[Wien] Bug report (impi & AVX512 register nodes)

[Laurence Marks]

There appears to be a bug connected to impi versions from 2017 or later
when running mpi on AVX512 capable machines. In 0.01-0.1% of cases for an
ill-conditioned problem the eigenvectors end up with slightly wrong total
charge (0.001-0.2 electrons total for 30 k-points), which will appear as an
anomaly in :NEC01.

The mixer will handle this adequately, but it is still not "right".

N.B., it may be something OS or fabric related rather than specific to
impi; I know that it has "gone away" when I switched to an earlier impi
version. Compiler options, mkl, shared/static, scalapck version (mkl or
Netlib) had no effect, only regressing the impi version. I am 99.999%
certain this is not a Wien2k bug.

-- 
Professor Laurence Marks
"Research is to see what everybody else has seen, and to think what nobody
else has thought", Albert Szent-Gyorgi
www.numis.northwestern.edu ; Corrosion in 4D: MURI4D.numis.northwestern.edu
Partner of the CFW 100% program for gender equity, www.cfw.org/100-percent
Co-Editor, Acta Cryst A
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://zeus.theochem.tuwien.ac.at/pipermail/wien/attachments/20180731/9d06ace5/attachment.html>