Learning Forum

[aitor.amatriain]

I need to perform some simulations in a cluster, and I have an estimate of the required RAM memory and number of cores. However, I will also need to postprocess the results (and maybe create the mesh) in my computer, and for that purpose I am not sure about the requirements.

As far as I am concerned, given a mesh size, mesh generation needs less RAM than the case simulation (around 50% less, based on my experience), so in this regard I can have an estimate based on what I can do in my current compute. But what about the requirements for post-processing? I understand that the increase of the number of cores will speed-up the calculations (read the results, compute surface integrals...), but is there any additional limitation that I should take into account? Does RAM memory have an impact in post-processing?

Thank you in advance,

Aitor

[Rob]

RAM limits the size of model you can open, but post processing generally uses less RAM than running apart from particle tracks as they're computed as they're plotted. Figure 2-3GB per million cells to run, a bit less will work for generating images in Fluent. Parallel may speed things up, but is less critical than for running: cpu gets you the solution speed.
Mesh generation is cheaper than solving (no chemistry or multiphase in the mesher!) but excessive use of size functions will effect RAM usage. Fluent Meshing is pretty efficient, and we've generated a few 100Million cells on 32/64GB RAM at various times.

[aitor.amatriain]

In my case I have 32GB of RAM and I am limited to less than 20GB in meshing, so I see that there is a strong dependence on the type of mesh.
In regards to post-processing, I would only like to display contours and compute surface integrals. Do you have an estimate of the required RAM memory in comparison with meshing? My intuition says that it should be much less, but I am not sure about that.

[Rob]

I don't see why you'd not be able to use all of the RAM when meshing. The tools typically take what they need which is why we occasionally finish up with memory issue and/or swapping to disc.
Hard to judge relative to meshing as you're storing the cell data in RAM and then (potentially) running some post processing routines to check min/max and integrate values. With single phase and no chemistry you may well be right, add in turbulence, species etc and it's much harder to judge. I'm not aware of any documents on this, and in most cases commercial users have a larger memory machine or ensure the head node on the cluster can use all the RAM on that box.

[aitor.amatriain]

Thank you for your answer. Maybe my message was not clear, but what I wanted to say is that a mesh of 20M cells requires more than 32GB of RAM in my particular case.

[Rob]

In Fluent Meshing or Ansys Meshing? I've built some very big Fluent Meshing meshes on my laptop, and that's only 64GB. Workbench (Ansys) Meshing is far less memory efficient as the size functions aren't as well scoped.

[DrAmine]

32Gb or 64GB are standard for now and for post-processing you can rely on Client / Server Connection if you need to deal with very large cases requiring a lot of power for post-processing. A GPU on top of that is required for some rendering if needed. I will be more selective when writing result files out of Fluent or CFX and do just select what you want to look into and use Ansys EnSight for post-processing using EnSight Gold Case Format (relying on solver built-in capabilities to automatically export images or out files if required).