[gpaw-users] Segfaults when testing gpaw-python
Hildur Guðmundsdóttir
hildur.gud at gmail.com
Fri Mar 22 16:52:31 CET 2013
On 03/22/2013 01:50 PM, Marcin Dulak wrote:
> On 03/22/13 14:42, Hildur Guðmundsdóttir wrote:
>> On 03/22/2013 11:42 AM, Marcin Dulak wrote:
>>> On 03/22/13 12:02, Hildur Guðmundsdóttir wrote:
>>>> On 03/22/2013 10:11 AM, Jens Jørgen Mortensen wrote:
>>>>> Den 21-03-2013 12:21, Hildur Guðmundsdóttir skrev:
>>>>>> Good morning
>>>>>>
>>>>>> I am trying to install GPAW on the NHPC cluster
>>>>>> (http://nhpc.hi.is/content/hardware) and when running the tests with
>>>>>> gpaw-python they stop on the fourth test, due to segmentation faults.
>>>>>> This happens for numpy_zdotc_graphite.py, eigh.py and blas.py and no
>>>>>> doubt others too. I do not get any output, the tests just stop
>>>>>> running.
>>>>>>
>>>>>> When doing for instance:
>>>>>> mpirun -np 2 gpaw-python gpaw/test/blas.py
>>>>>>
>>>>>> I get
>>>>>> --------------------------------------------------------------------------
>>>>>> mpirun noticed that process rank 1 with PID 11048 on node
>>>>>> gardar-1.local exited on signal 11 (Segmentation fault).
>>>>>> --------------------------------------------------------------------------
>>>>>>
>>>>>> I have compiled numpy and GPAW with gcc using the same mkl libraries
>>>>>> for blas and lapack ('mkl_blas95_lp64', 'mkl_lapack95_lp64'). I'm
>>>>>> using python2.7, numpy1.7.0, openmpi-1.4.4 and gpaw 0.9.0-8965. First
>>>>>> I installed with the system provided numpy and got the same problem.
>>>>>> Attached is my customize.py
>>>>>>
>>>>>> I saw this has happened to others before but couldn't find a solution
>>>>>> for the problem, I would appreciate any advice.
>>>>>
>>>>> Does all tests pass in serial? Also when you run them with "mpirun
>>>>> -np 1 gpaw-python"?
>>>>>
>>>>> You may want to try the latest development version which should
>>>>> get rid of warnings from numpy-1.7.
>>>>
>>>> I am running the serial tests but so far only two have failed
>>>> (aeatom and pw/interpol), no segfaults.
>>>> When I use "mpirun -np 1 gpaw-python" I get a segmentation fault on
>>>> the before mentioned tests (numpy_zdotc_graphite.py, eigh.py,
>>>> blas.py) and many others, whereas some work fine, like spinpol.py
>>>> and fermisplit.py.
>>> any numpy*py test suggest the old problem related to
>>> numpy/gpaw-python blas.
>>> Can you try numpy built with it's own internal blas/lapack?
>>> You need to manually patch it, as described for example here:
>>> https://wiki.fysik.dtu.dk/gpaw/install/Linux/r410_psmn.ens-lyon.html
>>> Also to make sure: do you export /OMP_NUM_THREADS=1?/
>>>
>>> Marcin
>>
>> I have tried the blas/lapack patch for numpy and /OMP_NUM_THREADS=1/.
>> I still get the same problem.
> did numpy actually built with it's own blas?
> Please check with something like:
> find /usr/lib64/python2.7/site-packages/numpy/ -name _dotblas.so |
> xargs ldd
> There should be no _dotblas.so is numpy uses it's own blas.
> If yes: i would try to build gpaw with acml while keeping the default
> numpy.
Thank you Marcin!
Numpy was actually built with its own blas, however when I built gpaw
with acml instead of mkl my segfaults went away.
Best regards,
Hildur
>
> Marcin
>>
>> Hildur
>>
>>>>
>>>>
>>>>>
>>>>> Jens Jørgen
>>>>>
>>>>>> Thanks in advance,
>>>>>>
>>>>>> Hildur
>>>>>>
>>>>>>
>>>>>> _______________________________________________
>>>>>> gpaw-users mailing list
>>>>>> gpaw-users at listserv.fysik.dtu.dk
>>>>>> https://listserv.fysik.dtu.dk/mailman/listinfo/gpaw-users
>>>>>
>>>>
>>>>
>>>>
>>>> _______________________________________________
>>>> gpaw-users mailing list
>>>> gpaw-users at listserv.fysik.dtu.dk
>>>> https://listserv.fysik.dtu.dk/mailman/listinfo/gpaw-users
>>>
>>>
>>> --
>>> ***********************************
>>>
>>> Marcin Dulak
>>> Technical University of Denmark
>>> Department of Physics
>>> Building 307, Room 229
>>> DK-2800 Kongens Lyngby
>>> Denmark
>>> Tel.: (+45) 4525 3157
>>> Fax.: (+45) 4593 2399
>>> email:Marcin.Dulak at fysik.dtu.dk
>>>
>>> ***********************************
>>
>
>
> --
> ***********************************
>
> Marcin Dulak
> Technical University of Denmark
> Department of Physics
> Building 307, Room 229
> DK-2800 Kongens Lyngby
> Denmark
> Tel.: (+45) 4525 3157
> Fax.: (+45) 4593 2399
> email:Marcin.Dulak at fysik.dtu.dk
>
> ***********************************
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://listserv.fysik.dtu.dk/pipermail/gpaw-users/attachments/20130322/d1e6b966/attachment-0001.html
More information about the gpaw-users
mailing list