[tiam]

Hello!

I have been running LES of channel flow for the meshes of following sizes: 1, 2, 4 and 8 million.
For each mesh I have 2 case which differ only in one aspect: one uses cubeRootVol delta, and the other one van Driest.

I decompose the cases in such a way that each processor has around 30k cells, which as far as I know a reasonably large amount.

For the cubeRootVol cases I have very nice scaling in the sense that I can compute (almost) the same simulation time for all the mesh sizes for a given wall-time.

However, for the van Driest cases I have problems. For the 1 and 2 millions meshes the performance is pretty much on par with cubeRootVol. But for 4 and 8 million, the difference becomes very significant.

For the 4 million case I have 0.52 hours per simulation-second for the cubeRootVol and for the van Driest 0.75 hours, so about 1.45 times slower.

For the 8 million case I have 0.47 h/sim-sec for the cubeRootVol (better than the 4 mill case !) and for the van Driest delta I have 1.76 h/sim-sec!!

As I said the cases are identical besides for the choice delta. Also by observing the log as the simulation runs I see that I have a pause in the end of each time-step, after the U and p solvers are done. As far as I know that is where turbulent quantities get updated.

I am using version 2.3.1.

Has anyone encountered similar issues? Is this to be expected? As I understand the van Driest delta has to calculate the wall distance. This should be done only once for a static mesh. Maybe the issue is that it is recalculated every time?

Best,
Timofey

[tiam]

As a follow up.
Indeed, the distance is evaluated at every time-step.
This is necessary, since we are interested in considering cells where y+<500 or some similar value.
Since y+ gets updated every time-step we have to also recalculate the distance.

I am wondering if this is needed, since the form of the damping function makes it dissappear away from the wall anyway.
So it should be possible to just caluclate the whole distance field once and then use it, whithout worrying about the damping acting only near the wall.
Or am I missing something?

[tiam]

An update.

The behaviour has now also been confirmed using a larger mesh (69 mil.) and a standard scaling test, i.e. doubling the amount of cores and comparing to the increase in the simulated time.

Fixing this is not very easy it seems. The damping function depends on y+, calculated for the interiour cells. On a general geometry it is not trivial which u_tau to pick up to do the scaling for each cell. And one is forced to recalculate this field every time-step.

[beatlejuice]

Hello tiam,

I have exactly the same problem!

I have to identical LES Smagorinsky cases which differ only in the choice of delta: cubeRootVol & van Driest.

I also observe a very long pause after one time step is solved (after solving for U & p and preparing for next time step) for the van Driest method. This pause does not exist in the cubeRootVol simulation.

Did you find any work around or does anybody has any solution to this.

I use OF 2.3.1

cheers!

[tiam]

Originally Posted by beatlejuice

Hello tiam,

I have exactly the same problem!

I have to identical LES Smagorinsky cases which differ only in the choice of delta: cubeRootVol & van Driest.

I also observe a very long pause after one time step is solved (after solving for U & p and preparing for next time step) for the van Driest method. This pause does not exist in the cubeRootVol simulation.

Did you find any work around or does anybody has any solution to this.

I use OF 2.3.1

cheers!

Hi! No, there is now workaround I think. I would consider using a different SGS model, WALE for instance. This should be looked i to by the devs.

[beatlejuice]

Hey tiam,

thx for your fast reply!

Did you use the WALE model? Were the results satisfying?

Do you know what the field B is?
What boundary condition did you use for the walls, inlet, outlet for the fields B, k, nut, nuTilda?

I would appreciate yout help!
Thanks! Beatle