Re: [ublas] Solver recommendation

From: Jonas (jonas.hagmar_at_[hidden])
Date: 2008-03-06 09:41:31

On Thu, Mar 6, 2008 at 3:11 PM, Jens Seidel <jensseidel_at_[hidden]> wrote:
> On Thu, Mar 06, 2008 at 02:32:50PM +0100, Jonas wrote:
>  > Which library/solver (that has uBLAS bindings, preferrably) would you
>  > recommend for solving a very sparse, real system Ax = b with size on
>  > the order of 10^5? My initial guess from monitoring the list would be
>  > umfpack. Would that be a good candidate?
>
>  Mmh, this is sufficient large to require special algorithms. Please
>  explain in more detail the properties of your matrix. Is it symmetric,
>  positive definite, an M matrix, ...? Does it result from
>  discretisation of a (partial) differential equation, which one?
>  Does there exist a hierarchy of matrices?
>
>  It is nearly impossible for me to answer your question without these
>  information!
>
>  You could also look at Dolfin (http://www.fenics.org/wiki/Documentation)
>  which contains some simple standard solvers and has also interfaces to
>  more difficult ones (such as Algebraic Multigrid).
>
>  Do you have a typical matrix somewhere available for download (in text
>  format)? Maybe I could test my Algebraic Multigrid solver (no publically
>  available, but I could give you some references) with it?
>
>  Jens
>  _______________________________________________
>  ublas mailing list
>  ublas_at_[hidden]
>  http://lists.boost.org/mailman/listinfo.cgi/ublas
>

The matrix is not symmetric, is real, and has typically less than ten
non-zero elements per row. It is not banded, but it would probably be
possible to perform a permutation that makes it banded with relatively
low cost. After speaking to a colleague of mine who is in the HPC
field, he recommended me to look into spooles, PETSc and umfpack in
that order. Which of these are there uBLAS bindings for?

Best Regards,
Jonas Hagmar