Where communities thrive


  • Join over 1.5M+ people
  • Join over 100K+ communities
  • Free without limits
  • Create your own community
People
Activity
  • Jul 07 2017 06:38
    victorsndvg closed #7
  • Jul 07 2017 06:38
    victorsndvg commented #7
  • Jul 06 2017 20:30
    prudhomm commented #7
  • Jul 06 2017 20:29
    prudhomm closed #1
  • Apr 06 2017 10:07
    victorsndvg commented #1
  • Apr 06 2017 07:54
    prudhomm commented #1
  • Apr 06 2017 07:53
    prudhomm labeled #6
  • Apr 06 2017 07:53
    prudhomm labeled #6
  • Apr 06 2017 07:53
    prudhomm commented #6
  • Apr 06 2017 07:53
    prudhomm assigned #6
  • Apr 06 2017 07:52
    prudhomm labeled #8
  • Apr 06 2017 07:51
    prudhomm assigned #8
  • Apr 06 2017 07:51
    prudhomm labeled #8
  • Apr 06 2017 07:51
    prudhomm assigned #8
  • Apr 06 2017 07:51
    prudhomm opened #8
  • Mar 31 2017 13:14
    victorsndvg commented #7
  • Mar 31 2017 13:13
    victorsndvg commented #7
  • Mar 31 2017 13:13
    victorsndvg opened #7
  • Mar 29 2017 16:08
    victorsndvg commented #5
  • Mar 29 2017 16:08
    victorsndvg closed #5
victorsndvg
@victorsndvg
I think we can talk during this afternoon meeting
Guillaume Dollé
@gdolle
Yes!
I agree we need a native version. That's why user should be able to recompile Feel++ easily if needed ;)
Guillaume Dollé
@gdolle
@victorsndvg little problem when trying to recompile Feel++
$ ls $PETSC_DIR
ls: cannot access /opt/cesga/feelpp/0.102.00/gcc/6.3.0/openmpi/2.0.2/packages//petsc/3.7.3/: Permission denied
victorsndvg
@victorsndvg
ls /opt/cesga/feelpp/0.102.00/gcc/6.3.0/openmpi/2.0.2/packages//petsc/3.7.3/ works
maybe is a problem with the double slash
I already fixed it :+1: . Thanks @gdolle
Guillaume Dollé
@gdolle
@victorsndvg humm seems not ok
 ls -lsh /opt/cesga/feelpp/0.102.00/gcc/6.3.0/openmpi/2.0.2/packages/
total 0
0 drwxrwxr-- 3 software cesga 80 May  2 10:56 cln
0 drwxrwxr-- 3 software cesga 80 May  2 10:57 gmsh
0 drwxrwxr-- 3 software cesga 80 May  2 10:46 petsc
ls -lsh /opt/cesga/feelpp/0.102.00/gcc/6.3.0/openmpi/2.0.2/packages/petsc/
ls: reading directory /opt/cesga/feelpp/0.102.00/gcc/6.3.0/openmpi/2.0.2/packages/petsc/: Permission denied
Guillaume Dollé
@gdolle
@victorsndvg maybe you should create specific modules for packages in
/opt/cesga/feelpp/0.102.00/gcc/6.3.0/openmpi/2.0.2/packages,
to avoid recompiling them for each version of feel++
victorsndvg
@victorsndvg
But them are compiled explicitly for feelpp ... for example petsc must be compiled with a lot of dependencies in order to be used by other people at FT2
Guillaume Dollé
@gdolle
but it can be used by other people also no ?
you could do a module petsc-externals/3.7.5-externals
victorsndvg
@victorsndvg
I don't understand the last sentence
Guillaume Dollé
@gdolle
I mean it is not really feel++ specific cln, gmsh, and petsc
other user can benefit of these version

it is quite like

petsc/3.7.5-complex-debug
petsc/3.7.5-complex-opt
petsc/3.7.5-real-debug

the library are compiled with some specific deps
isn't it possible to just do

petsc/3.7.3-externals
petsc/3.7.3-feelpp
...

(or better naming ? I put external for external petsc packages)

victorsndvg
@victorsndvg
Ok, perfectly explained
it's possible
Guillaume Dollé
@gdolle
At least you would not have to recompile them for each petsc/gmsh/cln version
victorsndvg
@victorsndvg
we have to take a decision on how to do it for the next version
yes, is the same with feelpp. I'm looking for a solution in order to not to do it more times than needed
Guillaume Dollé
@gdolle
I guess for the MSO4SC project, these package version should not change
only the feelpp lib
the most important thing is that we can recompile feel++ if we require to patch the library for the pilots :)
victorsndvg
@victorsndvg
I think it could be a solution. But I prefer, if it is possible, to work with a particular stable release
users will not need to recompile anything if it's already installed
It's launch time. See you in the meeting. We can talk about it
Guillaume Dollé
@gdolle
Final users no, but pilots devs till the end of the project might have to recompile it. For example we have HDG developments that are not included yet and "should" be merged in the version 0.104.
We can discuss later about it. Have a good lunch ! :)
Guillaume Dollé
@gdolle
For info, singularity 2.3 is realeased :sparkles: !
victorsndvg
@victorsndvg
Great @gdolle ! we will try with it
Have you check this new version??

Hi @prudhomm ,

I would like to get access to the HiFiMagnet container. How can I get it?

Guillaume Dollé
@gdolle
@victorsndvg for singularity 2.3, it seems users can create image with the combo singularity create/import/mount (but not bootstrap)
victorsndvg
@victorsndvg
ouch!
victorsndvg
@victorsndvg

Hi @prudhomm , @gdolle , @Trophime ,
In Budapest we talk about testing the scalability of feel++ using Singularity containers.
I would like to:

  • know which benchmark can I use
  • know how to launch this benchmark
  • which kind of scalability (weak, strong) we are going to check

Can someone give me some lights on this?
Thanks in advance!

Christophe Prud'homme
@prudhomm
@victorsndvg , I think @vincentchabannes has already some benchmarks ready to be run with the toolbox. It provides some timing informations to do weak and strong scalability if I remember correctly. However we are a bit busy these days with a deadline next week and I don't know if these tools are readily available or it requires some updates. @vincentchabannes ?
victorsndvg
@victorsndvg

Thanks @prudhomm ,
Ok, don't worry. It's only a reminder. Let me know when you are ready to do this tests.

Good luck with the deadline!

Christophe Prud'homme
@prudhomm
@gdolle @vincentchabannes we need to provide the benchmarks setup for @victorsndvg . @gdolle you can start from the turek/hron testcase and add some information regarding what we discussed last week
Trophime
@Trophime

Hi, I've tried hifimagnet singularity image build with openmpi 2.0.2 on a Debian Jessie with openmpi 1.6.5. I've got the follwing error message:

It looks like MPI_INIT failed for some reason; your parallel process is
likely to abort.  There are many reasons that a parallel process can
fail during MPI_INIT; some of which are due to configuration or environment
problems.  This failure appears to be an internal failure; here's some
additional information (which may only be relevant to an Open MPI
developer):

  ompi_mpi_init: ompi_rte_init failed
  --> Returned "(null)" (-43) instead of "Success" (0)

@victorsndvg is this the kind of error you were talking about during our latest WP3 meeting?

Guillaume Dollé
@gdolle
@Trophime I think it's interesting to follow this thread
https://groups.google.com/a/lbl.gov/forum/#!topic/singularity/lQ6sWCWhIWY
It seems different openmpi version inside/outside does not go very well. @victorsndvg can confirm/explain maybe ?
@Trophime does your image work on froggy in // ?
Trophime
@Trophime
I've only tried so far on Debian stretch, the newest stable, here in the lab.
I can confirm that it working great when openmpi is at least the same inside/outside the image.
victorsndvg
@victorsndvg

hi @Trophime ,

Yes we checked that you need to use the same OpenMPI version inside and outside of the container to have compatibility

Tamás Budai
@budait
@victorsndvg I also did a bit of research and it looks like the only way to go is the same mpi inside and outside the container. This is bad news, since it makes our sing. containers less portable
for the time being we could work around this problem by generating containers for all mpi versions that may be used on the different infrasctructures
this sounds a bit dumb but if we gonna use a filesystem with deduplication for container storage (image repository for the marketplace) we can keep the overall size manageable, hence many files are the same in these images
victorsndvg
@victorsndvg

Hi @budait ,

is good to think in possible future choices to deal with portability.

We are also working on it.

This approach seems promising (fingers crossed) to integrate Singularity with, at least, FinisTerrae II Slurm version:

http://singularity.lbl.gov/tutorial-gpu-drivers-open-mpi-mtls

It also seems that there are plugin (for the latest slurm versions) to integrate Slurm with Singularity, but we cannot check it

Tamás Budai
@budait
why not? version problems?
victorsndvg
@victorsndvg
Yes, exactly