[petsc-users] Help with ML/BoomerAMG

[Karthik Duraisamy]

Thanks. I was asking because I use PETSc+GMRES for RANS of turbulence and combustion (on unstructured grids). The most difficult problems don't get off the ground without a local LU preconditioner, but otherwise ASM+ILU or ILU(k) works OK for me.

- Karthik

----- Original Message -----
From: "Gaetan Kenway" <gaetank at gmail.com>
To: "Jed Brown" <jedbrown at mcs.anl.gov>
Cc: "petsc-users" <petsc-users at mcs.anl.gov>
Sent: Monday, April 29, 2013 7:25:11 PM
Subject: Re: [petsc-users] Help with ML/BoomerAMG

To Karthik 


Yes, GMRES is my nominal default. For systems that are easier to solve, TFQMR also seems to be competitive with GMRES. It is the high mach RANS problems with strong shocks shock induced flow separation that seem to give me the most trouble. 


Gaetan 


On Mon, Apr 29, 2013 at 10:07 PM, Gaetan Kenway < gaetank at gmail.com > wrote: 


That makes sense. Is there a reasonably easy way of doing that in PETSc currently for reasonably large systems? 


Gaetan 




On Mon, Apr 29, 2013 at 4:32 PM, Jed Brown < jedbrown at mcs.anl.gov > wrote: 



Gaetan Kenway < gaetank at gmail.com > writes: 

> I would be very interested in looking at the problematic eigenmodes as 
> well. On my end I use Tecplot for all my visualization. What sort of 
> visualization technique are you thinking about? Is it the KSP subspace 
> vectors you want to look at? 

No, it would be an eigensolve for either outliers or eigenvalues very 
close to zero. The cost to find the vector is like several solves, but 
it would tell us exactly what sort of functions are poorly approximated 
by the preconditioner. Then we would think about how we can change 
algorithms to make the preconditioner correct those functions better.