Hello,
I noticed this strange result. I wanted to calculate the total energy of a semiconducting system (50atoms).
ISTART = 0
ENCUT = 400
IALGO = 48
ISMEAR = -5
SIGMA = 0.05
LREAL= Auto
NSW = 200
IBRION = 2
ISIF = 3
I made two identical runs one using 2 nodes, and another using 16 nodes for testing purposes. Both runs ended successfully.
Etot_16cpu = -313.428eV
Etot_2cpu = -312.991eV
Isn't this quite a large diff, for exactly the same run?
The second point is that total cpu times where 5100sec for the 2cpu run and 6400sec for the 16cpu run. Any ideas?
Thank you in advance!
gV
total energy calculation
Moderators: Global Moderator, Moderator
-
gevolo
- Newbie
- Posts: 11
- Joined: Tue May 05, 2009 3:57 pm
total energy calculation
Last edited by gevolo on Wed Oct 13, 2010 10:15 am, edited 1 time in total.
-
forsdan
- Sr. Member
- Posts: 339
- Joined: Mon Apr 24, 2006 9:07 am
- License Nr.: 173
- Location: Gothenburg, Sweden
total energy calculation
To start with, the runs do not have to be equal since the number of bands will be adjusted so that NBANDS divided by NPAR is an integer. This can for a specific geometry give a small energy difference if the number of bands is not sufficiently large to start with.
However, for the larger energy difference as in your case, you have presumably ended up in two different geometries. Please improve the break conditions for both the electronic iteration and the ionic iteration (EDIFF, EDIFFG) to obtain more accurate forces and a converged geometry. Also check that all ionic steps are converged within NELM electronic steps (default 60).
To get very accuracte forces (if required) you can also set PREC=Accurate and ADDGRID =.TRUE.
The calculation time will e.g. depend on your settings (e.g. NPAR, NBANDS, the number of electronic/ionic steps), the interconnect and your cluster environment, and the compiler optimization. So your stated time difference cannot be evaluated without much more information.
Cheers,
/Dan
However, for the larger energy difference as in your case, you have presumably ended up in two different geometries. Please improve the break conditions for both the electronic iteration and the ionic iteration (EDIFF, EDIFFG) to obtain more accurate forces and a converged geometry. Also check that all ionic steps are converged within NELM electronic steps (default 60).
To get very accuracte forces (if required) you can also set PREC=Accurate and ADDGRID =.TRUE.
The calculation time will e.g. depend on your settings (e.g. NPAR, NBANDS, the number of electronic/ionic steps), the interconnect and your cluster environment, and the compiler optimization. So your stated time difference cannot be evaluated without much more information.
Cheers,
/Dan
Last edited by forsdan on Wed Oct 13, 2010 2:55 pm, edited 1 time in total.