Hello,
I am making a test jobs with respect to the number of processors used for one system: the same job (same files) using 16 or 32 processors. I found that the results differ:
for the 16 proc I obtained (step 92 of the relaxation):
92 FORCES: Max/Atom : 0.071775 RMS : 0.018997
92 free energy TOTEN = -912.039181 eV
for the 32 proc I obtained (step 92 of the relaxation):
92 FORCES: Max/Atom : 0.067238 RMS : 0.018041
92 free energy TOTEN = -912.037800 eV
The final relaxed geometry is not the same whethear I use 16 or 32 processors. And the relaxation at 0.05 eV/A is obtained after 158 steps for 16 processors:
158 F= -.91205496E+03 E0= -.91205496E+03 d E =-.264818E-03
FORCES: Max/Atom : 0.048071 RMS : 0.013404
free energy TOTEN = -912.054962 eV
And for 32 processors the relaxed geometry is found after 286 steps:
286 F= -.91218177E+03 E0= -.91218177E+03 d E =-.212863E-03
FORCES: Max/Atom : 0.049968 RMS : 0.012545
free energy TOTEN = -912.181769 eV
Can somebody explain me the difference due to the number of processors used
Thank you in advance.
Pierre.
Is it normal ?
Accuracy vs. Nb of porcessors
Moderators: Global Moderator, Moderator
-
- Newbie
- Posts: 29
- Joined: Mon Jul 03, 2006 11:19 am
Accuracy vs. Nb of porcessors
Last edited by pmignon on Thu Jun 05, 2008 12:56 pm, edited 1 time in total.
-
- Administrator
- Posts: 2921
- Joined: Tue Aug 03, 2004 8:18 am
- License Nr.: 458
Accuracy vs. Nb of porcessors
most probably, the difference is due to the following: the parameter NBANDS is automatically set to an integer multiple of the number of processors. this may lead to small differences in the results (in the range as you report). To be absolutely sure that this is the reason for the difference, please run the job on 16 processors, giving the same NBANDS as was used in the 32-processor job explicitely in INCAR.
Last edited by admin on Thu Jun 05, 2008 1:02 pm, edited 1 time in total.
-
- Newbie
- Posts: 29
- Joined: Mon Jul 03, 2006 11:19 am
Accuracy vs. Nb of porcessors
Ok thank you I launch the test job.
Another question: The calculation depends on the NBANDS. So I have to set it before the calculation starts.
For the system I study now, the formula: "NBANDS = NELECT/2 + NIONS/2" gives a NBANDS smaller than the one used in the 32 or 16 processors calculations:
NBANDS = NELECT/2 + NIONS/2 = 357
NBANDS (32 proc) = 384
NBANDS (16 proc) = 368
How to set NBANDS ? Is the choice "NBANDS = NELECT/2 + NIONS/2" is the best
Thank you
Pierre.
Another question: The calculation depends on the NBANDS. So I have to set it before the calculation starts.
For the system I study now, the formula: "NBANDS = NELECT/2 + NIONS/2" gives a NBANDS smaller than the one used in the 32 or 16 processors calculations:
NBANDS = NELECT/2 + NIONS/2 = 357
NBANDS (32 proc) = 384
NBANDS (16 proc) = 368
How to set NBANDS ? Is the choice "NBANDS = NELECT/2 + NIONS/2" is the best
Thank you
Pierre.
Last edited by pmignon on Thu Jun 05, 2008 2:16 pm, edited 1 time in total.