[Ipopt] Linear solver comparison

[Jonathan Hogg]

Pascal,

I'm actually in the middle of doing such a comparison myself. The 
results seem to largely depend on the characteristics of the problems 
involved.

First note that MA57 has a scaling built in that is applied on every 
iteration. It can be disabled by a setting in ipopt.opt. This scaling 
improves reliability (i.e. more problems converge to a solution) at the 
price of performance overheads (that are really only significant on 
small problems).

I also have an interface not yet publically available to our new code 
HSL_MA86, designed to do 'proper' pivoting on multicore machines. (This 
is currently sitting on my stack of things to tidy up and release, 
probably sometime in May).

To summarise some VERY preliminary results:

* For small and well behaved problems MA27 or MA57 without scaling is best.
* For small and badly behaved problems MA57 with scaling is best.
* For larger problems (where factoring a single system takes more than a 
second, say) HSL_MA86 or PARDISO is significantly better than MA27 or 
MA57 - generally because speedups available are significant on these 
problems, but much more limited on others.
* On some very small problems PARDISO seems to have a large fixed 
overhead for some reason, taking almost a second whereas all other 
solvers solve the problems in less than 0.1s. I haven't gotten to the 
bottom of this, and it may well be a mistake on my part.

Most solvers seems to suffer from a small handful of problems where for 
some reason they do considerably worse than other solvers.

I don't know if that helps you at all?

Jonathan Hogg
STFC

On 19/04/11 13:17, Cote, Pascal (RTA) wrote:
> Hi everyone,
>
> I recently use Ipopt for my project. I try ma27 solver and its work very
> well. I also try ma57 with metis but I observe that the results are
> worst (more slow and worst objective function value). I also buy a
> license for pardiso and it is the same. In fact pardiso is worst than
> ma57 (with openmp on a quad core).
>
> Did you know if it exist some papers or reports that present comparison
> results with Ipopt using different linear solver because it seems very
> strange to me why I observed this results?
>
> Pascal
>
>
> Avis:
> Ce message et toute pièce jointe sont la propriété de Rio Tinto et sont
> destinés seulement aux personnes ou à l'entité à qui le message est
> adressé. Si vous avez reçu ce message par erreur, veuillez le détruire
> et en aviser l'expéditeur par courriel. Si vous n'êtes pas le
> destinataire du message, vous n'êtes pas autorisé à utiliser, à copier
> ou à divulguer le contenu du message ou ses pièces jointes en tout ou en
> partie.
> Notice:
> This message and any attachments are the property of Rio Tinto and are
> intended solely for the named recipients or entity to whom this message
> is addressed. If you have received this message in error please inform
> the sender via e-mail and destroy the message. If you are not the
> intended recipient you are not allowed to use, copy or disclose the
> contents or attachments in whole or in part.
>
>
>
> _______________________________________________
> Ipopt mailing list
> Ipopt at list.coin-or.org
> http://list.coin-or.org/mailman/listinfo/ipopt

-- 
Scanned by iCritical.