reproductibility with tke / dependence on the number of subdomains

[julienjouanno]

Context

Setup of regional configuration with bdy / tke / ubs on momentum / fct for tracers / trunk (version 6980)

Results depends on the number of subdomains.

It seems to works fine when using small number of subdomain (e.g. jpnij=80; jpnj=10,jpni=8) whereas the use of large number of procs (> 100) leads to very strange results.
See figure that shows the surface meridional velocity when using 10x8 subdomains (top) and 20x8 subdomains (bottom).

Analysis

• The error is not dependent of the machine (same problem on both curie and occigen supercomputers)
• Using GLS instead of TKE resolve this issue (solution does not depend on the nuùmber of subdomains)
• We see a strong imprint of the wind stress field in the corrupted case (bottom figure) additionaly suggesting that this is an issue with zdf/tke.
• Figures are shown for a case that do not use key_agrif, but when using an agrif zoom, only the parent grid seems to be affected

[julienjouanno]

meridional surface velocity after a few days of simulation

[julienjouanno]

zonal surface velocity in the parent grid when using agrif : 4x10 subdomains (top) vs 14*10 subdomains (bottom)

[julienjouanno]

namelist

[nicolasmartin]

Probably the ticket is much readable now with images rendering directly in the ticket.
Just a trick for further use.

[cbricaud]

I tried to reproduce this bug with trunk rev 8583
I used AMM12 reference configuration ( bdy), switched it from gls to tke, and to UBS adv scheme for dynamics
I compared 4*8 and 8*4 for MPI splitting, then 10*12 and 12*10

The configuration is reproductible for these tests and the "strange" paterns doesn't appear

[gsamson]

As Clément Bricaud did not succeed in reproducing Julien bug with AMM12 and a newer revision of the trunk (r8583 vs r6980), I consider either that the bug was fixed somewhere between the 2 revisions, either that we don't have enough elements from Julien configuration to reproduce the bug. So I close the ticket.