It seems this is partly a problem with specifying the 'slots' keyword. If I
specify the number of cores in the binding, as opposed to 'slots', then I
see the binding displayed.

qsub -binding linear:28 ...

qstat -j JOB_ID
binding:                    set linear:28
job_type:                   NONE
usage         1:            cpu=07:49:55, mem=108064.47512 GB s, io=0.09399
GB, vmem=550.511G, maxvmem=556.599G
binding       1:
 
ScttcttcttcttcttcttcttcttcttcttcttcttcttcttScttcttcttcttcttcttcttcttcttcttcttcttcttctt

But that does not help if the number of cores used on the hosts are not all
the same.

Glenn

On Sat, Feb 25, 2017 at 9:52 PM, Glenn Johnson <glenn-john...@uiowa.edu>
wrote:

> I have a question about core binding behavior. I have SoGE version 8.1.9
> installed on a CentOS-7.3 system. I have
>
> ENABLE_BINDING=true
>
> in the execd_params and
>
> -binding linear:slots
>
> in the sge_request file.
>
> For jobs on a single node I see the core binding when looking at the
> output of qstat -j JOB_ID. However, when jobs run across nodes, it appears
> that SGE is not providing any core binding information. I am using
> openmpi-2.0.1 and have the following set in the openmpi-mca-params.conf
> file.
>
> rmaps_base_mapping_policy = node
> hwloc_base_binding_policy = core
>
> When using the --report-bindings option to mpirun I can see that openmpi
> is binding as expected. However, I thought that SGE would set the
> SGE_BINDING environment variable but it does not. Also, when looking at the
> qstat -j output the binding is listed as NONE.
>
> ...
> binding:                    set linear:slots
> job_type:                   NONE
> usage         1:            cpu=45:40:47, mem=637700.64073 GB s,
> io=0.11017 GB, vmem=546.217G, maxvmem=552.417G
> binding       1:            NONE
> ...
>
> Is there something else that needs to be done when setting core binding
> across nodes? If I am understanding, openmpi will not be able to honor the
> SGE core binding hints as there is no SGE_BINDING variable set.
>
> Thanks.
>
> Glenn
>
_______________________________________________
SGE-discuss mailing list
SGE-discuss@liv.ac.uk
https://arc.liv.ac.uk/mailman/listinfo/sge-discuss

Reply via email to