[Trilinos-Users] Large Generalized Symmetric Complex Eigenvalue Problem: From ARPACK to Anasazi?
Klaus Zimmermann
klaus.zimmermann at physik.uni-freiburg.de
Fri Aug 6 16:43:02 MDT 2010
Hi Heidi,
sorry, it is late, over here at least ;)
We are using pardiso with arpack. This isn't to much by design though, I
think.
When I joined the project last December we were working with this all
handwritten
Fortran code that descended through my phd advisor from his phd advisor.
It was a handwritten lanczos with an equally handwritten cholesky solver...
We then moved on to the arpack stuff with pardiso.
Since we now really want to increase the size of the system I think it
makes sense to
look at iterative solvers, because if we increase the size by, say, four
orders of
magnitude the out of core alone might not save us.
Thanks for your help,
Klaus
On 07.08.10 00:35, Thornquist, Heidi K wrote:
> Hi Klaus,
>
> That should be no problem, I guess I was just wondering what solver
> you were using for the linear system with ARPACK.
> Did you use a direct solver, like Pardiso, or an iterative one, like
> GMRES preconditioned by ILU? That will help us try to get you
> set up with something similar to what you have now, or at least give
> you an ETA on when that will be available.
>
> Thanks,
> Heidi
>
>
> On 8/6/10 4:26 PM, "Klaus Zimmermann"
> <klaus.zimmermann at physik.uni-freiburg.de> wrote:
>
> Heidi, Mike,
>
> thanks your response time is incredible!
> You are exactly right with the spectral transformation. Should
> have mentioned.
> We are doing shift-and-invert. I guess this is still applicable?
>
> Thanks again,
> Klaus
>
> On 07.08.10 00:23, Thornquist, Heidi K wrote:
>
> Re: [Trilinos-Users] Large Generalized Symmetric Complex
> Eigenvalue Problem: From ARPACK to Anasazi? Hi Klaus,
>
> For now we'd suggest Anasazi (BKS), Tpetra, Belos (GMRES),
> Tifpack (for preconditioning). Since Tifpack is a
> domain decomposition incomplete LU, it won't take nearly as
> much memory as Pardiso. This is because it
> is not performing an full factorization of the matrix, just an
> incomplete LU factorization of the parts of the domain that
> are on each processor. When you are using ARPACK, I'm
> assuming that you are using a spectral transformation,
> which is why you need the linear solver. If so, which linear
> solver did you use with ARPACK?
>
> Thanks,
> Heidi
>
>
> On 8/6/10 4:13 PM, "Klaus Zimmermann"
> <klaus.zimmermann at physik.uni-freiburg.de>
> <mailto:klaus.zimmermann at physik.uni-freiburg.de> wrote:
>
>
> Mike,
>
> thanks for your quick reply!
> I am glad to hear of these improvements. If I understand
> you correctly
> for now you'd suggest
> Anasazi, Tpetra, Belos, right?
> Then there is the question of the preconditioner: In our
> legacy code we
> did a LU factorization
> (via pardiso) which quickly led to storage requirements on
> the order of
> 50 gb. This is unfortunately
> prohibitive for the 16 gb of ram/node on the grid. We used
> to use out of
> core functionality for the
> solver.
> Is such out of core stuff available for Belos as well, or
> will Tifpack
> reduce the amount of memory
> needed so much I don't have to worry?
> For reference the matrices we deal with right now are on
> the order of
> n=1e8, complex symmetric
> with about 0.5% non zero elements.
>
> Klaus
>
> On 06.08.10 23:53, Heroux, Michael A wrote:
> > Klaus,
> >
> > The software stack that supports complex is growing
> quickly. Belos already
> > supports complex for Krylov methods for linear systems.
> There is a new
> > package called Tifpack that supports complex to do domain
> decomposition with
> > ILU and related preconditioners. So, using Trilinos 10.4
> (the current
> > release), there is a better answer than Komplex for
> iterative linear
> > solvers.
> >
> > In the next version of Trilinos (Oct 2010), we plan to
> have Amesos2 and KLU
> > 2, which support complex. How much of the linking
> support to complex
> > versions of MUMPS, UMFPACK, etc. will be done by October
> is not clear right
> > now, but we are moving in the right direction.
> >
> > BTW, all of the support for complex comes from using
> templated scalar types
> >
> > Mike
> >
> > On 8/6/10 4:41 PM, "Klaus Zimmermann"
> > <klaus.zimmermann at physik.uni-freiburg.de>
> <mailto:klaus.zimmermann at physik.uni-freiburg.de> wrote:
> >
> >
> >> Hi All,
> >>
> >> for my ab initio quantum calculations I have to find
> some eigenvalues
> >> for large, sparse complex symmetric generalized
> eigenvalue problems.
> >> I would really like to move from our legacy ARPACK
> fortran code to a
> >> more modern code base
> >> in order to make use of our local MPI based grid. I am a
> bit overwhelmed
> >> by the choices. But I now think
> >> I boiled my possibilities down to Anasazi or SLEPc. I'd
> prefer to go the
> >> C++ way.
> >> However I understand the support for complex systems in
> the linear
> >> solver part is limited.
> >>
> >> What would be the best way to go about tackling this
> kind of problem?
> >> Right now I am thinking: Anasazi<Tpetra,
> BlockKrylovSchur> +
> >> (Komplex+Amesos+MUMPS).
> >> Is this reasonable?
> >>
> >> Thanks for any hints in advance!
> >>
> >> Regards,
> >> Klaus Zimmermann
> >>
> >> _______________________________________________
> >> Trilinos-Users mailing list
> >> Trilinos-Users at software.sandia.gov
> >> http://software.sandia.gov/mailman/listinfo/trilinos-users
> >>
> >
> >
>
>
> _______________________________________________
> Trilinos-Users mailing list
> Trilinos-Users at software.sandia.gov
> http://software.sandia.gov/mailman/listinfo/trilinos-users
>
>
>
>
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: https://software.sandia.gov/pipermail/trilinos-users/attachments/20100807/af878c8d/attachment.html
More information about the Trilinos-Users
mailing list