<div class="gmail_quote">On Fri, Nov 25, 2011 at 05:45, Dominik Szczerba <span dir="ltr"><<a href="mailto:dominik@itis.ethz.ch">dominik@itis.ethz.ch</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex;">
<div id=":560">I tried -ksp_max_it 1 but then I get divergence<br>
because residue reduction is somehow very poor (?), which does not<br>
make sense, because I expect the system be solved down to the epsilon.<br></div></blockquote><div><br></div><div>This can happen if the linear system is very ill conditioned. You can run the direct solve inside -ksp_type richardson for a more classical "iterative refinement". Using a Krylov method is generally better. Because of the way GMRES preconditioning works, converging after one iteration still means two preconditioner applications (so much cheaper than factorization that you probably don't notice it). In contrast, one iteration of FGMRES only does one preconditioner application.</div>
</div>