[HakikiCanakkaleli]

Hi

== 1 ==

For a lowest-effort-possible grid using LES-pisoFoam, using sample utility for 'U' seems yielding different results giving the same coordinates for:

1. Single processor
2. 'simple' parallelization method - 4 processors
3. 'hierarchial' parallelization method - 4 processors
4. 'scotch' parallelization method - 4 processors

== 2 ==

Actually 2 and 3 (i.e. simple and hierarchial) give identical results. Thus, 1, 2-3 and 4 produce the different results.

== 3 ==

Let me also note that, only 2 time-steps I run: 0.001 and 0.002 secs.

== 4 ==

The lowest-effort case files are given in the following link:

http://speedy.sh/bnAQM/1-test-to-send-away.zip

== 5 ==

The obtained results (for the given 'sampleDict' inside the above case files) can be seen from the following link:

http://speedy.sh/NTzGr/Results.zip

== 6 ==

Assuming the sample utility gives the exact answer from the above computational domain, could anyone please tell me what would be the reasoning behind these different answers despite the solution settings are the same?

Also note that the cases are not converged at all. But, I attempt to understand how the parallelization would affect the results obtained from the same setting cases?

e.g. when the simulation cases are numerically converged (e.g. according to the final residual of pressure) , should I expect that sampling of U will give the same results for all types of parallelization methods?

== 7 ==

Many thanks for any comment and time being spent.

[wyldckat]

Greetings HakikiCanakkaleli,

Once the solution is converged, the results should be nearly identical, but not necessarily 100% identical.

As for the several decomposition methods: running in parallel sometimes requires changing the parameters in "fvSchemes" and "fvSolution", because what works fine in serial mode, doesn't mean that it will run fine in parallel mode.
Here's an example (I couldn't find another one I was looking for): http://www.cfd-online.com/Forums/ope...lllimited.html


And I'm unable to download the files from that website. Can you share the files via Dropbox, Google Drive, SkyDrive or something similar?

Best regards,
Bruno

[HakikiCanakkaleli]

Many thanks Bruno for your time and explanations,

Here the files:

"
== 4 ==

The lowest-effort case files are given in the following link:

https://drive.google.com/file/d/0BzV...it?usp=sharing

== 5 ==

The obtained results (for the given 'sampleDict' inside the above case files) can be seen from the following link:

https://drive.google.com/file/d/0BzV...it?usp=sharing
"

[wyldckat]

Hi HakikiCanakkaleli,

I've finally managed to have a look into your files, but I haven't managed to do any tests with them, due to insufficient time...

But here's what I've seen that might be the reason for the big differences in results:

1. The divScheme "Gauss vanLeer" is one of those schemes that shouldn't be used lightly. In other words, it's not as easy to use as "Gauss linearUpwind" and "Gauss limitedLinear". I suggest that you get into the formulation of the "vanLeer" scheme, to ensure that you know what's at play.
2. If you have a look at the "fvSchemes" located at "incompressible/pisoFoam/les/motorBike/lesFiles/", you'll see how they configured the gradSchemes:
   

   https://github.com/OpenFOAM/OpenFOAM...iles/fvSchemes
   

   gradSchemes
   {
       default         Gauss linear;
   
       grad(nuTilda)   cellLimited Gauss linear 1;
       grad(U)         cellLimited Gauss linear 1;
   }

   As I mentioned in the previous post, the "cellLimited" option is used a lot for multi-core simulations, because it makes it easier for the data to be sort-of stable between processor sub-domains. Try using that option to check if the results seem more consistent.

Since I don't have much time to study this deeper, feel free to test this yourself and get back to us here on the forum, since the more information is available, the easier is for others to pitch-in!

Best regards,
Bruno