Hello,
Sorry, I am not sure on which cases this was observed tested (this is based on recent work by colleagues).
We use multiblock meshes on some cases, and other types of hybrid meshes on others, so i! am not sure about this either (though multiblock based is usually preferred frequent for "high quality" meshes, since cell size progression rate is critical for many turbulence model, and often harder to control with tet or polyhedral meshes (or not controlled well by users who want a quick mesh).
Best regards,
Yvan
Clipping of the Reynolds-stress tensor
Forum rules
Please read the forum usage recommendations before posting.
Please read the forum usage recommendations before posting.
-
Yvan Fournier
- Posts: 4206
- Joined: Mon Feb 20, 2012 3:25 pm
-
StandardRANSUser001
- Posts: 45
- Joined: Tue Apr 09, 2024 3:26 am
Re: Clipping of the Reynolds-stress tensor
Hi Yvon,
Just as an update. The strange clipping phenomenon of the Reynolds-stress tensor (CS8.2) is no longer an issue in CS8.3.
I observed that this became more prominent as the number of cores increased (i.e. MPI issue). There were two variables in this test (CS8.2 on an AMD EX Cray system, and then CS8.3 on an intel HPC.). I wanted to verify that CS is being tested on EX CRAY?
In addition to this, I am running the EB-DRSM in CS8.3, and I am using a SOLU scheme with isstpc=0. I am not observing any clipping with the slope test, whereas I observed a clipping with the slope test enabled with the k-w SST model. I would have expected this slope test to work with the EBRSM + SOLU scheme. Hopefully I have understood how the slope test works. Is this a bug in CS8.3?
Best regards,
Sean hanrahan
Just as an update. The strange clipping phenomenon of the Reynolds-stress tensor (CS8.2) is no longer an issue in CS8.3.
I observed that this became more prominent as the number of cores increased (i.e. MPI issue). There were two variables in this test (CS8.2 on an AMD EX Cray system, and then CS8.3 on an intel HPC.). I wanted to verify that CS is being tested on EX CRAY?
In addition to this, I am running the EB-DRSM in CS8.3, and I am using a SOLU scheme with isstpc=0. I am not observing any clipping with the slope test, whereas I observed a clipping with the slope test enabled with the k-w SST model. I would have expected this slope test to work with the EBRSM + SOLU scheme. Hopefully I have understood how the slope test works. Is this a bug in CS8.3?
Best regards,
Sean hanrahan
-
Yvan Fournier
- Posts: 4206
- Joined: Mon Feb 20, 2012 3:25 pm
Re: Clipping of the Reynolds-stress tensor
Hello,
We do not have access to a Cray machine, though code_saturne has been run regularly at least on the Archer 2 machine by several researchers.
There have been a succession of changes in the RSM models in the last year, many of them oriented towards trying to avoid or reduce clippings, and I did not follow the details of those (mainly handled by 2 other colleagues).
The standard slope test should work on all models, though some specific TVD/NVD limiters enabled in the GUI since version 8.1 or 8.2 may work only in scalar variables (i.e. not Rij). But differences in the models themselves may lead to different clipping levels.
I'll suggest my colleagues to check this thread, because I really am not of much more help on this particular subject.
Best regards,
Yvan
We do not have access to a Cray machine, though code_saturne has been run regularly at least on the Archer 2 machine by several researchers.
There have been a succession of changes in the RSM models in the last year, many of them oriented towards trying to avoid or reduce clippings, and I did not follow the details of those (mainly handled by 2 other colleagues).
The standard slope test should work on all models, though some specific TVD/NVD limiters enabled in the GUI since version 8.1 or 8.2 may work only in scalar variables (i.e. not Rij). But differences in the models themselves may lead to different clipping levels.
I'll suggest my colleagues to check this thread, because I really am not of much more help on this particular subject.
Best regards,
Yvan
Re: Clipping of the Reynolds-stress tensor
Hello,
I allow myself to jump inside this topic.
I have faced the clipping bahavior of EBRSM with CS for some configurations.
I guess what you call clipping here is the fact that the stresses are clipped to very low values, to prevent zero or negative values?
Could you share further details about the flow configuration you are simulating (if I am not wrong you haven't written anything in the previous posts, sorry if I missed something)? It would be helpful and interesting to understand common flow configurations leading to this behavior of the model, and to discuss further.
Moreover, have you tried to visualize the regions where the solution features zero Re stresses? Does the the number of cells clipped reach a steady value, or does it increase constantly and eventually the entire domain liminarize?
Thank you.
Best regards,
Daniele
I allow myself to jump inside this topic.
I have faced the clipping bahavior of EBRSM with CS for some configurations.
I guess what you call clipping here is the fact that the stresses are clipped to very low values, to prevent zero or negative values?
Could you share further details about the flow configuration you are simulating (if I am not wrong you haven't written anything in the previous posts, sorry if I missed something)? It would be helpful and interesting to understand common flow configurations leading to this behavior of the model, and to discuss further.
Moreover, have you tried to visualize the regions where the solution features zero Re stresses? Does the the number of cells clipped reach a steady value, or does it increase constantly and eventually the entire domain liminarize?
Thank you.
Best regards,
Daniele
-
StandardRANSUser001
- Posts: 45
- Joined: Tue Apr 09, 2024 3:26 am
Re: Clipping of the Reynolds-stress tensor
Hi Daniele,
The conclusion that I came to was that there was an MPI issue in CS8.2 relating to the Rij tensor, as the number of clipped cells was increasing with core count. When I visualised this, the clipping seemed to be related to the regions of the interface between MPI blocks. As I increased the number of cores, the number of clipped cells also increased. This problem wouldn't really be evident if you were undertaking a small benchmark case, as it only becomes apparent when working on larger cell counts, where MPI is required (say +10M cells).
This clipping wasn't related to the physics - it was in similar locations to the MPI partitioning edges. It did vary slightly with time, but not by a lot.
In CS8.3, an update was pushed to the GIT to address this issue. I have run this code with upwind schemes, coupled Rij, and refined structured mesh. The Reynolds-stress tensor and epsilon now looks a lot better. Clipping occurs for a very very small number of cells, but I think that is because of the size of my timestep (this is also a condition of realisability).
Please let me know if you need more information on this.
Are you studying a buoyancy driven flow, or a flow where thermal effects are not a focus?
Best regards,
Sean Hanrahan
The conclusion that I came to was that there was an MPI issue in CS8.2 relating to the Rij tensor, as the number of clipped cells was increasing with core count. When I visualised this, the clipping seemed to be related to the regions of the interface between MPI blocks. As I increased the number of cores, the number of clipped cells also increased. This problem wouldn't really be evident if you were undertaking a small benchmark case, as it only becomes apparent when working on larger cell counts, where MPI is required (say +10M cells).
This clipping wasn't related to the physics - it was in similar locations to the MPI partitioning edges. It did vary slightly with time, but not by a lot.
In CS8.3, an update was pushed to the GIT to address this issue. I have run this code with upwind schemes, coupled Rij, and refined structured mesh. The Reynolds-stress tensor and epsilon now looks a lot better. Clipping occurs for a very very small number of cells, but I think that is because of the size of my timestep (this is also a condition of realisability).
Please let me know if you need more information on this.
Are you studying a buoyancy driven flow, or a flow where thermal effects are not a focus?
Best regards,
Sean Hanrahan
Re: Clipping of the Reynolds-stress tensor
Hi Sean,
Thank you for explaining in details your issue.
I misunderstood that it was actually purely a numerical problem, not linked with the physics.
The flows in which I faced the clipping issue were iso-thermal. And the clipping occurred in specific flow regions, not linked with the partition interfaces.
Best regards,
Daniele
Thank you for explaining in details your issue.
I misunderstood that it was actually purely a numerical problem, not linked with the physics.
The flows in which I faced the clipping issue were iso-thermal. And the clipping occurred in specific flow regions, not linked with the partition interfaces.
Best regards,
Daniele
-
StandardRANSUser001
- Posts: 45
- Joined: Tue Apr 09, 2024 3:26 am
Re: Clipping of the Reynolds-stress tensor
Hi Daniele,
Could you please post (or PM) a few screen shots of your problem? It would be good to see the components of the Reynolds-stress tensor, epsilon, alpha, MPI partitioning, and the velocity components as well. Could you please also show me your run_solver.log summary of the residuals and clipping?
Have you tried to reduce your timestep as well? I.e Is this behaviour subsiding as the timestep is reduced?
This is something that I am quite interested in resolving.
Best regards,
Sean Hanrahan
Could you please post (or PM) a few screen shots of your problem? It would be good to see the components of the Reynolds-stress tensor, epsilon, alpha, MPI partitioning, and the velocity components as well.
Could you please also show me your run_solver.log summary of the residuals and clipping?Have you tried to reduce your timestep as well? I.e Is this behaviour subsiding as the timestep is reduced?
This is something that I am quite interested in resolving.
Best regards,
Sean Hanrahan