[Ipopt] question about ipopt with ma86 or pardiso

[Jonathan Hogg]

My rule of thumb on this is that if your ma27 takes more than about
0.5s/iteration it is probably worth trying a shared memory parallel
solver. If you're able to dump the matrices out (ma97_dump_mat option,
probably needs a line uncommented in the Ma97SolverInterface.cpp file)
I'm happy to run them on our machines and advise.

Other thing to note is that sometimes the scalings used by default in
non-MA27 solvers can be cause of slowness - see if there are any solver
specific options to disable them if they aren't needed.

I'm actually currently looking at algorithms for making small problems
go much faster, but we've not got any results yet (basically driving
parallelism and vectorization into the sparse blas-2 like operations
that are significant in small problems).

Jonathan.

On 08/07/15 21:04, Ian Washington wrote:
> Hi All,
> 
> Quick question.
> 
> At what point in the problem size (variables/constraints) does using a
> parallel shared-memory linear solver like ma86, ma97 or pardiso become
> beneficial in terms of speedup.
> 
> I am trying to do some timings on a sparse discretized optimal control
> problem with about 30000 variables and 20000 equalities (i.e. 10000
> dof). Currently, ma86 and pardiso require more time than the serial ma27
> solver for the same amount of iterations. I am using openmp 3.1 with
> gcc-4.7.3 on linux 64bit.
> 
> I am guessing my problem is just too small at present.
> 
> Anyone have any thoughts?
> 
> Thanks,
> Ian.
> _______________________________________________
> Ipopt mailing list
> Ipopt at list.coin-or.org
> http://list.coin-or.org/mailman/listinfo/ipopt
>