[Q-e-developers] QE 6.1 vs. 6.2 performance still

Ari P Seitsonen Ari.P.Seitsonen at iki.fi
Mon Oct 30 13:05:34 CET 2017


Hello All,

   Adding a note to Paolo's reply on Saturday - sorry, I deleted the post 
and could no longer reply to it -

> > SCF steps last longer in QE 6.2
> 
> this is likely a consequence of the new FFT's, that have better scaling 
> on many (MPI) processes but are slower on few processors.
> I would try MPI parallelization instead of OpenMP, by the way.

I have encountered a slowdown of 25 % in a small system of 8-10 atoms. Is 
this possible? I looked in the timings, indeed it seemed to come from 
everywhere, not only from specific subroutines. I have a small 12-core AMD 
machine, dedicated, with the same gfortran (5.4.0), OpenMPI (MPI-only, no 
OpenMP), SCALAPACK both ATLAS and OpenBLAS libraries, with both versions 
(6.0 was also considerably faster than 6.2). Is there a way to switch back 
to the previous FFT (pre-processor)? I have not tried in bigger systems 
yet; if the difference is smaller there, I can live with it.

     Greetings,

        apsi

-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-=*=-
   Ari Paavo Seitsonen / Ari.P.Seitsonen at iki.fi / http://www.iki.fi/~apsi/
     Ecole Normale Supérieure (ENS), Département de Chimie, Paris
     Mobile (F) : +33 789 37 24 25    (CH) : +41 79 71 90 935


More information about the developers mailing list