[Trilinos-Users] Ramakrishna

[David M Day]

Ramakrishna,
As noone has gotten to this yet, I will a little something, realizing
that others are likely to have more to say later.
It is reasonable to use an iterative method, particularly for linear
systems with very large numbers of unknowns, even if it is not
completely satisfactory.
Within Belos, the GCRODR family solve a sequence of the linear systems
more efficiently than ordinary GMRES.
Are your linear systems symmetric and or positive definite?  Does the
coefficient matrix ever change, or just the right-hand side?

For direct methods, the key point is to re-use a factorization as long
as the coefficient matrix remains unchanged.
It might help to glace at
Trilinos/packages/amesos/test/Test_MultipleSolves
The distributed memory codes are MUMPS (Fortran 90) and SuperLU_DIST.
Pardiso is shared memory.  DSCPACK is only for symmetric positive
definite linear systems.  For linear systems with modest numbers of
unknowns, parallel sparse direct methods are great.  As the problem size
gradually increases, however, the methods tend to suddenly fail due to
insufficient memory of one type or another (e.g. MPI communication
buffers).
--David Day