Steffen --

What version of Open MPI are you using?


> On Apr 16, 2019, at 9:21 AM, Steffen Christgau <christ...@cs.uni-potsdam.de> 
> wrote:
> 
> Hi Tim,
> 
> it helps, up to four processes. But it has two drawbacks. 1) Using more
> cores/threads than the machine provides  (so the actual
> over-subscription) is still not possible.  2) it still requires an
> additional command line argument.
> 
> What I'd like to see is that a call of mpirun with an arbitrary number
> of processes that just works without any other command line options.
> However, an environment variable would be acceptable.
> 
> MPICH's (v3.3) mpirun of a plain installation with no further configure
> options (I think it uses the Hydra PM) just does what I want, but MPICH
> is not always an option ;-)
> 
> Regards, Steffen
> 
> On 16/04/2019 14.56, Tim Jim wrote:
>> Hi Steffen,
>> 
>> I'm not sure if this will help you (I'm by far no expert) but the
>> mailing group pointed by to using: 
>> mpirun --use-hwthread-cpus
>> 
>> to solve something similar.
>> 
>> Kind regards,
>> Tim
>> 
>> 
>> On Tue, 16 Apr 2019 at 19:01, Steffen Christgau
>> <christ...@cs.uni-potsdam.de <mailto:christ...@cs.uni-potsdam.de>> wrote:
>> 
>>    Hi everyone,
>> 
>>    on my 2 cores/4 threads development platform I want to start programs
>>    via mpirun with over-subscription enabled by default. I have some
>>    external packages which have tests that use more than 2 processes. They
>>    all fail because Open MPI refuses them to run due to over-subscription.
>>    I know that I can enable over-subscription with mpirun --oversubscribe
>>    and that works well, but that would require to modify the packages'
>>    autotools files or the generated Makefiles with I found hardly
>>    convenient.
>> 
>>    I also tried two of the RMAPS MCA parameters:
>> 
>>     - rmaps_base_no_oversubscribe
>>     - rmaps_base_oversubscribe
>> 
>>    (btw, are they redundant? Having two of them and one is the negation of
>>    the other is quite confusing. The description in ompi_info reads quite
>>    similar for the two.)
>> 
>>    $ mpirun --mca rmaps_base_no_oversubscribe 0 --mca
>>    rmaps_base_oversubscribe 1 -n 4 hostname
>>    --------------------------------------------------------------------------
>>    There are not enough slots available in the system to satisfy the 4
>>    slots that were requested by the application:
>>      hostname
>> 
>>    Either request fewer slots for your application, or make more slots
>>    available for use.
>>    --------------------------------------------------------------------------
>> 
>>    Setting the environment variables (OMPI_MCA_rmaps_...) did not help
>>    either. The same goes or the FAQ [1]:
>> 
>>    $ cat > my-hostfile
>>    localhost
>>    $ mpirun -np 4 --hostfile my-hostfile hostname
>>    --------------------------------------------------------------------------
>>    There are not enough slots available in the system to satisfy the 4
>>    slots...
>> 
>>    My Open MPI installation is a private build of version 4.0.0, configured
>>    with nothing more than ./configure --prefix=/some/where
>>    --enable-mpi-fortran
>> 
>>    How can I allow over-subscription by default? I am aware about
>>    performance implications, but it's only used for testing and
>>    development. I am not using a resource manager on that machine.
>> 
>>    Thanks in advance
>> 
>>    Regards, Steffen
>> 
>>    [1] https://www.open-mpi.org/faq/?category=running#oversubscribing
>> 
>>    _______________________________________________
>>    users mailing list
>>    users@lists.open-mpi.org <mailto:users@lists.open-mpi.org>
>>    https://lists.open-mpi.org/mailman/listinfo/users
>> 
>> 
>> 
>> -- 
>> 
>>      
>> 
>> *Timothy Jim
>> */PhD Researcher in Aerospace/
>> 
>> Creative Flow Research Division,
>> Institute of Fluid Science, Tohoku University
>> 
>> www.linkedin.com/in/timjim/ <http://www.linkedin.com/in/timjim/>
>> 
>> 
>> _______________________________________________
>> users mailing list
>> users@lists.open-mpi.org
>> https://lists.open-mpi.org/mailman/listinfo/users
>> 
> _______________________________________________
> users mailing list
> users@lists.open-mpi.org
> https://lists.open-mpi.org/mailman/listinfo/users


-- 
Jeff Squyres
jsquy...@cisco.com

_______________________________________________
users mailing list
users@lists.open-mpi.org
https://lists.open-mpi.org/mailman/listinfo/users

Reply via email to