Solver problem

[Tim Kroeger]

Dear Barry,

On Mon, 27 Jul 2009, Barry Smith wrote:

> On Jul 27, 2009, at 4:35 AM, Tim Kroeger wrote:
>
>> In my application, there is a linear system to be solved in every time 
>> step.  Steps 0 and 1 work well, but in step 2 PETSc fails to converge. I 
>> suspected that the system might be unsolvable in that step and checked that 
>> by writing matrix and the right hand side to files and loading them into 
>> "octave".  Surprisingly, "octave" does find a solution to the system 
>> without any problems.
>
>  Octave is using a direct solver. Did you try PETSc's direct solver using 
> -pc_type lu?

Good idea!  It works, and I'm actually surprised that it does.  I did 
try ILU(3) before (i.e., -pc_type ilu -pc_factor_levels 3), which took 
forever to compute.  Hence I thought that full LU would take even 
longer, but this turned out to be not true; LU is accetible in 
performance and solves the problem for that test case.  I'll see how 
it will behave on a larger-scale problem and on multiple cores.

Would you recommend to try MUMPS as well?  (I.e., will MUMPS have a 
change to be faster than ILU?)

Best Regards,

Tim

-- 
Dr. Tim Kroeger
tim.kroeger at mevis.fraunhofer.de            Phone +49-421-218-7710
tim.kroeger at cevis.uni-bremen.de            Fax   +49-421-218-4236

Fraunhofer MEVIS, Institute for Medical Image Computing
Universitaetsallee 29, 28359 Bremen, Germany