Dear Lucas,
there was indeed a bug with taskgroups in v 6.1, fixed in v6.2 (see
release notes:
https://gitlab.com/QEF/q-e/blob/develop/Doc/release-notes#L338 ).
Some time ago I gave a presentation about the parallelization scheme of
QE. Unfortunately the session was not recorded but you may extract some
information from the slides: https://tinyurl.com/y3qb7jql .
Hope this helps,
best regards,
Pietro
On 5/9/19 6:43 PM, Lucas Nicolás Lodeiro Moraga wrote:
Hi all!
I am performing a scaling test with QE (versions 6.1 and 6.3) for a big
sistem >2000 electrons with GGA/USPP. I found that when I use more than
one node the use of -nt is very usefull to improve the calculation time.
For this system nt=number of nodes is the best choise in most cases...
but sometimes nt = number of nodes + X, with X = 1,2,3,4... (no more
than 4) is better. As an example, if i use 440 processors, with 44
processors nodes, the best nt is 11.
Also, I read that the -nt is usefull when the FFT 3D grid < n°
processors, but in some cases when FFT 3D grid > n° processors, -nt 2 or
3 improves the calculation too.
Can someone help me with this topic, give me more information and
internalize me in this topic?
On the other hand, I try this with hybrid functional in a simple
semiconductor sistem with ~50 atoms with HSE/NCPP. With 6.1 version the
calculation CRASH when I use -nt 2, the error message is:
iteration # 1 ecut= 80.00 Ry beta=0.30
Davidson diagonalization with overlap
ethr = 1.00E-02, avg # of iterations = 8.0
negative rho (up, down): 1.101E-01 0.000E+00
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
Error in routine electrons (1):
charge is wrong: smearing is needed
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
stopping ...
When I use smearing:
iteration # 1 ecut= 80.00 Ry beta=0.30
Davidson diagonalization with overlap
ethr = 1.00E-02, avg # of iterations = 8.0
negative rho (up, down): 1.081E-01 0.000E+00
total cpu time spent up to now is 24.5 secs
WARNING: integrated charge= 312.31515712, expected= 320.00000000
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
Error in routine electrons (1):
charge is wrong
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
stopping ...
It is very weird, and with 6.3 version the calculation runs without
problem, but take more time, and the speedup is very very low in
comparison with one node calculation.
There is a bug in 6.1 version? Is it a bug in the 6.1 compilation?
Somebody knows some method to accelerate the hybrid calculations to be
able to occupy more nodes?
Regards
Lucas Lodeiro
Universidad de Chile
_______________________________________________
Quantum Espresso is supported by MaX (www.max-centre.eu/quantum-espresso)
users mailing list [email protected]
https://lists.quantum-espresso.org/mailman/listinfo/users
--
Pietro Bonfa'
Dipartimento di Scienze Matematiche, Fisiche e Informatiche
Parco Area delle Scienze, 7/A
43124 Parma
_______________________________________________
Quantum Espresso is supported by MaX (www.max-centre.eu/quantum-espresso)
users mailing list [email protected]
https://lists.quantum-espresso.org/mailman/listinfo/users
