Open MPI logo

Open MPI User's Mailing List Archives

  |   Home   |   Support   |   FAQ   |   all Open MPI User's mailing list

Subject: Re: [OMPI users] default num_procs of round_robin_mapper with cpus-per-proc option
From: tmishima_at_[hidden]
Date: 2014-01-23 00:37:07


Thanks for your explanation, Ralph.

But it's really subtle to understand for me ...
Anyway, I'd like to report what I found through verbose output.

"-map-by core" calls "bind in place" as below:
[mishima_at_manage work]$ mpirun -np 4 -hostfile pbs_hosts -report-bindings
-cpus-per-proc 4 -map-by core -mca rmaps_base_v
erbose 10 ~/mis/openmpi/demos/myprog
...
[manage.cluster:11362] mca:rmaps: compute bindings for job [8729,1] with
policy CORE
[manage.cluster:11362] mca:rmaps: bindings for job [8729,1] - core to core
[manage.cluster:11362] mca:rmaps: bind in place for job [8729,1] with
bindings CORE
...

On the other hand, "-map-by slot" calls "bind downward" as below:
[mishima_at_manage work]$ mpirun -np 4 -hostfile pbs_hosts -report-bindings
-cpus-per-proc 4 -map-by slot -mca rmaps_base_v
erbose 10 ~/mis/openmpi/demos/myprog
...
[manage.cluster:12032] mca:rmaps: compute bindings for job [8571,1] with
policy CORE
[manage.cluster:12032] mca:rmaps: bind downward for job [8571,1] with
bindings CORE
...

I think your best guess is right and something is wrong with
bind_in_place function. I have to say the logic of source code
is so complex that I could not figure it out.

Regards,
Tetsuya Mishima

> On Jan 22, 2014, at 8:08 PM, tmishima_at_[hidden] wrote:
>
> >
> >
> > Thanks, Ralph.
> >
> > I have one more question. I'm sorry to ask you many things ...
>
> Not a problem
>
> >
> > Could you tell me the difference between "map-by slot" and "map-by
core".
> > From my understanding, slot is the synonym of core.
>
> Not really - see below
>
> > But those behaviors
> > using openmpi-1.7.4rc2 with the cpus-per-proc option are quite
different
> > as shown below. I tried to browse the source code but I could not make
it
> > clear so far.
> >
>
> It is a little subtle, I fear. When you tell us "map-by slot", we assign
each process to an allocated slot without associating it to any specific
cpu or core. When we then bind to core (as we do by
> default), we balance the binding across the sockets to improve
performance.
>
> When you tell us "map-by core", then we directly associate each process
with a specific core. So when we bind, we bind you to that core. This will
cause us to fully use all the cores on the first
> socket before we move to the next.
>
> I'm a little puzzled by your output as it appears that cpus-per-proc was
ignored, so that's something I'd have to look at more carefully. Best guess
is that we aren't skipping cores to account for
> the cpus-per-core setting, and thus the procs are being mapped to
consecutive cores - which wouldn't be very good if we then bound them to
multiple neighboring cores as they'd fall on top of each
> other.
>
>
> > Regards,
> > Tetsuya Mishima
> >
> > [ un-managed environment] (node05,06 has 8 cores each)
> >
> > [mishima_at_manage work]$ cat pbs_hosts
> > node05
> > node05
> > node05
> > node05
> > node05
> > node05
> > node05
> > node05
> > node06
> > node06
> > node06
> > node06
> > node06
> > node06
> > node06
> > node06
> > [mishima_at_manage work]$ mpirun -np 4 -hostfile pbs_hosts
-report-bindings
> > -cpus-per-proc 4 -map-by slot ~/mis/openmpi/dem
> > os/myprog
> > [node05.cluster:23949] MCW rank 1 bound to socket 1[core 4[hwt 0]],
socket
> > 1[core 5[hwt 0]], socket 1[core 6[hwt 0]], so
> > cket 1[core 7[hwt 0]]: [./././.][B/B/B/B]
> > [node05.cluster:23949] MCW rank 0 bound to socket 0[core 0[hwt 0]],
socket
> > 0[core 1[hwt 0]], socket 0[core 2[hwt 0]], so
> > cket 0[core 3[hwt 0]]: [B/B/B/B][./././.]
> > [node06.cluster:22139] MCW rank 3 bound to socket 1[core 4[hwt 0]],
socket
> > 1[core 5[hwt 0]], socket 1[core 6[hwt 0]], so
> > cket 1[core 7[hwt 0]]: [./././.][B/B/B/B]
> > [node06.cluster:22139] MCW rank 2 bound to socket 0[core 0[hwt 0]],
socket
> > 0[core 1[hwt 0]], socket 0[core 2[hwt 0]], so
> > cket 0[core 3[hwt 0]]: [B/B/B/B][./././.]
> > Hello world from process 0 of 4
> > Hello world from process 1 of 4
> > Hello world from process 3 of 4
> > Hello world from process 2 of 4
> > [mishima_at_manage work]$ mpirun -np 4 -hostfile pbs_hosts
-report-bindings
> > -cpus-per-proc 4 -map-by core ~/mis/openmpi/dem
> > os/myprog
> > [node05.cluster:23985] MCW rank 1 bound to socket 0[core 1[hwt 0]]:
> > [./B/./.][./././.]
> > [node05.cluster:23985] MCW rank 0 bound to socket 0[core 0[hwt 0]]:
> > [B/././.][./././.]
> > [node06.cluster:22175] MCW rank 3 bound to socket 0[core 1[hwt 0]]:
> > [./B/./.][./././.]
> > [node06.cluster:22175] MCW rank 2 bound to socket 0[core 0[hwt 0]]:
> > [B/././.][./././.]
> > Hello world from process 2 of 4
> > Hello world from process 3 of 4
> > Hello world from process 0 of 4
> > Hello world from process 1 of 4
> >
> > (note) I have the same behavior in the managed environment by Torque
> >
> >> Seems like a reasonable, minimal risk request - will do
> >>
> >> On Jan 22, 2014, at 4:28 PM, tmishima_at_[hidden] wrote:
> >>
> >>>
> >>> Hi Ralph, I want to ask you one more thing about default setting of
> >>> num_procs
> >>> when we don't specify the -np option and we set the cpus-per-proc >
1.
> >>>
> >>> In this case, the round_robin_mapper sets num_procs = num_slots as
> > below:
> >>>
> >>> rmaps_rr.c:
> >>> 130 if (0 == app->num_procs) {
> >>> 131 /* set the num_procs to equal the number of slots on
> > these
> >>> mapped nodes */
> >>> 132 app->num_procs = num_slots;
> >>> 133 }
> >>>
> >>> However, because of cpus_per_rank > 1, this num_procs will be refused
> > at
> >>> the
> >>> line 61 in rmaps_rr_mappers.c as below, unless we switch on the
> >>> oversubscribe
> >>> directive.
> >>>
> >>> rmaps_rr_mappers.c:
> >>> 61 if (num_slots < ((int)app->num_procs *
> >>> orte_rmaps_base.cpus_per_rank)) {
> >>> 62 if (ORTE_MAPPING_NO_OVERSUBSCRIBE &
> > ORTE_GET_MAPPING_DIRECTIVE
> >>> (jdata->map->mapping)) {
> >>> 63 orte_show_help("help-orte-rmaps-base.txt",
> >>> "orte-rmaps-base:alloc-error",
> >>> 64 true, app->num_procs, app->app);
> >>> 65 return ORTE_ERR_SILENT;
> >>> 66 }
> >>> 67 }
> >>>
> >>> Therefore, I think the default num_procs should be equal to the
number
> > of
> >>> num_slots divided by cpus/rank:
> >>>
> >>> app->num_procs = num_slots / orte_rmaps_base.cpus_per_rank;
> >>>
> >>> This would be more convinient for most of people who want to use the
> >>> -cpus-per-proc option. I already confirmed it worked well. Please
> > consider
> >>> to apply this fix to 1.7.4.
> >>>
> >>> Regards,
> >>> Tetsuya Mishima
> >>>
> >>> _______________________________________________
> >>> users mailing list
> >>> users_at_[hidden]
> >>> http://www.open-mpi.org/mailman/listinfo.cgi/users
> >>
> >> _______________________________________________
> >> users mailing list
> >> users_at_[hidden]
> >> http://www.open-mpi.org/mailman/listinfo.cgi/users
> >
> > _______________________________________________
> > users mailing list
> > users_at_[hidden]
> > http://www.open-mpi.org/mailman/listinfo.cgi/users
>
> _______________________________________________
> users mailing list
> users_at_[hidden]
> http://www.open-mpi.org/mailman/listinfo.cgi/users