<div class="gmail_quote">On Wed, Dec 28, 2011 at 07:02, Alexander Grayver <span dir="ltr"><<a href="mailto:agrayver@gfz-potsdam.de">agrayver@gfz-potsdam.de</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<div bgcolor="#ffffff" text="#000000">Hi Jed,<br>
If you mean input parameters they are as following:<br>
-ksp_view -ksp_monitor -mat_type mpisbaij -ksp_type preonly -pc_type
cholesky -pc_factor_mat_solver_package pastix -mat_pastix_verbose 2
<br><div class="im">
<br>
<blockquote type="cite">
<div class="gmail_quote">
<div> </div>
<blockquote class="gmail_quote" style="margin:0pt 0pt 0pt 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">
<div>The output is attached.<br>
The matrix is SPD and has condition number of 16, there
should be no problem.<br>
I use petsc-dev.<br>
How could I figure out where is the problem?</div>
</blockquote>
</div>
<br>
<div>This sounds like a small example matrix. Can you
-ksp_view_binary and send us the output file ('binaryoutput').</div>
</blockquote>
<br></div>
Maybe it was too preliminary to say that it is positive definite
since CG diverges with reason DIVERGED_INDEFINITE_MAT,</div></blockquote><div><br></div><div>Oh, I also see in the PaStiX output that it is reporting an error (but just a number). If you run with -ksp_converged_reason, it should say that it diverged. You should always run with -ksp_converged_reason -ksp_monitor_true_residual when investigating whether a solver is converging. From code, call KSPGetConvergedReason().</div>
<div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div bgcolor="#ffffff" text="#000000"> however GMRES
with default parameters converges successfully.<br>
Matrix is ~200000x200000, not so small, but zipped occupies 4.5 Mb.
Shall I send it?</div></blockquote></div><br><div>Try solving with LU instead. It sounds like Cholesky is just breaking down, maybe because it does a slightly different method than the other direct solver packages.</div>