[juanblanco]

Hello, I'm currently coupling OpenFOAM (OF) with other code. The latter is compiled from OF directly. The problem is that the code (of type Monte Carlo) imports the full mesh of the original geometry and parallelizes in a different way from usual OF running. This parallelization is already implemented and working but then I want to split the mesh and run OF routines with the usual mpirun. At this point I have an inconvenience, I need to split the mesh, turn on mpirun, run the OF routine, recompose the mesh and turn mpirun off.
To summarize, the scheme of what I want to do goes like the following

1. Imports geometry and mesh specifications
2. While loop
--------3. Run code of type Monte Carlo.
--------4. Split mesh.
--------5. Turn on mpirun.
--------6. Run classic OpenFOAM routine parallelized.
--------7. Reconstruct mesh.
--------8. Turn off mpirun
9.End loop

If I try to use decomposePar and mpirun with -np processors from console I obtain that I run step 2 to 9 -np times simultaneously. So is like running the case -np times independently.

Also, with the command system() I tried to use decomposePar, runParallel and reconstructPar inside the code from linux scripts but I couldn't make it work. I've ended with the same problem as explained before.

In conclusion, I want a way to turn on and off mpirun from inside the code and not from terminal.

Observation: The coupled code need the data recompose in order to work, so it is necessary to gather the data at the end of each loop step.

I appreciate your collaboration and thank you for your time.
Juan

[floquation]

I don't know any details, but this might help you...
What I recall from an HPC course during my studies is that you should detect whether a certain processor is the "master" or a "slave". This should also be the case for OpenFOAM, as the beginning of a simulation outputs:

/*---------------------------------------------------------------------------*\
| =========                 |                                                 |
| \\      /  F ield         | OpenFOAM: The Open Source CFD Toolbox           |
|  \\    /   O peration     | Version:  4.1                                   |
|   \\  /    A nd           | Web:      www.OpenFOAM.org                      |
|    \\/     M anipulation  |                                                 |
\*---------------------------------------------------------------------------*/
Build  : 4.1
Exec   : interFoam -parallel
Date   : May 10 2017
Time   : 14:02:03
Host   : "n77"
PID    : 186229
Case   : /foo/bar
nProcs : 4
Slaves :
3
(
"n77.186230"
"n77.186231"
"n77.186232"
)

If you can trace where this message comes from, you immediately know how to separate the master from the slaves.

Then, if something should run in series, you can have something like (pseudocode):

if(master)
{
  do serial stuff
  send message to slaves to synchronise
}else{
  wait for message from master
}
continue with parallel stuff

[juanblanco]

Hi, thank you for the answer and sorry for delay.
For selecting the master is as you've said. Unfortunately, I had other problems inherent to my code so I couldn't make it work properly.

If it is useful for anyone the command I've used for the master is
Pstream::master()

Also, I've found a related thread. I attach the link
Access patch points (on different processor) in parallel

Regards