt] Can you make a completely open source build?             
                                                                                                                                        
                      07/05/05 01:32 PM                                                                                                 
                                                                                                                                        
                                                                                                                                        



As it stands right now, there is no tested, open source substitution for
the Harwell routines. We are looking at some open linear solvers, but we
do not have anything ready for prime time as yet. As I understand, the
Harwell routines are free for "research" purposes - please consult the
Harwell website for a more precise definition.

Of course, we agree that there *should* be a way to build IPOPT completely
from open source, however, this is not a trivial task. The IPOPT algorithm
requires a sparse, indefinite linear solver (symmetric) that can provide
the inertia of the linear system (the number of positive and negative
eigenvalues). So far, we have found VERY few open source linear solvers
that can provide this functionality.

If anyone on the list has knowledge of some alternatives that we could
consider, please let us know :)

Cheers,

Carl.

On Tue, 5 Jul 2005 kevin.c.furman at exxonmobil.com wrote:

> Are there any libraries that can be substituted for the Harwell routines?
>
> Is there a way to build IPOPT completely from open source?  If not,
> shouldn't there be one?
>
> Regards,
>
> Kevin Furman
>
> _______________________________________________
> Coin-ipopt mailing list
> Coin-ipopt at list.coin-or.org
> http://list.coin-or.org/mailman/listinfo/coin-ipopt
>
>