Gilles --
Can you send a stack trace from one of these crashes?
I am *guessing* that the following is happening:
1. coll selection begins
2. coll ml is queried, and disqualifies itself (but is not dlclosed yet)
3. coll hcol is queried, which ends up calling down into libhcol. libhcol
calls a coll_ml_* symbol (which is apparently in a different .o file in the
library), but the linker has already resolved that coll_ml_* symbol in the coll
ml DSO. So the execution transfers back up into the coll ml DSO, and ...
kaboom.
A simple stack trace will confirm this -- it should show execution going down
into libhcol and then back up into coll ml.
> On Jun 25, 2015, at 1:03 AM, Gilles Gouaillardet <gil...@rist.or.jp> wrote:
>
> Folks,
>
> this is a followup on an issue reported by Daniel on the users mailing list :
> OpenMPI is built with hcoll from Mellanox.
> the coll ml module has default priority zero.
>
> on my cluster, it works just fine
> on Daniel's cluster, it crashes.
>
> i was able to reproduce the crash by tweaking mca_base_component_path and
> ensure
> the coll ml module is loaded first.
>
> basically, i found two issues :
> 1) libhcoll.so (vendor lib provided by Mellanox, i tested
> hpcx-v1.3.336-gcc-OFED-1.5.4.1-redhat6.2-x86_64) seems to include its own
> coll ml, since there are some *public* symbols that are common to this module
> (ml_open, ml_coll_hier_barrier_setup, ...)
> 2) coll ml priority is zero, and even if the library is dlclose'd, it seems
> this is uneffective
> (nothing changed in /proc/xxx/maps before and after dlclose)
>
>
> there are two workarounds :
> mpirun --mca coll ^ml
> or
> mpirun --mca coll ^hcoll ... (probably not what is needed though ...)
>
> is it expected the library is not unloaded after dlclose ?
>
> Mellanox folks,
> can you please double check how libhcoll is built ?
> i guess it would work if the ml_ symbols were private to the library.
> if not, the only workaround is to mpirun --mca coll ^ml
> otherwise, it might crash (if coll_ml is loaded before coll_hcoll, which is
> really system dependent)
>
> Cheers,
>
> Gilles
> On 6/25/2015 10:46 AM, Gilles Gouaillardet wrote:
>> Daniel,
>>
>> thanks for the logs.
>>
>> an other workaround is to
>> mpirun --mca coll ^hcoll ...
>>
>> i was able to reproduce the issue, and it surprisingly occurs only if the
>> coll_ml module is loaded *before* the hcoll module.
>> /* this is not the case on my system, so i had to hack my
>> mca_base_component_path in order to reproduce the issue */
>>
>> as far as i understand, libhcoll is a proprietary software, so i cannot dig
>> into it.
>> that being said, i noticed libhcoll defines some symbols (such as
>> ml_coll_hier_barrier_setup) that are also defined by the coll_ml module, so
>> it is likely hcoll coll_ml and openmpi coll_ml are not binary compatible
>> hence the error.
>>
>> i will dig a bit more and see if this is even supposed to happen (since
>> coll_ml_priority is zero, why is the module still loaded ?)
>>
>> as far as i am concerned, you *have to* mpirun --mca coll ^ml or update your
>> user/system wide config file to blacklist the coll_ml module to ensure this
>> is working.
>>
>> Mike and Mellanox folks, could you please comment on that ?
>>
>> Cheers,
>>
>> Gilles
>>
>>
>>
>> On 6/24/2015 5:23 PM, Daniel Letai wrote:
>>> Gilles,
>>>
>>> Attached the two output logs.
>>>
>>> Thanks,
>>> Daniel
>>>
>>> On 06/22/2015 08:08 AM, Gilles Gouaillardet wrote:
>>>> Daniel,
>>>>
>>>> i double checked this and i cannot make any sense with these logs.
>>>>
>>>> if coll_ml_priority is zero, then i do not any way how
>>>> ml_coll_hier_barrier_setup can be invoked.
>>>>
>>>> could you please run again with --mca coll_base_verbose 100
>>>> with and without --mca coll ^ml
>>>>
>>>> Cheers,
>>>>
>>>> Gilles
>>>>
>>>> On 6/22/2015 12:08 AM, Gilles Gouaillardet wrote:
>>>>> Daniel,
>>>>>
>>>>> ok, thanks
>>>>>
>>>>> it seems that even if priority is zero, some code gets executed
>>>>> I will confirm this tomorrow and send you a patch to work around the
>>>>> issue if that if my guess is proven right
>>>>>
>>>>> Cheers,
>>>>>
>>>>> Gilles
>>>>>
>>>>> On Sunday, June 21, 2015, Daniel Letai <d...@letai.org.il> wrote:
>>>>> MCA coll: parameter "coll_ml_priority" (current value: "0", data source:
>>>>> default, level: 9 dev/all, type: int)
>>>>>
>>>>> Not sure how to read this, but for any n>1 mpirun only works with --mca
>>>>> coll ^ml
>>>>>
>>>>> Thanks for helping
>>>>>
>>>>> On 06/18/2015 04:36 PM, Gilles Gouaillardet wrote:
>>>>>> This is really odd...
>>>>>>
>>>>>> you can run
>>>>>> ompi_info --all
>>>>>> and search coll_ml_priority
>>>>>>
>>>>>> it will display the current value and the origin
>>>>>> (e.g. default, system wide config, user config, cli, environment
>>>>>> variable)
>>>>>>
>>>>>> Cheers,
>>>>>>
>>>>>> Gilles
>>>>>>
>>>>>> On Thursday, June 18, 2015, Daniel Letai <d...@letai.org.il> wrote:
>>>>>> No, that's the issue.
>>>>>> I had to disable it to get things working.
>>>>>>
>>>>>> That's why I included my config settings - I couldn't figure out which
>>>>>> option enabled it, so I could remove it from the configuration...
>>>>>>
>>>>>> On 06/18/2015 02:43 PM, Gilles Gouaillardet wrote:
>>>>>>> Daniel,
>>>>>>>
>>>>>>> ML module is not ready for production and is disabled by default.
>>>>>>>
>>>>>>> Did you explicitly enable this module ?
>>>>>>> If yes, I encourage you to disable it
>>>>>>>
>>>>>>> Cheers,
>>>>>>>
>>>>>>> Gilles
>>>>>>>
>>>>>>> On Thursday, June 18, 2015, Daniel Letai <d...@letai.org.il> wrote:
>>>>>>> given a simple hello.c:
>>>>>>>
>>>>>>> #include <stdio.h>
>>>>>>> #include <mpi.h>
>>>>>>>
>>>>>>> int main(int argc, char* argv[])
>>>>>>> {
>>>>>>> int size, rank, len;
>>>>>>> char name[MPI_MAX_PROCESSOR_NAME];
>>>>>>>
>>>>>>> MPI_Init(&argc, &argv);
>>>>>>> MPI_Comm_size(MPI_COMM_WORLD, &size);
>>>>>>> MPI_Comm_rank(MPI_COMM_WORLD, &rank);
>>>>>>> MPI_Get_processor_name(name, &len);
>>>>>>>
>>>>>>> printf("%s: Process %d out of %d\n", name, rank, size);
>>>>>>>
>>>>>>> MPI_Finalize();ffff
>>>>>>> }
>>>>>>>
>>>>>>> for n=1
>>>>>>> mpirun -n 1 ./hello
>>>>>>> it works correctly.
>>>>>>>
>>>>>>> for n>1 it segfaults with signal 11
>>>>>>> used gdb to trace the problem to ml coll:
>>>>>>>
>>>>>>> Program received signal SIGSEGV, Segmentation fault.
>>>>>>> 0x00007ffff6750845 in ml_coll_hier_barrier_setup()
>>>>>>> from <path to openmpi 1.8.5>/lib/openmpi/mca_coll_ml.so
>>>>>>>
>>>>>>> running with
>>>>>>> mpirun -n 2 --mca coll ^ml ./hello
>>>>>>> works correctly
>>>>>>>
>>>>>>> using mellanox ofed 2.3-2.0.5-rhel6.4-x86_64, if it's at all relevant.
>>>>>>> openmpi 1.8.5 was built with following options:
>>>>>>> rpmbuild --rebuild --define 'configure_options --with-verbs=/usr
>>>>>>> --with-verbs-libdir=/usr/lib64 CC=gcc CXX=g++
>>>>>>> FC=gfortran CFLAGS="-g -O3" --enable-mpirun-prefix-by-default
>>>>>>> --enable-orterun-prefix-by-default --disable-debug
>>>>>>> --with-knem=/opt/knem-1.1.1.90mlnx --with-platform=optimized
>>>>>>> --without-mpi-param-check
>>>>>>> --with-contrib-vt-flags=--disable-iotrace --enable-builtin-atomics
>>>>>>> --enable-cxx-exceptions --enable-sparse-groups
>>>>>>> --enable-mpi-thread-multiple --enable-memchecker
>>>>>>> --enable-btl-openib-failover --with-hwloc=internal --with-verbs
>>>>>>> --with-x --with-slurm --with-pmi=/opt/slurm
>>>>>>> --with-fca=/opt/mellanox/fca --with-mxm=/opt/mellanox/mxm
>>>>>>> --with-hcoll=/opt/mellanox/hcoll' openmpi-1.8.5-1.src.rpm
>>>>>>>
>>>>>>> gcc version 5.1.1
>>>>>>>
>>>>>>> Thanks in advance
>>>>>>> _______________________________________________
>>>>>>> users mailing list
>>>>>>> us...@open-mpi.org
>>>>>>> Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/users
>>>>>>> Link to this post:
>>>>>>> http://www.open-mpi.org/community/lists/users/2015/06/27154.php
>>>>>>>
>>>>>>>
>>>>>>> _______________________________________________
>>>>>>> users mailing list
>>>>>>>
>>>>>>> us...@open-mpi.org
>>>>>>>
>>>>>>> Subscription:
>>>>>>> http://www.open-mpi.org/mailman/listinfo.cgi/users
>>>>>>>
>>>>>>> Link to this post:
>>>>>>> http://www.open-mpi.org/community/lists/users/2015/06/27155.php
>>>>>>
>>>>>>
>>>>>>
>>>>>> _______________________________________________
>>>>>> users mailing list
>>>>>>
>>>>>> us...@open-mpi.org
>>>>>>
>>>>>> Subscription:
>>>>>> http://www.open-mpi.org/mailman/listinfo.cgi/users
>>>>>>
>>>>>> Link to this post:
>>>>>> http://www.open-mpi.org/community/lists/users/2015/06/27157.php
>>>>>
>>>>>
>>>>>
>>>>> _______________________________________________
>>>>> users mailing list
>>>>>
>>>>> us...@open-mpi.org
>>>>>
>>>>> Subscription:
>>>>> http://www.open-mpi.org/mailman/listinfo.cgi/users
>>>>>
>>>>> Link to this post:
>>>>> http://www.open-mpi.org/community/lists/users/2015/06/27169.php
>>>>
>>>>
>>>>
>>>> _______________________________________________
>>>> users mailing list
>>>>
>>>> us...@open-mpi.org
>>>>
>>>> Subscription:
>>>> http://www.open-mpi.org/mailman/listinfo.cgi/users
>>>>
>>>> Link to this post:
>>>> http://www.open-mpi.org/community/lists/users/2015/06/27170.php
>>>
>>>
>>>
>>> _______________________________________________
>>> users mailing list
>>>
>>> us...@open-mpi.org
>>>
>>> Subscription:
>>> http://www.open-mpi.org/mailman/listinfo.cgi/users
>>>
>>> Link to this post:
>>> http://www.open-mpi.org/community/lists/users/2015/06/27183.php
>>
>>
>>
>> _______________________________________________
>> devel mailing list
>>
>> de...@open-mpi.org
>>
>> Subscription:
>> http://www.open-mpi.org/mailman/listinfo.cgi/devel
>>
>> Link to this post:
>> http://www.open-mpi.org/community/lists/devel/2015/06/17528.php
>
> _______________________________________________
> devel mailing list
> de...@open-mpi.org
> Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/devel
> Link to this post:
> http://www.open-mpi.org/community/lists/devel/2015/06/17529.php
--
Jeff Squyres
jsquy...@cisco.com
For corporate legal information go to:
http://www.cisco.com/web/about/doing_business/legal/cri/
