[petsc-users] iterative solvers

[Manav Bhatia]

Thanks Barry. 

Your message provides a good starting point. I will need to do more reading to understand my options. 

Thanks,
Manav



> On Apr 7, 2015, at 3:00 PM, Barry Smith <bsmith at mcs.anl.gov> wrote:
> 
> 
>   You cannot just use algebraic multigrid directly on the transonic Euler flow problem discretized with SUPG you must break the problem into pieces (probably using PCFIELDSPLIT) and then use appropriate preconditioners for each piece. 
> 
>   Barry
> 
> Note that ILU is a bottom feeder preconditioner, it is only working because you are using a huge fill factor (so it is a lot like a direct solver) and even then it is working really poorly. You should google for good preconditioners for transonic Euler flow  and SUPG (and ignore anything that mentions ILU) to get a handle on how PCFIELDSPLIT could be used for your problem.
> 
> 
>> On Apr 7, 2015, at 7:06 AM, Manav Bhatia <bhatiamanav at gmail.com> wrote:
>> 
>> Hi,
>> 
>>   I am solving a transonic Euler flow problem discretized with SUPG. The mesh is made of Tet4 elements and there are about 7M dofs, which I am trying to solve over 192 cores. 
>> 
>>   I had earlier written about the linear solver returning with Inf, and have since moved beyond that such that I am able to get a solution with the following command line parameters:
>> 
>> -ksp_gmres_restart 100  -pc_type asm -sub_pc_type ilu -sub_pc_factor_levels 4 -sub_ksp_type preonly
>> 
>>   The problem now is that I am limited to very small time-steps (~10e-6), outside of which the solver starts to choke and the solution diverges. I have tried other command line options and this above combination is what seems to provide the best possible solution strategy, albeit constrained in time steps. 
>> 
>>    I would appreciate any recommendations on solver configurations that may allow me to take larger time-steps. 
>> 
>>    Would coding up AMG be a possible alternative (more robust and dependable)? 
>> 
>>    Please let me know if you need additional information. 
>> 
>> Regards,
>> Manav
>> 
>