[petsc-users] BoomerAMG Howto

[Gong Ding]

----- Original Message ----- 
From: "Barry Smith" <bsmith at mcs.anl.gov>
To: "PETSc users list" <petsc-users at mcs.anl.gov>
Sent: Sunday, September 05, 2010 11:10 AM
Subject: Re: [petsc-users] BoomerAMG Howto



>You can call MatSetBlockSize() on the matrix with a value of 3. This information will then be transmitted to BoomerAMG so that it will know it is a 3 component problem. It may work a bit better with this information.

Thanks, I will try it.

>   But I suspect it will still not work well semiconductor problems are notoriously difficult for iterative solvers. You may just go best with a direct solver. We recommend the Mumps solver. You can use it by configuring PETSc with --download-mumps --download->scalapack --download-blacs and then run the program with -pc_type lu -pc_factor_mat_solver_package mumps

Yes, MUMPS is the default linear solver in my code. I found it is the most stable one in practical. However, direct solver become too slow for large 3D problem. 
I am trying to find some fast solver instead.


On Sep 4, 2010, at 9:41 PM, Gong Ding wrote:

> 
> I had tried BoomerAMG for poisson equation, which works well.
> But when I preform it on my semiconductor system, it does not convergence.
> I guess the reason is semiconductor equations has three variables, the potential, electron and hole density,
> the algebraic multigrid arithmetic does not consider this and treat coarsen/smooth between different variables.
> 
> Is there any way to specify coarsen arithmetic by user? 
> Since each mesh node has 3 variables, I would like to do point based coarsen operation.
> 
> 
> 
> 
>  BoomerAMG can not handle the linear system you are giving it. Each algebraic multigrid solver can only handle a certain class of problems.
> 
> 
>    Barry
> 
> On Sep 3, 2010, at 9:55 AM, Gong Ding wrote:
> 
>> Dear all,
>> I am trying to use Boomer AMG of hypre + GMRES for my nonlinear problem, which is solved by MUMPS or GMRES +ILU previously.
>> However AMG does not convergence.
>> 
>> The -ksp_monitor shows that there are huge residual during inner GMRES iteration, as large as 1e30.
>> I only change PC from ILU to boomeramg.
>> Can anyone tell me what's wrong with it?
>> 
>> Gong Ding