Welcome to the forums. Please post in English or French.

You are not logged in.

#1 2020-07-02 10:33:36

bpaillard
Member
Registered: 2015-10-16
Posts: 21

OpenTURNS parallel evaluations

Hi all,

I'm starting to use the very promising tool OpenTURNS, going great so far.

I'd like to be able to parallelize function evaluations, which I guess should be provided by the option "Number of parallel evaluations"

Unfortunately this options does not look functional when I try to use it. And changing # of cores or # of nodes in advanced parameters does not seem to work either.

Has anybody managed to use this parallelization ?

Offline

#2 2020-07-09 10:57:03

Irvise
Member
Registered: 2019-12-07
Posts: 9

Re: OpenTURNS parallel evaluations

Hi bpaillard,

I tried to do the exact same thing and I have gotten the exact same result as you. Neither variable seemed to do what it indicates as you indicate.

I suppose that the developers are aware of this issue, since it is pretty obvious. With the time I have had with OT, I can tell it is quite rough around the edges (it has some very basic issues such as not being able to detect that a point has already been run).

Do you have any experience with YACS? After all, it seems it is the system that OT is using to run the simulation (at the very least the Python case has to be set up as such).

I will keep playing with it. Regards,

Fernando

EDIT: I can see that several python3 processes are spawned if I tell it to do a parallel evaluation. However, only one seems active

EDIT 2: after trying to run a design of experiments I go this error at the very end. The Python file it is pointing to, may have some answers. It also confirms it is a pure YACS simulation that is controlled form OT. Here is the error:

No results are available. An error has occurred during the execution of the analysis.
An error occured while fetching the results.
Failed to execute python function.
An error occured during the execution of the YACS schema.
The results you get may be incomplete or incorrect.
For further details, see /tmp/salome_localres_workdir_fernando/logs directory on localhost.
Traceback (most recent call last):
  File "/home/fernando/salome_meca/appli_V2019.0.3_universal/lib/python3.6/site-packages/salome/pydefx/pystudy.py", line 260, in getResult
    raise Exception(errorMessage)
Exception: An error occured during the execution of the YACS schema.
The results you get may be incomplete or incorrect.
For further details, see /tmp/salome_localres_workdir_fernando/logs directory on localhost

Last edited by Irvise (2020-07-09 11:48:30)

Offline

#3 2020-07-09 13:10:38

GPSalachs
Member
Registered: 2018-03-10
Posts: 185

Re: OpenTURNS parallel evaluations

Hello,

i am interested in the discussion because i am getting more or less the same errors . I do get these errors on Salome Meca 2019. Are you too using the 2019 version too?

Offline

#4 2020-07-09 17:28:19

Irvise
Member
Registered: 2019-12-07
Posts: 9

Re: OpenTURNS parallel evaluations

After playing with OT for a while, I have realized that there is no parallel execution under a "normal" simulation case. By "normal" I refer to a direct OT model (no python script, no YACS model).

When creating a OT case from a Python script there is a "multiprocessing" option when defining the model. It does not make my cpus go to 100% but it makes them all go to about 50%. This last cpu usage is very similar to the one I got by using the "normal" mode. This is probably due to my model being too simple and not letting the cpu be used at its max.

I would suppose then that OT does not have a parallel option by default and it is only added to the YACS module. This option may then refer to "network" parallelization (mpi) and not local cpus. However, mpi also works locally, which invalidates this theory. It is probably that it may not be set up correctly.

Parallel computations in OT may then be obtained by using extra internal parameter under the "advanced parameter" tab. However, this is just a guess, since I could not find any info on "how to make OT run cases in parallel".

Regards,

Fernando

EDIT: and yes, I am using the 2019 version

Last edited by Irvise (2020-07-09 17:28:52)

Offline