Hello.

Generally speaking: if we want to run a huge case on a cluster (i.e. a distributed-memory system), do we then need to have ONE machine with a huge amount of RAM to read in the mesh, and partition it, in order to run the job?

We have 10 powerful machines, each with 16G of RAM. Say we want to run a job which consumes almost all RAM on each machine. We are able to generate a very big mesh in ICEM using only 16G-20G on that particular machine, but when we read it into Fluent, a 16 G cluster-machine is nowhere near enough.

Seems odd that you should need to have all that RAM available both on one host-machine AND the on all the separate client-machines. That is, one machine with 64G and then our 10 machines with 16G each. Is this really needed?

We start-up Fluent in parallel mode, but it seems to want the entire mesh to reside on the host, before partitioning it. This probably makes sense, but I wonder if there is some workaround we could try, other than buying lots of (more) RAM for one of the machines.

Thanks!

/Mads

How many Million cells are you trying? If you are over 160 million cells then it might be an issue as collectively your machines may not have enough RAM. The mesh size in ICEM will use much less memory than FLUENT when it imports, initializes and runs.

The biggest I have done is a 14 Million cell case, and that is on a 16GB total cluster, and it read fine. We had it running parallel for the import, and just did the normal import routine.

I have found on another machine, that some meshes require me to import the mesh and set up the case file and run a couple of iterations on one node, before saving and switching to parallel.

Regards,

Scott

FLUENT should have the auto partitioning without requiring any big memory on a specific machine. Which version of FLUENT are you using?

Okay. Running 6.3.26 for production, tried 6.3.35, and 6.4/12 betas. Still, requirements are quite high.

/Mads