code_saturne User's Forum

Hi,
I am trying to run a thermal simulation
with radiative transfers.
My geometry is mesh with 1.46 millions
of cells. A first calculation with convection (no radiative) take 13
second by iteration.
The same mesh take 1000 seconds by
iteration with radiative heat transfers (Discrete ordinates method).
The mesh has prisms layer on wall
boundary (1 st cell = 1mm and and 10 layers of prisms)

In order to know if I don't have
numerical issue, I make a second coarse mesh with 0.32 millions of
cells (1 st cell=3mm and 5 layers of prisms). The time
calculation is 84 second by iteration.

Do you think the rapport of time
calculation between the convection calculation and the radiative
calculation is normal?

Merci
Best Regards

Alexandre

Hello,
You have over 20 additional equations to solve, so going from 13 to 100 would not be surprising, though 1000 is.
Could you post your "listing" file to check info on linear solvers ?
You may also try reducing the solver resolution precision for radiative variables.
Also, the radiative variables are solved using Jacobi, which can be very slow in the first time steps (it is usually faster after the first iterations for most variables, as linear systems become more diagonal dominant)
An interesting test would be to add the raysol.f90 source file from the Code's source tree (ncs/src/rayt/raysol.f90) to your user Fortran subroutines, and replace the "iresol = 1" (line 319) by "iresol = 2" so as to solve Radaitive equations using a Bi-CGSTAB solver. I do not know if this will help, but am interested in your feedback on this: either it may fail and diverge, or you may get faster calculation times.
Best regards,
  Yvan Fournier

Hello,
I have attached 3 listing files:
   - Convection calculation
   - Radiative calculation with Jacobi
   - Radiative calculation with Bi-CGSTAB (divergence)

By the way, do you know if it's possible to write output in the terminal and in a listing file (not a batch file but a listing file with a specific name for each calculation, like listing.09291640) Which value for ARG_CS_OUTPUT?

I don't how I can reduce precision for radiative variables. I have reduce the precision of Pression,VitesseX, VitesseY, VitesseZ, Enerturb, omega and TempK to 1e-05, in the GUI interface.

Best Regards
Alexandre

Hello,
In raysol.f90, you may reduce the value of epsilp (defined line 158 to 1.e-8) to change the precision of the linear solver.
I believe there are different options to the DOM model, with a different number of computed directions depending on the precision required, and you are probably using the most precise (but costly) variant.
Otherwise, it is not currently possible to write output both in a terminal and in a listing file (you may set ARG_CS_OUTPUT="--log 0" in the runcase file, either directly or telling the GUI not to redirect the output to "listing", but this does not always work well on some compute clusters in parallel). If you direct the output to a terminal, you do not have a "listing" file anymore, so you would need to modify the script to add a "tee" command, which might impact performance).
Best regards,
  Yvan Fournier

Hello,
I have reduce the value of epsilp to 1.e-5 and I use 32 directions for angular discretisation.
The time calculation is now 200 seconds by iterations, it's more acceptable.
Thanks for your answers
Best Regards
Alexandre