Thanks, Ralph,
The .../etc/mca-params.conf doesn't want the shell version with the
export and OMPI_MCA_ prefix, does it?
$ tail -3 $MPI_HOME/etc/openmpi-mca-params.conf
# See "ompi_info --param all all" for a full listing of Open MPI MCA
# parameters available and their default values.
orte_hetero_nodes=1
Yes, it appears the man page may have been outdated, as ompi_info -a shows:
MCA hwloc: parameter "hwloc_base_binding_policy" (current value: "",
data source: default, level: 9 dev/all, type: string)
Policy for binding processes [none |
hwthread | core (default) | l1cache | l2cache | l3cache | socket |
numa | board] (supported qualifiers: overload-allowed,if-supported)
and the default is hwloc_base_binding_policy=core, I believe.
Thanks, again, and sorry to be dense.
-- bennet
On Mon, Jan 11, 2016 at 9:39 AM, Ralph Castain <r...@open-mpi.org> wrote:
> For the 1.10 series, putting "export
> OMPI_MCA_hwloc_base_binding_policy=none” into your default MCA param file
> will solve the problem. I believe that is true for all of the 1.8 series as
> well, and suspect the man page for 1.8.2 was simply out-of-date. You could
> verify that if you are using something that old.
>
>
>
>> On Jan 11, 2016, at 5:32 AM, Bennet Fauber <ben...@umich.edu> wrote:
>>
>> We have an issue with binding to cores with some applications and the
>> default causes issues. We would, therefore, like to set the
>> equivalent of
>>
>> mpirun --bind-to none
>>
>> globally. I tried search for combinations of 'openmpi global
>> settings', 'site settings', and the like on the web and ended up
>> several times at
>>
>> https://www.open-mpi.org/faq/?category=sysadmin#sysadmin-mca-params
>>
>> That makes it look very much like MCA parameters are for network
>> settings; see, specifically, section 4. What are MCA Parameters? Why
>> would I set them?
>>
>> At some point, though, the mpirun man page,
>> https://www.open-mpi.org/doc/v1.8/man1/mpirun.1.php, where at the end
>> of the section titled, Mapping, Ranking, and Binding: Oh My!, it says:
>>
>> -----------------------------
>> Process binding can also be set with MCA parameters. Their usage is
>> less convenient than that of mpirun options. On the other hand, MCA
>> parameters can be set not only on the mpirun command line, but
>> alternatively in a system or user mca-params.conf file or as
>> environment variables, as described in the MCA section below. Some
>> examples include:
>>
>> mpirun option MCA parameter key value
>>
>> --map-by core rmaps_base_mapping_policy core
>> . . . .
>> --bind-to none hwloc_base_binding_policy none
>> -----------------------------
>>
>> Am I correct in interpreting this to mean that, if I
>>
>> export OMPI_MCA_hwloc_base_binding_policy=none
>>
>> from the module file, the default binding will be 'none'?
>>
>> Equivalently, if I add a line to /ompi/install/path/etc/mca-params.conf
>>
>> -----
>> hwloc_base_binding_policy = none
>> -----
>>
>> that would do the same?
>>
>> The web version of the man page is for 1.8.8, and it agrees with the
>> installed man page for our 1.8.7. However, it appears that our system
>> man page for mpirun(1) for OpenMPI 1.8.2 has slightly different
>> parameters. Specifically,
>>
>> Process binding can also be set with MCA parameters. Their usage is
>> less convenient than that of mpirun options. On the other hand, MCA
>> parameters can be set not only on the mpirun command line, but alterna-
>> tively in a system or user mca-params.conf file or as environment vari-
>> ables, as described in the MCA section below. The correspondences are:
>>
>> mpirun option MCA parameter key value
>>
>> -bycore rmaps_base_schedule_policy core
>> -bysocket rmaps_base_schedule_policy socket
>> -bind-to-core orte_process_binding core
>> -bind-to-socket orte_process_binding socket
>> -bind-to-none orte_process_binding none
>>
>> So for version 1.8.2, the equivalent incantations would be
>>
>> export OMPI_MCA_orte_process_binding=none
>>
>> or
>>
>> /ompi/install/path/v1.8.2/etc/mca-params.conf
>> -----
>> orte_process_binding = none
>> -----
>>
>> Yes?
>>
>> Sorry to be dense about this.
>>
>> Thanks, -- bennet
>> _______________________________________________
>> users mailing list
>> us...@open-mpi.org
>> Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/users
>> Link to this post:
>> http://www.open-mpi.org/community/lists/users/2016/01/28243.php
>
> _______________________________________________
> users mailing list
> us...@open-mpi.org
> Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/users
> Link to this post:
> http://www.open-mpi.org/community/lists/users/2016/01/28244.php
