<div class="gmail_quote">On Mon, May 16, 2011 at 14:54, Tian(ICT) <span dir="ltr"><<a href="mailto:rongtian@ncic.ac.cn">rongtian@ncic.ac.cn</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex;">
<div> 12 KSP preconditioned resid norm
5.794625326821e+000 true resid norm 5.490686672936e+006 ||Ae||/||Ax||
7.765003561961e+004<br> 13 KSP preconditioned resid norm
7.239823936251e-001 true resid norm 4.262009749173e+005 ||Ae||/||Ax||
6.027391992073e+003<br> 14 KSP preconditioned resid norm
6.718858761053e-001 true resid norm 7.089916382344e+005 ||Ae||/||Ax||
1.002665590704e+004<br> 15 KSP preconditioned resid norm
6.284315560280e-001 true resid norm 4.441924320726e+005 ||Ae||/||Ax||
6.281829619309e+003<br> 16 KSP preconditioned resid norm
6.861179511523e-011 true resid norm 1.032548577293e+006 ||Ae||/||Ax||
1.460244202260e+004</div></blockquote></div><br><div>Notice how the true residual is blowing up while the GMRES estimate converges. (The language here is not correct, "preconditioned resid norm" in this case is actually the unpreconditioned norm as estimated by FGMRES.) This indicates that the preconditioner is singular which could be due to lack of pivoting. If you run with -pc_type none, it should converge correctly. Do that and see if SNES is converging properly. Also try -pc_type lu -pc_factor_mat_solver_package superlu (or mumps or umfpack) if you have any of these packages installed.</div>
<div><br></div><div>What equations are you solving and with what discretization?</div>