open TELEMAC-MASCARET The mathematically superior suite of solvers

Hello,

I use Telemac to compute several siting projections in parallel. I have a docker image with Telemac installed on it, and I launch my jobs with --ncsize=2. The parallelization works well, I see file splitting, and CPU usage repartition is on 2 of 8 device CPU.

But when I launch a second job, it is the same CPU which are used, so I have 6 lazy ones that never work, and 2 braves that handle the 2 jobs in parallel. How can I set the repartition to use full capability of my device ?

Thanks in advance,

Patrick

Hi,

This mpight be an issue either from your image configuration or from MPI configuration.

What does the command nproc returns ?

Could you post your systel.cfg file ?

Hi,

Here the systel.cfg used. nproc returns 8.

We don't modify the docker image, we use this one :
github.com/flussplan/docker-telemac

Thanks for your answer !

Hi,

I am puzzled by the option --oversubscribe for the mpi command.
This option allows to submit more core than are available on the machine this could be the issue ans is usualy not recommended.
Try removing it.


If that does not work I would suggest contacting the owner of the the docker image as this is a docker issue.

I will try that, thank you for your help.

The workaround we imagined is to create several docker instances with an orchestrator which can handle the repartition between CPU, but we haven't tried yet...