[mpich-discuss] Problems Running WRF on Ubuntu 11.10, MPICH2

Sukanta Basu sukanta.basu at gmail.com
Thu Feb 9 08:25:45 CST 2012 

Hi Gus,

The Cray has 4 nodes (each containing 8 cores, 24 GB RAM). The nodes
are connected by gigE. I need to use dmpar option.

Regards,
Sukanta

On Thu, Feb 9, 2012 at 8:59 AM, Gustavo Correa <gus at ldeo.columbia.edu> wrote:
> Hi Basu
>
> Sorry, I missed the 'dmpar' information.
> I am not familiar to it, but I guess it is the Cray trick to make the CX1 machine
> look like a standard distributed memory environment?
> [As opposed to a full shared memory environment across all nodes,
> which would be 'smpar', right?]
>
> If 'dmpar' it is a standard distributed memory environment, I presume all that
> I said before still holds.  I would just try to set KMP_STACKSIZE to 16m or more on
> all nodes, and run WRF again.
>
> FYI, I had some issues compiling some models with Intel 12.0, and in other mailing
> lists I saw people that had issues with version 12.1.
> However, I compiled some models with Intel 11.1 correctly, but as I said before, not WRF.
>
> BTW, we're cheap here.  No funding for fancy machines, no Cray, no IBM, no SGI.
> The top thing we can buy is a standard Linux cluster once in a while. :)
>
> Good luck,
> Gus Correa
>
> On Feb 9, 2012, at 8:25 AM, Sukanta Basu wrote:
>
>> Hi Gus,
>>
>> Thanks for your email.
>>
>> I am compiling WRF with dmpar option (distributed memory). WRF has a
>> different option for hybrid openmp+mpi (they call it dmpar+smpar). To
>> the best of my knowledge, openmp is not invoked.
>>
>> I do understand the distinction between openmp and openmpi. Yesterday,
>> I uninstalled mpich2 and installed openmpi. I compiled and ran wrf
>> jobs. As I mentioned before, I faced different types of problems.
>>
>> I have been using WRF on various clusters for ~6-7 years. I bought a
>> Cray CX1 recently and trying to set it up myself for running WRF
>> locally. Now, I am suspecting that there is some compatibility issues
>> between WRF and the Intel Composer. I used to use Intel 11.1 compiler.
>>
>> I will set KMP_STACKSIZE and re-run the simulations with wrf+mpich2+intel.
>>
>> Best regards,
>> Sukanta
>>
>> On Thu, Feb 9, 2012 at 8:02 AM, Gustavo Correa <gus at ldeo.columbia.edu> wrote:
>>> Hi Sukanta
>>>
>>> Did you read the final part of my previous email about KMP_STACKSIZE?
>>> This is how Intel calls the OpenMP threads stack size.
>>> I think you misspelled that environment variable [it is not MP_STACKSIZE as your email says].
>>>
>>> Did you compile WRF with OpenMP turned on and with the Intel compiler?
>>> If you did, you certainly need to increase also the threads' stack size.
>>>
>>> I had experiences similar to yours, with other models compiled with Intel ifort,
>>> and OpenMP, i.e., unexplained segmentation faults, even though the stacksize was
>>> set to unlimited.
>>>
>>> Some time ago I posted this same solution in this mailing list to somebody
>>> at LLNL or ANL, I think, who was having this type of problem as well.
>>> It is common in hybrid MPI+OpenMP programs.
>>>
>>> I would set KMP_STACKSIZE to 16m  at least *on all nodes*, maybe in your .bashrc, or in the script that launches the job.  I don't remember the syntax on top of my head,
>>> but the MPICH2 mpiexec [hydra] probably has a way to export the environment variables
>>> to all processes.  Check 'man mpiexec'.
>>> You must ensure that the environment variable is set *on all nodes*.
>>>
>>> You may need more than 16m, depending on how fine a grid you are using.
>>> In another model here I had to use 512m, but this also depends
>>> on how much memory/RAM your nodes have available per core.
>>> You could try increasing it step by step, say, doubling each  time:
>>> 16m, 32m, 64m, ...
>>>
>>> Anyway, this is a guess based on what happened here.
>>> There is no guarantee that it will work, although it may be worth trying it.
>>> The problem you see may also be a bug in WRF, or an input/forcing file that is missing, etc.
>>>
>>> I hope this helps,
>>> Gus Correa
>>>
>>> PS - Note:  Just to avoid confusion with names.
>>> OpenMP and OpenMPI  [or Open MPI] are different things.
>>> The former is the thread-based standard for parallelization:
>>> http://openmp.org/wp/
>>> The latter is another open source  MPI, like MPICH2:
>>> http://www.open-mpi.org/
>>>
>>>
>>> On Feb 8, 2012, at 10:33 PM, Sukanta Basu wrote:
>>>
>>>> Hi Gus,
>>>>
>>>> I tried setting the stack option in limits.conf. No change. I logged
>>>> on to each nodes and checked that the ulimit is indeed unlimited.
>>>>
>>>> I just installed openmpi and recompiled WRF. It now runs with any
>>>> array sizes. However, I have a different problem. Now, one of the
>>>> processes quits suddenly during the run (with a segmentation fault
>>>> error). I think both the mpich2 and openmpi problems are somewhat
>>>> related.
>>>>
>>>> Best regards,
>>>> Sukanta
>>>>
>>>> On Wed, Feb 8, 2012 at 6:20 PM, Gustavo Correa <gus at ldeo.columbia.edu> wrote:
>>>>> Hi Sukanta
>>>>>
>>>>> Did you set the stacksize [not only memlock] to unlimited in
>>>>> /etc/security/limits.conf on all nodes?
>>>>>
>>>>> Not sure this will work, but you could try to run 'ulimit -s'  and 'ulimit -l' via mpiexec, just to check:
>>>>>
>>>>> mpiexec -prepend-rank -f hostfile -np 32 ulimit -s
>>>>> mpiexec -prepend-rank -f hostfile -np 32 ulimit -l
>>>>>
>>>>> Or just login to each node and check.
>>>>>
>>>>> Also, if your WRF is compiled with OpenMP,
>>>>> I think the Intel-specific environment variable for OMP_STACKSIZE is
>>>>> KMP_STACKSIZE [not MP_STACKSIZE], although they should also accept
>>>>> the portable/standard OMP_STACKSIZE [but I don't know if they do].
>>>>> For some models here I had to make is as big as 512m [I don't run wrf, though].
>>>>> 'man ifort' should tell more about it [at the end of the man page].
>>>>>
>>>>> I hope this helps,
>>>>> Gus Correa
>>>>>
>>>>> On Feb 8, 2012, at 4:23 PM, Anthony Chan wrote:
>>>>>
>>>>>>
>>>>>> There is fpi, Fortran counterpart of cpi, you can try that.
>>>>>> Also, there is MPICH2 testsuite which is located in
>>>>>> mpich2-xxx/test/mpi can be invoked by "make testing".
>>>>>> It is unlikely those tests will reveal anything.
>>>>>> The testsuite is meant to test the MPI implementation
>>>>>> not your app.
>>>>>>
>>>>>> As what you said earlier, your difficulty in running WRF
>>>>>> with larger dataset is memory related.  You should contact WRF
>>>>>> emailing list for more pointers.
>>>>>>
>>>>>> ----- Original Message -----
>>>>>>> Hi Anthony,
>>>>>>>
>>>>>>> Is there any other mpi example code (other than cpi.c) that I could
>>>>>>> test which will give me more information about my mpich setup?
>>>>>>>
>>>>>>> Here is the output from cpi (using 32 cores on 4 nodes):
>>>>>>>
>>>>>>> mpiuser at crayN1-5150jo:~/Misc$ mpiexec -f mpd.hosts -n 32 ./cpi
>>>>>>> Process 1 on crayN1-5150jo
>>>>>>> Process 18 on crayN2-5150jo
>>>>>>> Process 2 on crayN2-5150jo
>>>>>>> Process 26 on crayN2-5150jo
>>>>>>> Process 5 on crayN1-5150jo
>>>>>>> Process 14 on crayN2-5150jo
>>>>>>> Process 21 on crayN1-5150jo
>>>>>>> Process 22 on crayN2-5150jo
>>>>>>> Process 25 on crayN1-5150jo
>>>>>>> Process 6 on crayN2-5150jo
>>>>>>> Process 9 on crayN1-5150jo
>>>>>>> Process 17 on crayN1-5150jo
>>>>>>> Process 30 on crayN2-5150jo
>>>>>>> Process 10 on crayN2-5150jo
>>>>>>> Process 29 on crayN1-5150jo
>>>>>>> Process 13 on crayN1-5150jo
>>>>>>> Process 8 on crayN3-5150jo
>>>>>>> Process 20 on crayN3-5150jo
>>>>>>> Process 4 on crayN3-5150jo
>>>>>>> Process 12 on crayN3-5150jo
>>>>>>> Process 0 on crayN3-5150jo
>>>>>>> Process 24 on crayN3-5150jo
>>>>>>> Process 16 on crayN3-5150jo
>>>>>>> Process 28 on crayN3-5150jo
>>>>>>> Process 3 on crayN4-5150jo
>>>>>>> Process 7 on crayN4-5150jo
>>>>>>> Process 11 on crayN4-5150jo
>>>>>>> Process 23 on crayN4-5150jo
>>>>>>> Process 27 on crayN4-5150jo
>>>>>>> Process 31 on crayN4-5150jo
>>>>>>> Process 19 on crayN4-5150jo
>>>>>>> Process 15 on crayN4-5150jo
>>>>>>> pi is approximately 3.1416009869231249, Error is 0.0000083333333318
>>>>>>> wall clock time = 0.009401
>>>>>>>
>>>>>>> Best regards,
>>>>>>> Sukanta
>>>>>>>
>>>>>>> On Wed, Feb 8, 2012 at 1:19 PM, Anthony Chan <chan at mcs.anl.gov> wrote:
>>>>>>>>
>>>>>>>> Hmm.. Not sure what is happening.. I don't see anything
>>>>>>>> obviously wrong in your mpiexec verbose output (though
>>>>>>>> I am not hydra expert). Your code now is killed because of
>>>>>>>> segmentation fault. Naively, I would recompile WRF with -g
>>>>>>>> and use a debugger to see where segfault is. If you don't want
>>>>>>>> to mess around WRF source code, you may want to contact WRF
>>>>>>>> developers to see if they have encountered similar problem
>>>>>>>> before.
>>>>>>>>
>>>>>>>> ----- Original Message -----
>>>>>>>>> Dear Anthony,
>>>>>>>>>
>>>>>>>>> Thanks for your response. Yes, I did try MP_STACK_SIZE and
>>>>>>>>> OMP_STACKSIZE. The error is still there. I have attached a log file
>>>>>>>>> (I
>>>>>>>>> ran mpiexec with -verbose option). May be this will help.
>>>>>>>>>
>>>>>>>>> Best regards,
>>>>>>>>> Sukanta
>>>>>>>>>
>>>>>>>>> On Tue, Feb 7, 2012 at 3:28 PM, Anthony Chan <chan at mcs.anl.gov>
>>>>>>>>> wrote:
>>>>>>>>>>
>>>>>>>>>> I am not familar with WRF, and not sure if WRF uses any thread
>>>>>>>>>> in dmpar mode. Did you try setting MP_STACK_SIZE or OMP_STACKSIZE
>>>>>>>>>> ?
>>>>>>>>>>
>>>>>>>>>> see: http://forum.wrfforum.com/viewtopic.php?f=6&t=255
>>>>>>>>>>
>>>>>>>>>> A.Chan
>>>>>>>>>>
>>>>>>>>>> ----- Original Message -----
>>>>>>>>>>> Hi,
>>>>>>>>>>>
>>>>>>>>>>> I am using a small cluster of 4 nodes (each with 8 cores + 24 GB
>>>>>>>>>>> RAM).
>>>>>>>>>>> OS: Ubuntu 11.10. The cluster uses nfs file system and gigE
>>>>>>>>>>> connections.
>>>>>>>>>>>
>>>>>>>>>>> I installed mpich2 and ran cpi.c program successfully.
>>>>>>>>>>>
>>>>>>>>>>> I installed WRF (http://www.wrf-model.org/index.php) using the
>>>>>>>>>>> intel
>>>>>>>>>>> compilers (dmpar option)
>>>>>>>>>>> I set ulimit -l and -s to be unlimited in .bashrc (all nodes)
>>>>>>>>>>> I set memlock to be unlimited in limits.conf (all nodes)
>>>>>>>>>>> I have password-less ssh (public key sharing) on all the nodes
>>>>>>>>>>> I ran parallel jobs with 40x40x40, 40x40x50, and 40x40x60 grid
>>>>>>>>>>> points
>>>>>>>>>>> successfully. However, when I utilize 40x40x80 grid points, I
>>>>>>>>>>> get
>>>>>>>>>>> the
>>>>>>>>>>> following MPI error:
>>>>>>>>>>>
>>>>>>>>>>> **********************************************************
>>>>>>>>>>> Fatal error in PMPI_Wait: Other MPI error, error stack:
>>>>>>>>>>> PMPI_Wait(183)............: MPI_Wait(request=0x34e83a4,
>>>>>>>>>>> status=0x7fff7b24c400) failed
>>>>>>>>>>> MPIR_Wait_impl(77)........:
>>>>>>>>>>> dequeue_and_set_error(596): Communication error with rank 8
>>>>>>>>>>> **********************************************************
>>>>>>>>>>> Given that I can run the exact simulation with slightly lesser
>>>>>>>>>>> number
>>>>>>>>>>> of grid points without any problem, this error is related to
>>>>>>>>>>> stack
>>>>>>>>>>> size. What could be the problem?
>>>>>>>>>>>
>>>>>>>>>>> Thanks,
>>>>>>>>>>> Sukanta
>>>>>>>>>>>
>>>>>>>>>>> --
>>>>>>>>>>> Sukanta Basu
>>>>>>>>>>> Associate Professor
>>>>>>>>>>> North Carolina State University
>>>>>>>>>>> http://www4.ncsu.edu/~sbasu5/
>>>>>>>>>>> _______________________________________________
>>>>>>>>>>> mpich-discuss mailing list mpich-discuss at mcs.anl.gov
>>>>>>>>>>> To manage subscription options or unsubscribe:
>>>>>>>>>>> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss
>>>>>>>>>> _______________________________________________
>>>>>>>>>> mpich-discuss mailing list mpich-discuss at mcs.anl.gov
>>>>>>>>>> To manage subscription options or unsubscribe:
>>>>>>>>>> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> --
>>>>>>>>> Sukanta Basu
>>>>>>>>> Associate Professor
>>>>>>>>> North Carolina State University
>>>>>>>>> http://www4.ncsu.edu/~sbasu5/
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> --
>>>>>>> Sukanta Basu
>>>>>>> Associate Professor
>>>>>>> North Carolina State University
>>>>>>> http://www4.ncsu.edu/~sbasu5/
>>>>>> _______________________________________________
>>>>>> mpich-discuss mailing list     mpich-discuss at mcs.anl.gov
>>>>>> To manage subscription options or unsubscribe:
>>>>>> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss
>>>>>
>>>>> _______________________________________________
>>>>> mpich-discuss mailing list     mpich-discuss at mcs.anl.gov
>>>>> To manage subscription options or unsubscribe:
>>>>> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss
>>>>
>>>>
>>>>
>>>> --
>>>> Sukanta Basu
>>>> Associate Professor
>>>> North Carolina State University
>>>> http://www4.ncsu.edu/~sbasu5/
>>>> _______________________________________________
>>>> mpich-discuss mailing list     mpich-discuss at mcs.anl.gov
>>>> To manage subscription options or unsubscribe:
>>>> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss
>>>
>>> _______________________________________________
>>> mpich-discuss mailing list     mpich-discuss at mcs.anl.gov
>>> To manage subscription options or unsubscribe:
>>> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss
>>
>>
>>
>> --
>> Sukanta Basu
>> Associate Professor
>> North Carolina State University
>> http://www4.ncsu.edu/~sbasu5/
>> _______________________________________________
>> mpich-discuss mailing list     mpich-discuss at mcs.anl.gov
>> To manage subscription options or unsubscribe:
>> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss
>
> _______________________________________________
> mpich-discuss mailing list     mpich-discuss at mcs.anl.gov
> To manage subscription options or unsubscribe:
> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss



-- 
Sukanta Basu
Associate Professor
North Carolina State University
http://www4.ncsu.edu/~sbasu5/

More information about the mpich-discuss mailing list