[FlyBob91]

Hello to all,
I'm moving my first steps in the world of parallel simulation and HPC. If i understood corectly, with parallel simulation i can divide my domain and let the single core of my CPU solve that part of domain assigned. So, if i had a quad core CPU, i could divide the domain in 4 part, and not more. Is it right?
I run the pitzDailyExptInlet case, with my own PC, in the normal and parallel way but i didn't notice a substantial difference of velocity. Is my sensation wrong?

Now a question for people who have already worked with HPC system.
It is the first time i work with HPC system and i'm not sure about the structure of the system. To submit a job i have to specify:
1- The number of node NN
2- The number of CPUs (ncpus) max 16 in my case
3 -The number of MPI tasks per node (mpiprocs)

I don't understand the meaning of 1 and 3.
Suppose to choose these numbers (1,16,1) for the three value, this means that i have to decompose my domain in 16 regions, is it right?

Last question: how does decomposePar work? Which criterion i should use to divide my domain? Which directions i should choose?


Thanks for help.

[Svensen]

Originally Posted by FlyBob91

Hello to all,
Last question: how does decomposePar work? Which criterion i should use to divide my domain? Which directions i should choose?

decomposePar subdivides your computational domain on subdomains according to system/decomposeParDict.
I suggest you to use 'scotch' method, which will do subdivision automatically without any extra input data.

The number of subdomains in system/decomposeParDict could be arbitrary, but normally it is N or 2*N, where N is a total number of cores.

Also, in HPC system you don't need to specify all (1)-(3) parameters. You should know which management system runs on your HPC station. If it is SLURM, then you can type 'sbatch -n 16 icoFoam -parallel', for example, where '16' is a number of threads (it should be equal to a number of subdomains (system/decomposeParDict)).

[FlyBob91]

Thank you Svensen, you clarified me some doubts.
To be precise my scheduling system is PBS.

Best Regards,
Roberto

[Svensen]

Normally for every HPC cluster there are some scripts, which used to submit a task for computing. The scripts generally are prepared by cluster administrators.