[floquation]

Background:
With decomposePar it is possible to cut the domain in the x, y and z directions such that each processor receives (roughly) the same number of cells. That's the simple/hierarchical method.
The scotch method does something similar, but then automated.
Either way, each processor will receive (roughly) the same number of cells.

Question:
Is there, in OpenFoam, a way to distribute the cells unevenly over the processors? That is, how can I enforce some processors to receive less cells than others?
(Preferably except for using "manually".)


(
"Why would you do that?"
In case you wonder why I would want to do that:
You'd like to spread the load over all processors, not cells! Without any further information, your best guess to distributing the load will be to give all processors the same number of cells, so the default behaviour makes sense.
However, when you do know that a certain region will be computationally more expensive, it will be desirable to assign less cells of that region to a processor and more cells of the uninteresting ambient region.
In my specific case, this is because I use dynamicRefineFvMesh with interDyMFoam. I know in advance roughly what region will be refined. Therefore I'd like to assign less cells of that region to each processor, such that in the refined mesh I obtain roughly the same load per processor.
);

[Zbynek]

I never needed to consider what you are trying to do so I maybe won't be of much help. Anyway, there is an option in decomposeParDict to choose method "manual" where you allocate each cell to a particular processor. I did not see any example of how to define this so I am afraid you would have to have a look in the code what input it requires.

A possible workaround that comes into my mind is that you would terminate a calculation before/after each refine step, reconstruct the domain and decompose it again. This would be applicable if the refinement does not happen too often so the overhead caused by reconstruct/decompose steps would not slow down your computation too much.

[akidess]

Hi Kevin, manual distribution is an option as Zbynek said; dynamic load balancing is another:
Run-Time Parallel Load Balancing

[Zeppo]

Originally Posted by floquation

Question:
Is there, in OpenFoam, a way to distribute the cells unevenly over the processors? That is, how can I enforce some processors to receive less cells than others?
(Preferably except for using "manually".)

// "system/decomposeParDict"
numberOfSubdomains 4;
method                      scotch;
scotchCoeffs
{
    processorWeights    (10 1 1 1);
}

[guin]

Hi Kevin,
I recently ported the runtime load-balancing utility to OF-4.1. Maybe it is useful for you:
Run-Time Parallel Load Balancing