[Coin-ipopt] MA27 error message

[Andreas Waechter]

Hi Zhiwen,

That is a good question, it seems that a number of users run into this.

The output means that MA27 requires a lot of fill-in during the 
factorization, which can happen when the matrix is very badly scaled.  It 
therefore requires more memory than it already has, and Ipopt prints this 
message, indicating, how much memory it is now allocating for MA27.

There is an option to equilibriate the linear system give to MA27, using 
MC19 (which must be included when you compile Ipopt to be able to use 
this).  In the current C++ version of Ipopt, this scaling is done only 
when iterative refinement fails.  However, you can tell Ipopt to always 
use MC19 to scale the linear system, by setting the option

linear_scaling_on_demand no

Maybe we should make this the default...  (I didn't do this, because in my 
tests, it wasn't leading to considerable more robust results, but using 
MC19 makes the computation quite a bit slower...)  We might implement some 
other heuristics later, but for now, please try if the above option helps. 
For some other users, using thie option avoided the huge fill-in in MA27.

Of course, you can also try to use some other linear solver (there are new 
interfaces in the most recent release of Ipopt).

Regards,

Andreas

On Mon, 10 Apr 2006, Zhiwen Chong wrote:

> Hello all,
>
> Forgive me for asking what is possibly a very naïve question, but what does 
> this mean? (I get it a lot)
>
>
> MA27BD returned iflag=-4 and requires more memory.
>  Increase liw from 1607065 to 16070650 and la from   1663110 to 16980600 and 
> factorize again.
>
>
> Ipopt 3.1.0 does eventually converge to the optimal solution, but I'm worried 
> if this error is indicative of something nasty in my formulation? (some 
> badness in the linear system, perhaps?) Or is it just a benign warning about 
> needing to adjust the initial memory allocation for MA27?
>
> My problem is fairly big:
>
> 32823 variables:
>         9488 nonlinear variables
>         23335 linear variables
> 32837 constraints; 93126 linear nonzeros
>         4850 nonlinear constraints
>         27987 linear constraints
> 1 linear objective; 1 nonzero.
>
> I tried to google 'ma27 and iflag' but this is about the most useful link I 
> could find.
> http://tinyurl.com/oyujz
>
> I'd appreciate any comments.
>
> Thanks!
>
> Zhiwen
>
> _______________________________________________
> Coin-ipopt mailing list
> Coin-ipopt at list.coin-or.org
> http://list.coin-or.org/mailman/listinfo/coin-ipopt
>
>