<div class="gmail_quote">On Tue, Nov 1, 2011 at 09:21, Matthew Knepley <span dir="ltr"><<a href="mailto:knepley@gmail.com">knepley@gmail.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex;">
<div>It could be that:</div><div><br></div><div> a) Your system is really ill-conditioned, and the PC is not that good</div>
<div><br></div><div> b) Your system has a null space, and the initial guess had no component in it</div><div><br></div><div> c) Your Jacobian is not actually the Jacobian of the system, so the PC is not good</div><div>
<br>
</div><div> d) The FD approximation to the Jacobian action is not good</div><div><br></div><div>When you have so many different possible problems, the right thing to do is simplify the system until you can narrow down the cause</div>
<div><br></div><div> a) Run in serial</div><div><br></div><div> b) Shrink the problem</div><div><br></div><div> c) Fully form the Jacobian using -snes_fd</div><div><br></div><div> d) Run with -pc_type lu</div></blockquote>
</div><br><div>We keep a pretty comprehensive list here, I should have linked it:</div><div><br></div><div><a href="http://www.mcs.anl.gov/petsc/petsc-as/documentation/faq.html#kspdiverged">http://www.mcs.anl.gov/petsc/petsc-as/documentation/faq.html#kspdiverged</a></div>