Re: iterative vs. direct methods for linear systems

On Apr 1, 9:00 pm, s...@xxxxxxxxxxxxxxx (Victor Eijkhout) wrote:
Evgenii Rudnyi <use...@xxxxxxxxx> wrote:
But believe me, if you try modern direct solvers, you
will be surprised.

Can they improve on the N^{5/3} space bound for 3D problems?

Unfortunately I do not not know a theoretical bound for multifrontal
solvers. Sorry for that.

Han uses a crude band solver, so he indeed attains this bound. If other
direct solvers only improve this by a factor, I wouldn't consider it
worth his while to rewrite his code.

In my knowledge finite elements codes now use multifrontal solvers.
For example in ANSYS there is a frontal solver but it is much slower
as compared with multifrontal. Nowadays no one uses the frontal solver
in ANSYS anymore.

Also I am not quite sure if I understand why it is necesssery to
rewrite a solver. It would be much simpler just to call some already
developed solver. I can recommend MUMPS.

Evgenii http://MatrixProgramming.com
.