[Couenne] MUMPS memory errors when running Couenne in GAMS

[Stefan Vigerske]

Hi,

> Thanks for your help. I was making a dumb mistake on the options files
> when running locally, I wasn't putting them in the default GAMS project
> directory so it was indeed ignoring my changes to mumps_mem_percent. I'm
> continuing to do more tests locally but it takes several hours to run each
> instance. The NEOS server was much faster but the GAMS run via the BARON
> page didn't seem to recognize Couenne options files, whether I called it
> couenne.opt or baron.opt on upload. I wasn't seeing the "List of user-set
> options:" display in those runs.

Hmm, maybe the NEOS server restricts to baron.opt if it thinks that one 
runs BARON.

> GAMS/CONVERT is mighty useful, thanks for that tip! Looks like the NEOS
> server's installation of Ipopt via AMPL is using Ma27, but via GAMS is
> using MUMPS. The memory error when I run Couenne is also happening when I
> use Couenne via AMPL on NEOS - my last page of node status messages is
> interspersed with these errors (in this order, one of each, but with
> several iteration messages between them):
>
> Error: terminate called after throwing an instance of 'std::bad_alloc'
> Error:   what():  St9bad_alloc
> Error: error running couenne:
> Error: 	termination code 6
> <BREAK>
>
> I guess I'll try increasing ma27_la_init_factor, ma27_liw_init_factor,
> and/or ma27 _meminc_factor, unless anyone's got any better ideas.

I think you should also think about reformulating your model. It may 
work a bit better by parameter tuning, but if you could make the hessian 
more sparse, that may help more.

> I was able to get the source for Ma57, but I'm running Windows here so I
> didn't get very far compiling it. I tried make in cygwin using gcc-g77,
> but got a few "linker input file unused because linking not done",
> "Invalid token at (2) in expression or subexpression at (1)", and
> "Reference to unimplemented intrinsic 'EPSILON'/'TINY'/'HUGE'/'RADIX' at
> (^) (assumed EXTERNAL)". I think I saw two versions of Ma57, one called
> HSL_Ma57 - I'll try whichever one I didn't download yet, see if that
> behaves any differently.
>
> Still quite confused why these runs are taking about 10x as long to get to
> the same number of iterations when I do them locally vs on the NEOS
> server. The server was only about 1.5x faster than my machine when I was
> trying BARON on the same model. Maybe it's a number of cores /
> parallelization thing?

Current Couenne does not do any parallelization stuff.
Maybe the GAMS and thus Couenne versions differ, and the one on the NEOS 
server is more lucky on your instance than the one you run locally.

Stefan

> Anyway, would appreciate any advice on this memory hangup from you or others.
>
> Thanks,
> Tony
>


-- 
Stefan Vigerske
Humboldt University Berlin, Numerical Mathematics
http://www.math.hu-berlin.de/~stefan