Hi All, I want to us OpenFo
I want to us OpenFoam for LES combustion in the low Mach-number regime. There are two questions to start with:
1) which is the most appropriate solver in Foam to start with? Is it PISO or should we use a different scheme?
2) I am also interested to solve the equations in the zero Mach-number limit. In that case the energy equation reduces to a divergence constraint for the velocity leading to a pressure poisson equation. This will lead to a projection method type of scheme. Is there a fast solver for the poisson equation implemented? If yes, what kind of solver is it (AMG, Krylov subspace?).
How difficult is it to set up and solve a poisson problem with the unstructured grid in foam?
Dr. Michael Oevermann
Technische Universität Berlin
Institut für Energietechnik
Fasanenstr. 89, 10623 Berlin, Germany
Phone: +49 (0) 30 314 22452
Fax: +49 (0) 30 314 22157
1) Xoodles - sounds like exac
1) Xoodles - sounds like exactly what you need. It is a pressure-based compressible formulation using PISO.
2) Compressible PISO also solves a pressure equation - for the incompressibility limit you just lose the convection and the ddt terms. In this formulation, the varying compressibility just changes the nature of the pressure equation from hyperbolic (compressible) to elliptic (incompressible) and the solver can deal with both.
Foam contains both ICCG (Incomplete Cholesky preconditioned Conjugate Gradient) and AMG (Algebraic Multigrid) solvers. Both are excellent and the performance for elliptic problems is really not an issue.
Solving a Poisson equation is very easy - study the tutorials, because I suspect you'll spend more time learning how to use foam than actually solving the Poisson problem.
Hrvoje, thanks for the in
thanks for the information! Do you know if the AMG and ICCG are running on parallel machines? What is the speedup?
Can I do a Rolls-Royce trick
Can I do a Rolls-Royce trick :-) and say the speed-up is "sufficient"?
Foam is massively parallel and has been used up to 256 CPU-s (maybe more), with decent scaling results. The problem with concrete numbers is that it is very difficult to pick a representative case and a machine that will give proper data: if a case is too small it does not have enough work per node and if it is too big it does not fit onto a single CPU. I have done a test on a big Silicon Graphics (in a domain decomposition mode) and got speed up of about 11 on 16 CPU on a non-dedicated machine - there was other jobs running, the bus was loaded etc. so this is not the mest you can get.
As a general rule, you don't have to worry about parallel scaling - it is appropriate for the purpose and foam does parallelism well (that's how all the LES work is done).
As a guideline on the solver choice, ICCG scales better in parallel than AMG. AMG tends to talk a lot on the top-level (it does not reconstruct the top-level matrix on one node). This can be improved (it's been sitting on my to-do list for ages) but I don't see the priority. AMG is bets used whan ICCG tends to use a lot of iterations, either due to the mesh quality or tight solution tolerances.
|All times are GMT -4. The time now is 10:18.|