Open MPI logo

Open MPI User's Mailing List Archives

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

Subject: Re: [OMPI users] Problem with mpiexec --cpus-per-proc in multiple nodes in OMPI 1.6.4
From: Ralph Castain (rhc_at_[hidden])
Date: 2013-03-29 09:33:02


Just an update: I have this fixed in the OMPI trunk. It didn't make 1.7.0, but will be in 1.7.1 and beyond.

On Mar 21, 2013, at 2:09 PM, Gus Correa <gus_at_[hidden]> wrote:

> Thank you, Ralph.
>
> I will try to use a rankfile.
>
> In any case, the --cpus-per-proc option is a very useful feature:
> for hybrid MPI+OpenMP programs, for these processors with one FPU
> shared by two cores, etc.
> If it gets fixed in a later release of OMPI that would be great.
>
> Thank you,
> Gus Correa
>
>
> On 03/21/2013 04:03 PM, Ralph Castain wrote:
>> I've heard this from a couple of other sources -
> it looks like there is a problem on the daemons when
> they compute the location for -cpus-per-proc.
> I'm not entirely sure why that would be as the code
> is supposed to be common with mpirun, but there are
> a few differences.
>>
>> I will take a look at it - I don't know of any workaround,
> I'm afraid.
>>
>> On Mar 21, 2013, at 12:01 PM, Gus Correa<gus_at_[hidden]> wrote:
>>
>>> Dear Open MPI Pros
>>>
>>> I am having trouble using mpiexec with --cpus-per-proc
>>> on multiple nodes in OMPI 1.6.4.
>>>
>>> I know there is an ongoing thread on similar runtime issues
>>> of OMPI 1.7.
>>> By no means I am trying to hijack T. Mishima's questions.
>>> My question is genuine, though, and perhaps related to his.
>>>
>>> I am testing a new cluster remotely, with monster
>>> dual socket 16-core AMD Bulldozer processors (32 cores per node).
>>> I am using OMPI 1.6.4 built with Torque 4.2.1 support.
>>>
>>> I read that on these processors each pair of cores share an FPU.
>>> Hence, I am trying to run *one MPI process* on each
>>> *pair of successive cores*.
>>> This trick seems to yield better performance
>>> (at least for HPL/Linpack) than using all cores.
>>> I.e., the goal is to use "each other core", or perhaps
>>> to allow each process to wobble across two successive cores only,
>>> hence granting exclusive use of one FPU per process.
>>> [BTW, this is *not* an attempt to do hybrid MPI+OpenMP.
>>> The code is HPL with MPI+BLAS/Lapack and NO OpenMP.]
>>>
>>> To achieve this, I am using the mpiexec --cpus-per-proc option.
>>> It works on one node, which is great.
>>> However, unless I made a silly syntax or arithmetic mistake,
>>> it doesn't seem to work on more than one node.
>>>
>>> For instance, this works:
>>>
>>> #PBS -l nodes=1:ppn=32
>>> ...
>>> mpiexec -np 16 \
>>> --cpus-per-proc 2 \
>>> --bind-to-core \
>>> --report-bindings \
>>> --tag-output \
>>>
>>> I get a pretty nice process-to-cores distribution, with 16 processes, and each process bound to a couple of successive cores,
>>> as expected:
>>>
>>> [1,7]<stderr>:[node33:04744] MCW rank 7 bound to socket 0[core 14-15]: [. . . . . . . . . . . . . . B B][. . . . . . . . . . . . . . . .]
>>> [1,8]<stderr>:[node33:04744] MCW rank 8 bound to socket 1[core 0-1]: [. . . . . . . . . . . . . . . .][B B . . . . . . . . . . . . . .]
>>> [1,9]<stderr>:[node33:04744] MCW rank 9 bound to socket 1[core 2-3]: [. . . . . . . . . . . . . . . .][. . B B . . . . . . . . . . . .]
>>> [1,10]<stderr>:[node33:04744] MCW rank 10 bound to socket 1[core 4-5]: [. . . . . . . . . . . . . . . .][. . . . B B . . . . . . . . . .]
>>> [1,11]<stderr>:[node33:04744] MCW rank 11 bound to socket 1[core 6-7]: [. . . . . . . . . . . . . . . .][. . . . . . B B . . . . . . . .]
>>> [1,12]<stderr>:[node33:04744] MCW rank 12 bound to socket 1[core 8-9]: [. . . . . . . . . . . . . . . .][. . . . . . . . B B . . . . . .]
>>> [1,13]<stderr>:[node33:04744] MCW rank 13 bound to socket 1[core 10-11]: [. . . . . . . . . . . . . . . .][. . . . . . . . . . B B . . . .]
>>> [1,14]<stderr>:[node33:04744] MCW rank 14 bound to socket 1[core 12-13]: [. . . . . . . . . . . . . . . .][. . . . . . . . . . . . B B . .]
>>> [1,15]<stderr>:[node33:04744] MCW rank 15 bound to socket 1[core 14-15]: [. . . . . . . . . . . . . . . .][. . . . . . . . . . . . . . B B]
>>> [1,0]<stderr>:[node33:04744] MCW rank 0 bound to socket 0[core 0-1]: [B B . . . . . . . . . . . . . .][. . . . . . . . . . . . . . . .]
>>> [1,1]<stderr>:[node33:04744] MCW rank 1 bound to socket 0[core 2-3]: [. . B B . . . . . . . . . . . .][. . . . . . . . . . . . . . . .]
>>> [1,2]<stderr>:[node33:04744] MCW rank 2 bound to socket 0[core 4-5]: [. . . . B B . . . . . . . . . .][. . . . . . . . . . . . . . . .]
>>> [1,3]<stderr>:[node33:04744] MCW rank 3 bound to socket 0[core 6-7]: [. . . . . . B B . . . . . . . .][. . . . . . . . . . . . . . . .]
>>> [1,4]<stderr>:[node33:04744] MCW rank 4 bound to socket 0[core 8-9]: [. . . . . . . . B B . . . . . .][. . . . . . . . . . . . . . . .]
>>> [1,5]<stderr>:[node33:04744] MCW rank 5 bound to socket 0[core 10-11]: [. . . . . . . . . . B B . . . .][. . . . . . . . . . . . . . . .]
>>> [1,6]<stderr>:[node33:04744] MCW rank 6 bound to socket 0[core 12-13]: [. . . . . . . . . . . . B B . .][. . . . . . . . . . . . . . . .]
>>>
>>>
>>> ***************
>>>
>>> However, when I try to use eight nodes,
>>> the job fails and I get the error message below (repeatedly from
>>> several nodes):
>>>
>>> #PBS -l nodes=8:ppn=32
>>> ...
>>> mpiexec -np 128 \
>>> --cpus-per-proc 2 \
>>> --bind-to-core \
>>> --report-bindings \
>>> --tag-output \
>>>
>>>
>>> Error message:
>>>
>>> --------------------------------------------------------------------------
>>> An invalid physical processor ID was returned when attempting to bind
>>> an MPI process to a unique processor on node:
>>>
>>> Node: node18
>>>
>>> This usually means that you requested binding to more processors than
>>> exist (e.g., trying to bind N MPI processes to M processors, where N>
>>> M), or that the node has an unexpectedly different topology.
>>>
>>> Double check that you have enough unique processors for all the
>>> MPI processes that you are launching on this host, and that all nodes
>>> have identical topologies.
>>>
>>> You job will now abort.
>>> --------------------------------------------------------------------------
>>>
>>> Oddly enough, the binding map *is* shown on STDERR,
>>> and it sounds *correct*, pretty much the same binding map above
>>> that I get for a single node.
>>>
>>> *****************
>>>
>>> Finally, replacing "--cpus-per-core 2" by "--npernode 16"
>>> works to some extent, but doesn't reach my goal.
>>> I.e., the job doesn't fail, and each node gets 16 MPI
>>> processes indeed.
>>> However, it doesn't bind the processes the way I want.
>>> Regardless of whether I continue to use "--bind-to-core"
>>> or replace it by "--bind-to-socket"
>>> all 16 processes on each node always bind to socket 0,
>>> and nothing goes to socket 1.
>>>
>>> ************
>>>
>>> Is there any simple workaround to this
>>> (other than using a --rankfile),
>>> to make --cpus-per-proc work with multiple nodes,
>>> using "each other core"?
>>>
>>> [Only if it is simple workaround. I must finish this
>>> remote test soon. Otherwise I can revisit this issue later.]
>>>
>>> Thank you,
>>> Gus Correa
>>>
>>>
>>>
>>>
>>>
>>> _______________________________________________
>>> 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