[Ipopt] Problems with Pardiso solver in Ipopt

Fabian Wein fabian.wein at lse.e-technik.uni-erlangen.de
Mon Dec 15 04:31:24 EST 2008


Rui Huang wrote:
> Thanks all for your replying.
> I checked the output messages from pardiso and monitored with htop, it 
> seems clear that it is not dong parallel computing. One of my 4 
> processors is always at 100% usage and the others are idol.
> And the statistics of pardiso output says:
> Parallel Direct Factorization with #procesors : 1.
> 
> Do we really have to compiled it with parallel version of Lapack, as 
> suspected by Rony?

Olaf Schenk suggested me to use serial BLAS/LAPACK.

But anyway - if it goes for performance you need a fast BLAS/LAPACK.
I suggest to use Intel_MKL (NOT MKL Pardiso, that doesn't work)
or to use gotoBLAS. The is also AMDs ACML but gotoBLAS was faster
in our scenarios (even faster than MKL).

Are you sure you link the gomp and phtread lib?

Normally I suggest to create a callgraph with

valgrind --tool=callgraph <your application>

and check the results with kcachegrind.

But as valgrind simulates no parallel system ... it might be worth a try.

Fabian




More information about the Ipopt mailing list