CFD Online Discussion Forums

CFD Online Discussion Forums (http://www.cfd-online.com/Forums/)
-   OpenFOAM Running, Solving & CFD (http://www.cfd-online.com/Forums/openfoam-solving/)
-   -   AMG solver and other queries (http://www.cfd-online.com/Forums/openfoam-solving/59742-amg-solver-other-queries.html)

msrinath80 March 29, 2007 00:23

1. Why is there a limit of 501
 
1. Why is there a limit of 501 iterations on the AMG and 5001 on the ICCG solvers? Is this is a known and fixed constraint?

2. For non-conformal structured block meshes, does increasing the number of non-orthogonal correctors really help? What is the ballpark percentage limit (as reported by checkMesh) beyond which non-orthogonal correctors are absolutely necessary. Is it 25%, 50%, 75% ??

3. In the lid-driven cavity case, dirichlet boundary conditions are prescribed for the stationary walls. Quoting from F & P:

"At a wall the no-slip boundary condition applies, i.e. the velocity of the fluid is equal to the wall velocity, a Dirichlet boundary condition. However, there is another condition that can be directly imposed in a FV method; the normal viscous stress is zero at a wall."

Is this implicitly done in the OF solvers? If that is the case, then the zero value dirichlet boundary conditions specified in the subdirectory 0/U should apply to the continuity equation?

Please excuse my naiveness. Any thoughts/corrections are much appreciated. Thanks!

msrinath80 April 12, 2007 23:44

Answering my own question. I'm
 
Answering my own question. I'm sure it will be useful for other n00bs like myself:

1. Possibly because a well-posed problem (proper BC and/or discretization) should definitely not require more than that many iterations. How do I know this? After fooling around with all kinds of meshes for the flow past a bluff body, I have concluded that the time invested on proper meshing is well rewarded. Case in point, when I tried to reduce the mesh size by introducing non-conformal blocks into my domain giving a mesh size of approximately 1 million cells, I also introduced more problems for the solvers. As a result, most of the iterations topped around 450 to 500 for the AMG solver and around 1500-2000 for the ICCG solver. Not only that, I also had to reduce the time-step to a very very low value (sometimes even 0.00025) to keep the Courant number from blowing up (stability requirement).

After properly creating the mesh (and by that I mean not exceeding an aspect ratio of 1:5 on any cell in the domain and keeping it strictly orthogonal), the multigrid solver took only 100-120 iterations despite the new mesh being 4 times as big as what it was before!

To sum up crappy discretization is identically equal to crappy performance.

msrinath80 April 12, 2007 23:51

Update: On the proper mesh, th
 
Update: On the proper mesh, the number of iterations for the AMG solver has reduced to around 40 now. Only the first 5-10 iterations topped around 100-150.

paka April 13, 2007 05:40

This is interesting. By sayin
 
This is interesting.
By saying 'the new mesh being 4 times as big as what it was before' you mean around 4 million cells?

BTW. just because of my curiosity, what kind of hardware/machine are you using?

msrinath80 April 13, 2007 09:20

Exactly. 4 million cells with
 
Exactly. 4 million cells with optimal discretization solves faster than 1 million cells with crappy discretization.

Check one of the earlier messages on this forum concerning Super-linear speedup. The machine info is listed there.


All times are GMT -4. The time now is 07:33.