[Attesz]

Hi,

I'm dealing with an axial fan of 3 regions, and I'm currently using the scotch decomposition method. In your opinion, what is the best decomp. method for such cases (long domain, 2 interface). I'm thinking about the structured method...

[olivierG]

hello, just my 2 cents

From some test i made some time ago, ... it depend of your geometry.
- With "simple" one like pipe and / or very elongated in one or two direction, try the simple/ hierarchical method in the corresponding direction.
- with no clear direction, scotch will do a very good job.


regards,
olivier

[Attesz]

Thanks Olivier. My geometry is elongated in y direction, but I'll run on 16cores or more. Decomposing the geometry as (2 4 2) could be a good start in your opinion?

[olivierG]

No, if your geometry is on the y axis, i would try (1 16 1).
But make a small bench, and you would get your answer ...

regards,
olivier

[Attesz]

But make a small bench

I've 20millions of cells and transient simulation, so doing a small benchmark is not so easy but thank you, I'll try it!

best,
Attila

[olivierG]

then decompose using different setting, and sum the total of exchange face: the less = the best.