[Wien] segmentation fault in lapwso

Thu Aug 19 13:44:50 CEST 2021

I'm still on vacations, so cannot test myself.

However, I experienced such problems before. It has to do with 
multithreading (1 thread works always fine) and the mkl routine zheevr.

In my case I could fix the problem by enlarging the workspace beyond 
what the routine calculates itself. (see comment in hmsec on line 841).

Right below, the workspace was enlarged by a factor 10, which fixed my 
problem. But I can easily envision that it might not be enough in some 
other cases.

An alternative is to switch back to zheevx (commented in the code).

Peter Blaha

Am 18.08.2021 um 20:01 schrieb Pavel Ondračka:
> Right, I think that the reason deallocate is failing because the memory
> has been corrupted at some earlier point is quite clear, the only other
> option why it should crash would be that it was not allocated at all,
> which seem not to be the case here... The question is what corrupted
> the memory and even more strange is why does it work if we disable MKL
> multithreading?
>
> It could indeed be that we are doing something wrong. I can imagine the
> memory could be corrupted in some BLAS call if the number of
> columns/rows passed to the specific BLAS call is more than the actual
> size of the matrix, than this could easily happen (and the
> multithreading is somehow influencing what the final value of the
> corrupted memory, and depending on the final value the deallocate could
> fail or pass somehow). This should be possible to diagnose with
> valgrind as suggested.
>
> Luis, can you upload the testcase somewhere, or recompile with
> debuginfo as suggested by Laurence earlier, run "valgrind --track-
> origins=yes lapwso lapwso.def" and send the output? Just be warned,
> there is a massive slowdown with valgrind (up to 100x) and the logfile
> can get very large.
>
> Best regards
> Pavel
>
>
> On Wed, 2021-08-18 at 12:10 -0500, Laurence Marks wrote:
>> Correction, I was looking at an older modules.F. It looks like it
>> should be
>>
>> DEALLOCATE(vect,stat=IV) ; if(IV .ne. 0)write(*,*)IV
>>
>>
>> On Wed, Aug 18, 2021 at 11:23 AM Laurence Marks
>> <laurence.marks at gmail.com> wrote:
>>> I do wonder about this. I suggest editing module.F and changing
>>> lines 118 and 119 to
>>>       DEALLOCATE(en,stat=Ien) ; if(Ien .ne. 0)write(*,*)'Err en
>>> ',ien
>>>       DEALLOCATE(vnorm,stat=Ivn ; ) if(Ivn .ne. 0)write(*,*)'Err
>>> vnorm ',Ivn
>>>
>>> There is every chance that the bug is not in those lines, but
>>> somewhere completely different. SIGSEV often means that the code
>>> has been overwritten, for instance arrays going out of bounds.
>>>
>>> You can also recompile with -g (don't change other options)
>>> added, and/or -C. Sometimes this is better. Or use other things
>>> like debuggers or valgrind.
>>>
>>> On Wed, Aug 18, 2021 at 10:47 AM Pavel Ondračka
>>> <pavel.ondracka at email.cz> wrote:
>>>> I'm CCing the list back as the crash was now diagnosed to a
>>>> likely
>>>> MKL
>>>> problem, see below for more details.
>>>>>
>>>>>> So just to be clear, explicitly setting OMP_STACKSIZE=1g does
>>>> not
>>>>>> help
>>>>>> to solve the issue?
>>>>>>
>>>>>
>>>>> Right! OMP_STACKSIZE=1g with OMP_NUM_THREADS=4 does not solve
>>>>> the
>>>>> problem!
>>>>>   
>>>>>> The problem is that the OpenMP code in lapwso is very simple,
>>>> so I'm
>>>>>> having problems seeing how it could be causing the problems.
>>>>>>
>>>>>> Could you also try to see what happens if run with:
>>>>>> OMP_NUM_THREADS=1
>>>>>> MKL_NUM_THREADS=4
>>>>>>
>>>>>
>>>>> It does not work with these values, but I checked and it works
>>>>> reverting them:
>>>>> OMP_NUM_THREADS=4
>>>>> MKL_NUM_THREADS=1
>>>> This was very helpfull and IMO points to a problem with MKL
>>>> instead
>>>> of
>>>> Wien2k.
>>>>
>>>> Unfortunatelly setting MKL_NUM_THREADS=1 globally will reduce the
>>>> OpenMP performance, mostly in lapw1 but also at other places. So
>>>> if
>>>> you
>>>> want to keep the OpenMP BLAS/lapack level parallelism you have to
>>>> either find some MKL version that works (if you do please report
>>>> it
>>>> here), link with OpenBLAS (using it for lapwso is enough) or
>>>> create
>>>> a
>>>> simple wrapper that sets the MKL_NUM_THREADS=1 just for lapwso,
>>>> i.e.,
>>>> rename lapwso binary in WIENROOT to lapwso_bin and create new
>>>> lapwso
>>>> file there with:
>>>>
>>>> #!/bin/bash
>>>> MKL_NUM_THREADS=1 lapwso_bin $1
>>>>
>>>> and set it to executable with chmod +x lapwso.
>>>>
>>>> Or maybe MKL has a non-OpenMP version which you could link with
>>>> just
>>>> lapwso and use standard one in other parts, but dunno, I mostly
>>>> use
>>>> OpenBLAS. If you need some further help, let me know.
>>>>
>>>> Reporting the issue to intel could be also nice, however I never
>>>> had
>>>> any real luck there and it is also a bit problematic as you can't
>>>> provide testcase due to Wien2k being proprietary code...
>>>>
>>>> Best regards
>>>> Pavel
>>>>
>>>>>   
>>>>>> This should disable the Wien2k-specific OpenMP parallelism
>>>>>> but
>>>> still
>>>>>> keep the rest of paralellism at the BLAS/lapack level.
>>>>>>
>>>>>
>>>>> So, perhaps, the problem is related to MKL!
>>>>>   
>>>>>> Another option is that something is going wrong before lapwso
>>>> and the
>>>>>> lapwso crash is just the symptom. What happens if you run
>>>> everything
>>>>>> up
>>>>>> to lapwso without OpenMP (OMP_NUM_THREADS=1) and than enable
>>>>>> it
>>>> just
>>>>>> for lapwso?
>>>>>>
>>>>>
>>>>> If I run lapw0 and lapw1 with OMP_NUM_THREADS=4 and then change
>>>> it to 1
>>>>> just before lapwso, it works.
>>>>> If I do the opposite, starting with OMP_NUM_THREADS=1 and then
>>>> change
>>>>> it to 4 just before lapwso, it does not work.
>>>>> So I believe that the problem is really at lapwso.
>>>>>   
>>>>>     If you need more information, please, let me know!
>>>>>     All the best,
>>>>>              Luis
>>>>
>>>> _______________________________________________
>>>> Wien mailing list
>>>> Wien at zeus.theochem.tuwien.ac.at
>>>> https://urldefense.com/v3/__http://zeus.theochem.tuwien.ac.at/mailman/listinfo/wien__;!!Dq0X2DkFhyF93HkjWTBQKhk!H_VXJmyf6v2ZSCmTICvdVDv1QuKxPqCDcjbbytr7Fh51-KF5rv8A2uvyMlW3x3YA4jSb3A$
>>>>
>>>> SEARCH the MAILING-LIST at:
>>>> https://urldefense.com/v3/__http://www.mail-archive.com/wien@zeus.theochem.tuwien.ac.at/index.html__;!!Dq0X2DkFhyF93HkjWTBQKhk!H_VXJmyf6v2ZSCmTICvdVDv1QuKxPqCDcjbbytr7Fh51-KF5rv8A2uvyMlW3x3aDFmAN4g$
>>>>
>>>
>>> -- 
>>> Professor Laurence Marks
>>> Department of Materials Science and Engineering
>>> Northwestern University
>>> www.numis.northwestern.edu
>>> "Research is to see what everybody else has seen, and to think what
>>> nobody else has thought" Albert Szent-Györgyi
>>
>> _______________________________________________
>> Wien mailing list
>> Wien at zeus.theochem.tuwien.ac.at
>> http://zeus.theochem.tuwien.ac.at/mailman/listinfo/wien
>> SEARCH the MAILING-LIST at:
>> http://www.mail-archive.com/wien@zeus.theochem.tuwien.ac.at/index.html
>
> _______________________________________________
> Wien mailing list
> Wien at zeus.theochem.tuwien.ac.at
> http://zeus.theochem.tuwien.ac.at/mailman/listinfo/wien
> SEARCH the MAILING-LIST at:  http://www.mail-archive.com/wien@zeus.theochem.tuwien.ac.at/index.html

-- 
-----------------------------------------------------------------------
Peter Blaha,  Inst. f. Materials Chemistry, TU Vienna, A-1060 Vienna
Phone: +43-158801165300
Email: peter.blaha at tuwien.ac.at
WWW:   http://www.imc.tuwien.ac.at      WIEN2k: http://www.wien2k.at
-------------------------------------------------------------------------