[mpich-discuss] MPICH v3.2 and SLRUM
Andreas Gocht
andreas.gocht at tu-dresden.de
Tue Dec 1 11:22:21 CST 2015
Hey
yeah mpiexec is working quite well. I'd just liked to use slurm, as our
implantation allows to set the cpu frequency on a node. Is there a way
to pass flags to srun using mpiexec?
Thanks four your help.
Kind Regards
Andreas
Am 01.12.2015 um 17:41 schrieb Balaji, Pavan:
> Looks like SLURM is telling MPICH that two processes are on the same node, even though they are on different nodes. It looks like a bug in the SLURM PMI implementation. Did you try simply using mpiexec instead? You'll need to remove the --with-pmi, --with-pm, and LDFLAGS/LIBS options and rebuild mpich for that. Note that mpiexec will internally use srun on slurm environments.
>
> -- Pavan
>
>> On Dec 1, 2015, at 5:56 AM, Andreas Gocht <andreas.gocht at tu-dresden.de> wrote:
>>
>> Hey
>>
>> I tried to build an use mpich with slurm, sbatch and srun. Unfortunate it looks like MPI_Init doesn't work with srun.
>>
>> I got the following error:
>>
>> Fatal error in MPI_Init: Other MPI error, error stack:
>> MPIR_Init_thread(474).................:
>> MPID_Init(190)........................: channel initialization failed
>> MPIDI_CH3_Init(89)....................:
>> MPID_nem_init(272)....................:
>> MPIDI_CH3I_Seg_commit(366)............:
>> MPIU_SHMW_Hnd_deserialize(324)........:
>> MPIU_SHMW_Seg_open(865)...............:
>> MPIU_SHMW_Seg_create_attach_templ(637): open failed - No such file or directory
>> In: PMI_Abort(4239887, Fatal error in MPI_Init: Other MPI error, error stack:
>> MPIR_Init_thread(474).................:
>> MPID_Init(190)........................: channel initialization failed
>> MPIDI_CH3_Init(89)....................:
>> MPID_nem_init(272)....................:
>> MPIDI_CH3I_Seg_commit(366)............:
>> MPIU_SHMW_Hnd_deserialize(324)........:
>> MPIU_SHMW_Seg_open(865)...............:
>> MPIU_SHMW_Seg_create_attach_templ(637): open failed - No such file or directory)
>>
>> I configured MPICH with "./configure --prefix=<some/prefix> --with-pmi=slurm --with-pm=none --with-slurm=<path/to/slurm>" and compiled my application with the "-L<path_to_slurm_lib> -lpmi" command.
>>
>> (as described in
>>
>> https://wiki.mpich.org/mpich/index.php/Frequently_Asked_Questions#Note_that_the_default_build_of_MPICH_will_work_fine_in_SLURM_environments._No_extra_steps_are_needed.
>>
>> and
>>
>> https://computing.llnl.gov/linux/slurm/mpi_guide.html#mpich2
>>
>> )
>>
>> I am running with 10 nodes and one task per node. Is there something I am missing during the configuration of MPICH?
>>
>> Best,
>>
>> Andreas
>>
>> --
>> M.Sc. Andreas Gocht
>>
>> Technische Universität Dresden
>> Center for Information Services and
>> High Performance Computing (ZIH)
>> D-01062 Dresden
>> Germany
>>
>> Contact:
>> Willersbau, Room A 104
>> Phone: (+49) 351 463-36415
>> Fax: (+49) 351 463-3773
>> e-mail: andreas.gocht at tu-dresden.de
>>
>>
>> _______________________________________________
>> discuss mailing list discuss at mpich.org
>> To manage subscription options or unsubscribe:
>> https://lists.mpich.org/mailman/listinfo/discuss
> _______________________________________________
> discuss mailing list discuss at mpich.org
> To manage subscription options or unsubscribe:
> https://lists.mpich.org/mailman/listinfo/discuss
--
M.Sc. Andreas Gocht
Technische Universität Dresden
Center for Information Services and
High Performance Computing (ZIH)
D-01062 Dresden
Germany
Contact:
Willersbau, Room A 104
Phone: (+49) 351 463-36415
Fax: (+49) 351 463-3773
e-mail: andreas.gocht at tu-dresden.de
-------------- next part --------------
A non-text attachment was scrubbed...
Name: smime.p7s
Type: application/pkcs7-signature
Size: 5149 bytes
Desc: S/MIME Cryptographic Signature
URL: <http://lists.mpich.org/pipermail/discuss/attachments/20151201/e4c2c92a/attachment.p7s>
-------------- next part --------------
_______________________________________________
discuss mailing list discuss at mpich.org
To manage subscription options or unsubscribe:
https://lists.mpich.org/mailman/listinfo/discuss
More information about the discuss
mailing list