[QE-users] Immediate segfault in pw.x when compiled on SGI machine

[Nikolas Antolin]

Having read the FAQ before posting, this response is approximately what I
expected. Unfortunately, more analysis on my part shows it is very likely
incorrect.

On the same machine built against the same libraries with the exception of
the specific MKL (blacs_openmpi vs blacs_intelmpi) library but with OpenMPI
3.1.6, pw.x runs inputs from the test-suite without issue. This suggests to
me that some part of the code is in loose compliance with the MPI standard,
causing memory issues/incompatibilities with certain implementations.

This is further supported by a full stack trace when the software is built
against HPE MPT, which indicates a memory issue with the in-place AllReduce
operation on line 253 of FFTXlib/stick_base.f90. I can't be certain that
this is the same issue encountered with IntelMPI, but it's certainly
possible even with slightly different output given application output
buffering on the machine.

Obviously, I will be looking further into this tomorrow but welcome
suggestions as to flags to loosen the standards interpretation of IntelMPI.

Regards,
Nik Antolin

On Thu, Oct 14, 2021, 00:23 Paolo Giannozzi <p.giannozzi at gmail.com> wrote:

> On Thu, Oct 14, 2021 at 1:09 AM Nikolas Antolin <nantolin at gmail.com>
> wrote:
>
> (basically worthless as far as I can tell):
>> #0  0x00002af86df05817 in atan () from /lib64/libm.so.6
>>
>
> well, maybe not completely worthless: the segfault happens in the arctan
> function, presumably because the compiler links a wrong set of mathematical
> libraries. There is something weird in the software of your machine.
>
> QE does not use any nonstandard, "advanced", exotic features of MPI, as
> far as I know, All MPI libraries should work.
>
> Paolo
>
> --
> Paolo Giannozzi, Dip. Scienze Matematiche Informatiche e Fisiche,
> Univ. Udine, via delle Scienze 206, 33100 Udine, Italy
> Phone +39-0432-558216, fax +39-0432-558222
>
> _______________________________________________
> Quantum ESPRESSO is supported by MaX (www.max-centre.eu)
> users mailing list users at lists.quantum-espresso.org
> https://lists.quantum-espresso.org/mailman/listinfo/users
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.quantum-espresso.org/pipermail/users/attachments/20211014/e5f4a7e5/attachment.html>