<div class="gmail_quote">On Sun, Nov 6, 2011 at 06:28, robert <span dir="ltr"><<a href="mailto:robert.bodner@unil.ch">robert.bodner@unil.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=":i1">Do you have experience with lumping of the mass matrix?<br>
<br>
The following options semm to work fine if I don't use lumping:<br>
<br>
mpiexec -np 4 ./ThermoPaine3d-opt -ksp_type fgmres -pc_type ksp<br>
-ksp_ksp_type cg -ksp_pc_type jacobi -ksp_monitor_true_residual<br>
-ksp_converged_reason<br></div></blockquote><div><br></div><div>Is your matrix actually symmetric (including boundary conditions)? This is running a complete solve as a preconditioner.</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex;">
<div id=":i1">
<br>
(any comments to the options???)<br>
<br>
However, for heat diffusion, especially for non-steady starting<br>
conditions, oscillations seem to be a common problem (AFAIK).<br>
If I apply lumping of the mass matrix - which means I put the row-sum of<br>
the mass matrix on the main diagonal - the solver doesn't converge any<br>
more. Before it converged after 2 to 5 iterations, now I was running<br>
several hundreds.</div></blockquote></div><br><div>I suspect a mistake in assembly causing the matrix to be singular or indefinite. I would try a very small problem (e.g. < 1000 dofs) and run with -pc_type svd -pc_svd_monitor. A preconditioned residual of exactly 0 is likely a mistake (and the true residual is 1e17). Is this coming out of FGMRES? From your convergence history, it looks like you have set things up in your code to use a nonzero initial guess?</div>