Open MPI logo

Open MPI User's Mailing List Archives

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

Subject: Re: [OMPI users] Problems on large clusters
From: Gilbert Grosdidier (Gilbert.Grosdidier_at_[hidden])
Date: 2011-06-22 05:25:14


Bonjour Thorsten,

  I'm not surprised about the cluster type, indeed,
but I do not remember getting such specific hang up you mention.

  Anyway, I suspect SGI Altix is a little bit special for OpenMPI,
and I usually run with the following setup:
- there is need to create for each job a specific tmp area,
like "/scratch/ggg/uuu/run/tmp/pbs.${PBS_JOBID}"
- then use something like that:

setenv TMPDIR "/scratch/ggg/uuu/run/tmp/pbs.${PBS_JOBID}"
setenv OMPI_PREFIX_ENV "/scratch/ggg/uuu/run/tmp/pbs.${PBS_JOBID}"
setenv OMPI_MCA_mpi_leave_pinned_pipeline 1

- then, for running, many of these -mca options are probably useless
with your app,
while many of them may show to be useful. Your own way ...

mpiexec -mca coll_tuned_use_dynamic_rules 1 -hostfile $PBS_NODEFILE -
mca rmaps seq -mca btl_openib_rdma_pipeline_send_length 65536 -mca
btl_openib_rdma_pipeline_frag_size 65536 -mca
btl_openib_min_rdma_pipeline_size 65536 -mca
btl_self_rdma_pipeline_send_length 262144 -mca
btl_self_rdma_pipeline_frag_size 262144 -mca plm_rsh_num_concurrent
4096 -mca mpi_paffinity_alone 1 -mca mpi_leave_pinned_pipeline 1 -mca
btl_sm_max_send_size 128 -mca
coll_tuned_pre_allocate_memory_comm_size_limit 1048576 -mca
btl_openib_cq_size 128 -mca btl_ofud_rd_num 128 -mca
mpi_preconnect_mpi 0 -mca mpool_sm_min_size 131072 -mca btl
sm,openib,self -mca btl_openib_want_fork_support 0 -mca
opal_set_max_sys_limits 1 -mca osc_pt2pt_no_locks 1 -mca
osc_rdma_no_locks 1 YOUR_APP

  (Watch the step : only one line only ...)

  This should be suitable for up to 8k cores.

  HTH, Best, G.

Le 22 juin 11 à 09:13, Thorsten Schuett a écrit :

> Sure. It's an SGI ICE cluster with dual-rail IB. The HCAs are Mellanox
> ConnectX IB DDR.
>
> This is a 2040 cores job. I use 255 nodes with one MPI task on each
> node and
> use 8-way OpenMP.
>
> I don't need -np and -machinefile, because mpiexec picks up this
> information
> from PBS.
>
> Thorsten
>
> On Tuesday, June 21, 2011, Gilbert Grosdidier wrote:
>> Bonjour Thorsten,
>>
>> Could you please be a little bit more specific about the cluster
>> itself ?
>>
>> G.
>>
>> Le 21 juin 11 à 17:46, Thorsten Schuett a écrit :
>>> Hi,
>>>
>>> I am running openmpi 1.5.3 on a IB cluster and I have problems
>>> starting jobs
>>> on larger node counts. With small numbers of tasks, it usually
>>> works. But now
>>> the startup failed three times in a row using 255 nodes. I am using
>>> 255 nodes
>>> with one MPI task per node and the mpiexec looks as follows:
>>>
>>> mpiexec --mca btl self,openib --mca mpi_leave_pinned 0 ./a.out
>>>
>>> After ten minutes, I pulled a stracktrace on all nodes and killed
>>> the job,
>>> because there was no progress. In the following, you will find the
>>> stack trace
>>> generated with gdb thread apply all bt. The backtrace looks
>>> basically the same
>>> on all nodes. It seems to hang in mpi_init.
>>>
>>> Any help is appreciated,
>>>
>>> Thorsten
>>>
>>> Thread 3 (Thread 46914544122176 (LWP 28979)):
>>> #0 0x00002b6ee912d9a2 in select () from /lib64/libc.so.6
>>> #1 0x00002b6eeabd928d in service_thread_start (context=<value
>>> optimized out>)
>>> at btl_openib_fd.c:427
>>> #2 0x00002b6ee835e143 in start_thread () from /lib64/
>>> libpthread.so.0
>>> #3 0x00002b6ee9133b8d in clone () from /lib64/libc.so.6
>>> #4 0x0000000000000000 in ?? ()
>>>
>>> Thread 2 (Thread 46916594338112 (LWP 28980)):
>>> #0 0x00002b6ee912b8b6 in poll () from /lib64/libc.so.6
>>> #1 0x00002b6eeabd7b8a in btl_openib_async_thread (async=<value
>>> optimized
>>> out>) at btl_openib_async.c:419
>>> #2 0x00002b6ee835e143 in start_thread () from /lib64/
>>> libpthread.so.0
>>> #3 0x00002b6ee9133b8d in clone () from /lib64/libc.so.6
>>> #4 0x0000000000000000 in ?? ()
>>>
>>> Thread 1 (Thread 47755361533088 (LWP 28978)):
>>> #0 0x00002b6ee9133fa8 in epoll_wait () from /lib64/libc.so.6
>>> #1 0x00002b6ee87745db in epoll_dispatch (base=0xb79050,
>>> arg=0xb558c0,
>>> tv=<value optimized out>) at epoll.c:215
>>> #2 0x00002b6ee8773309 in opal_event_base_loop (base=0xb79050,
>>> flags=<value
>>> optimized out>) at event.c:838
>>> #3 0x00002b6ee875ee92 in opal_progress () at runtime/
>>> opal_progress.c:189
>>> #4 0x0000000039f00001 in ?? ()
>>> #5 0x00002b6ee87979c9 in std::ios_base::Init::~Init () at
>>> ../../.././libstdc++-v3/src/ios_init.cc:123
>>> #6 0x00007fffc32c8cc8 in ?? ()
>>> #7 0x00002b6ee9d20955 in orte_grpcomm_bad_get_proc_attr
>>> (proc=<value
>>> optimized out>, attribute_name=0x2b6ee88e5780 " \020322351n+",
>>> val=0x2b6ee875ee92, size=0x7fffc32c8cd0) at grpcomm_bad_module.c:500
>>> #8 0x00002b6ee86dd511 in ompi_modex_recv_key_value (key=<value
>>> optimized
>>> out>, source_proc=<value optimized out>, value=0xbb3a00, dtype=14
>>> '\016') at
>>> runtime/ompi_module_exchange.c:125
>>> #9 0x00002b6ee86d7ea1 in ompi_proc_set_arch () at proc/proc.c:154
>>> #10 0x00002b6ee86db1b0 in ompi_mpi_init (argc=15,
>>> argv=0x7fffc32c92f8,
>>> requested=<value optimized out>, provided=0x7fffc32c917c) at
>>> runtime/ompi_mpi_init.c:699
>>> #11 0x00007fffc32c8e88 in ?? ()
>>> #12 0x00002b6ee77f8348 in ?? ()
>>> #13 0x00007fffc32c8e60 in ?? ()
>>> #14 0x00007fffc32c8e20 in ?? ()
>>> #15 0x0000000009efa994 in ?? ()
>>> #16 0x0000000000000000 in ?? ()
>>> _______________________________________________
>>> users mailing list
>>> users_at_[hidden]
>>> http://www.open-mpi.org/mailman/listinfo.cgi/users
>>
>> --
>> *---------------------------------------------------------------------*
>> Gilbert Grosdidier Gilbert.Grosdidier_at_[hidden]
>> LAL / IN2P3 / CNRS Phone : +33 1 6446 8909
>> Faculté des Sciences, Bat. 200 Fax : +33 1 6446 8546
>> B.P. 34, F-91898 Orsay Cedex (FRANCE)
>> *---------------------------------------------------------------------*
>
>
> _______________________________________________
> users mailing list
> users_at_[hidden]
> http://www.open-mpi.org/mailman/listinfo.cgi/users

--
*---------------------------------------------------------------------*
   Gilbert Grosdidier                 Gilbert.Grosdidier_at_[hidden]
   LAL / IN2P3 / CNRS                 Phone : +33 1 6446 8909
   Faculté des Sciences, Bat. 200     Fax   : +33 1 6446 8546
   B.P. 34, F-91898 Orsay Cedex (FRANCE)
*---------------------------------------------------------------------*