[cfdonline2mohsen]

Dear All

After reviewing all the posts regarding decomposition methods in the forums, I realized that: (Please Correct me if I'm wrong)

1. Metis method has some restrictions about licensing problems.
2. scotch encounters some problems with snappyHexMesh, so ptscotch must be used at first for decomposition and after that scotch must be used for running the case.

I did not generate my mesh with snappyHexMesh (Fluent2Foam) so either scotch or ptscotch can be used. Here are my questions:

1. which Method (scotch or ptscotch) is better from the viewpoint of accuracy and speed? Please share your experiences about this 2 methods in here.
2. How much my running time will be shortened using either of these 2 methods in comparison with Simple decomposition method? Please Share your experience?
3. I don't know the differences between scotch and ptscotch methods so can anybody introduce some fundamental references about these methods (Thesis, books or papers)?

Thanks in Advance

[cfdonline2mohsen]

Please Some body help!

[wyldckat]

Greetings Kia,

I've got the feeling that you did very little research into this forum on this topic.

In a nutshell:

• "scotch" is used in serial/sequential mode.
• "ptscotch" can only be used in parallel, namely when the domain is already decomposed.

As of OpenFOAM 2.2, the two decomposition algorithms share the same name in the "decomposeParDict", i.e. they both are called "scotch". Which one is used depends on whether the application is executed in parallel or serial/sequential mode.


Many have shared their experience with "simple", "hierarchical", "metis" and "scotch" here on the forum. Feel free to search for the information

But again, in a nutshell:

• Use the simpler algorithms for when you have a simpler mesh.
• Use the complex algorithms when you have a complex mesh.

Best regards,
Bruno

[cfdonline2mohsen]

Thanks Bruno,
I used the scotch method but neither my run time nor the results did not change significantly in comparison with the simple method. (for the flow around a cube)
but for more complex geometries (film cooling of turbine blades) our run time reduced to 1/3 using scotch method. isn't that strange?

[wyldckat]

Like I said: simple meshes only need simple algorithms. Complex meshes, need complex algorithms. Scotch is a complex algorithm.

And since you didn't specify the number of cells on your meshes, nor the number of sub-domains (aka number of cores) in those cases, my guess is that: no, it's not strange, not strange at all.

[nsf]

Hi,

This might be covered better in other parts of the forum. In my experience metis might yield really bad decompositions. At work we saw a drastic decrease in performance with metis (using fluent) on more than 64 domains. When we changed the decomposition method we saw good speed up until 224 domains so it was not due to too small case. Since then we stopped using metis so I have no more experience of it.

I've seen others having problems with metis and OpenFOAM though, there is an article on this here
http://link.springer.com/chapter/10....642-24669-2_12. (I'm not affiliated with any of the authors).

Their main conclusion is not that metis is to be blamed but they did notice that the domains was not continuous on some processors. I.e. one processor was handling two separate domains.

If you know of an extensive study the please let me know.
Best
Nicolas

[cfdonline2mohsen]

Thanks Nicolas for sharing your experiences and also for introducing the article.