[Trilinos-Users] Sparse matrix and Anasazi Solvers,
Mike Atambo
mikeat4999 at gmail.com
Thu Feb 5 11:50:22 MST 2015
Alicia,
Im still trying to figure out what exactly im facing. Maybe what you
described be the issue, but at this
point i would prefer to try out what you have suggested, if it works, then
it would be certain thats what im facing,
if not, then i can rule that out.
Please let me know how to set the krylov-shur solver as you suggested,
here are the parameters i pass to the anasaziPL for Krylov-Shur at the
moment:
const double tol = 1.0e-8;
const int maxIters = 500;
anasaziPL.set ("Which", "SR");
anasaziPL.set ("Block Size",1 );
//anasaziPL.set ("Maximum Iterations", maxIters);
anasaziPL.set ("Convergence Tolerance", tol);
anasaziPL.set ("Full Ortho", true);
anasaziPL.set ("Use Locking", true);
Mike.
On Thu, Feb 5, 2015 at 5:47 PM, Alicia Klinvex <aklinvex at purdue.edu> wrote:
> Hi Mike,
>
> If ARPACK is working for you and block Krylov-Schur is not, I suspect that
> one is using a spectral transformation and the other is not.
>
> I don't know much about scipy's eigs, but I assume it is similar to
> Matlab's eigs (which also uses ARPACK). When you ask Matlab's eigs to find
> the smallest eigenvalues of a matrix, it will automatically transform the
> problem of finding the smallest eigenpairs of A x = \lambda x into the
> equivalent problem of finding the largest eigenpairs of A^{-1} x =
> \lambda^{-1}x. Anasazi's Krylov Schur does not do that unless you
> explicitly tell it to. Could that be the problem you're having? If so, we
> can discuss how to tell Krylov-Schur that's what you want to do.
>
> Best wishes,
> Alicia Klinvex
>
> On Thu, Feb 5, 2015 at 11:21 AM, Mike Atambo <mikeat4999 at gmail.com> wrote:
>
>> Hi,
>> Im unable to get matrices, ( say a sparse matrix with
>> 885500 non zeros out of about 2822796900 ) to converge
>> with the block davison or the krylov solver,
>>
>> I had a look at the condition number reported by the 'ILUT'
>> preconditioner and it seems to grow rather quickly for my problems, ( 9x9
>> with 39 non zeros has condition number= 10, and at 25*25 and 100 non
>> zeros the condition number is at 250).
>>
>> Im able to solve all these examples here using scipy.sparse.eigs (it
>> uses ARPACK), so there is probably something that im not doing right.
>>
>> The block-davidson solver with 'ILUT' pre-conditioner work for the
>> small systems above, but beyond 300*300 matrix size, this is not
>> longer true.
>>
>> At the moment im following these steps.
>>
>> 1. Create and initialize a matrix,
>> 2. Create an anasazi basic eigenproblem,
>> 3. Create a solver (either block davidson or krylov-shur)
>> 4. Get solution,
>>
>> Id like to find out whether there is a way to tell why the system im
>> working on seems
>> to be unable to converge. Any help would be appreciated.
>>
>> Mike
>>
>>
>> --
>> M. O. Atambo
>> mikeat4999 at gmail.com
>> matambo at ictp.it
>> Ext .139
>>
>>
>>
>> _______________________________________________
>> Trilinos-Users mailing list
>> Trilinos-Users at software.sandia.gov
>> https://software.sandia.gov/mailman/listinfo/trilinos-users
>>
>>
>
--
M. O. Atambo
mikeat4999 at gmail.com
matambo at ictp.it
Ext .139
Room 209.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://software.sandia.gov/pipermail/trilinos-users/attachments/20150205/2be2c4ec/attachment.html>
More information about the Trilinos-Users
mailing list