<div class="gmail_quote">Try using a multilevel method for the Poisson solve. If you use DM, you can just do -pc_type mg -pc_mg_levels 4 (or whatever) to do geometric multigrid. Otherwise, start with algebraic multigrid, -pc_type gamg, or configure with ML or Hypre and -pc_type ml or -pc_type hypre. The iteration count should be much lower, though each iteration will be somewhat more expensive.</div>
<div class="gmail_quote"><br></div><div class="gmail_quote"><br></div><div class="gmail_quote">On Sat, Jun 30, 2012 at 12:16 AM, Bao Kai <span dir="ltr"><<a href="mailto:paeanball@gmail.com" target="_blank">paeanball@gmail.com</a>></span> wrote:<br>
<blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">Hi, all,<br>
<br>
I am currently trying to solve a solute transport problem with PETSC<br>
and finite difference method with structured grids. Two equations are<br>
involved, one is a transport equation with advection and diffusion<br>
terms, another one is the flow equation with gravity, which is<br>
actually a poisson equation. Both the equations use the same solver.<br>
<br>
Since we eventually want to handle non-linear problems, we us SNES<br>
here to solve the equations. I use the BICGSTAB ksp linear solver and<br>
block Jacobi preconditioner. All the tolerances are the default ones<br>
from PETSC. The range of the solution values for transport equation<br>
is about 0-1100, and the range of the solution values of the flow<br>
equations should be between -5000-5000.<br>
<br>
The problem is that the solution of flow equations takes about ten<br>
times of the time than the transport equations. I am wondering if I<br>
should set different tolerances for the two equations.<br>
<br>
Attached please find the convergence processes of the two equations.<br>
Generally it takes about ten times of iterations to solve the flow<br>
equations than transport equations. I only do applications and do not<br>
have good knowledge in the solution of linear system, so I am here to<br>
find some suggestions on this issue.<br>
<br>
The size of the mesh is 500X500X500 (125 million unknowns ) and 512<br>
nodes (2048 processors ) on BlueGene/P machine are used.<br>
<br>
Any suggestions will be much appreciated.<br>
<br>
Thank you .<br>
<span class="HOEnZb"><font color="#888888"><br>
Kai<br>
</font></span></blockquote></div><br>