Dear Admin,
Single point calculations with MBJ functional (not LMBJ) seems about 4 times slower in vasp6.3.2 compared to vasp5.4.4. I see that quite a few things have changed in the src code between these two versions. But I could not figure out which part is responsible for slowing down.
I also see some bug fixes related to LMBJ in the new vasp6.4 release. But at the moment I do not have access to the latest version in my HPC center. So, I do not know if the problem has already been fixed in the latest version.
Please find my test examples in the attached zip file.
Any suggestions or comments on this issue.
Thank you
Performance issue with MBJ functional in vasp6.3.2
Moderators: Global Moderator, Moderator
-
ralf_tonner1
- Newbie
- Posts: 17
- Joined: Mon Dec 21, 2020 2:19 pm
Performance issue with MBJ functional in vasp6.3.2
You do not have the required permissions to view the files attached to this post.
-
fabien_tran1
- Global Moderator
- Posts: 418
- Joined: Mon Sep 13, 2021 11:02 am
Re: Performance issue with MBJ functional in vasp6.3.2
Hi,
In the OUTCAR files that you provided the elapsed time is
VASP5.4.4/without_SOC/OUTCAR: Elapsed time (sec): 193.947
VASP5.4.4/with_SOC/OUTCAR: Elapsed time (sec): 1128.975
VASP6.3.2/without_SOC/OUTCAR: Elapsed time (sec): 315.258
VASP6.3.2/with_SOC/OUTCAR: Elapsed time (sec): 1978.097
Thus, with 6.3.2 the calculations are roughly 1.7 times slower (and not 4 times slower as you mentioned), and I could reproduce this trend myself by repeating your MBJ calculations. I was not aware of this difference in the timing with MBJ between the various versions of VASP. I will investigate to figure out why it is the case and if this is specific to MBJ or if the same behavior is obtained for other functionals.
Indeed, some bugs in LMBJ have been fixed. Starting from VASP.6.4.1, LMBJ should be bug free.
In the OUTCAR files that you provided the elapsed time is
VASP5.4.4/without_SOC/OUTCAR: Elapsed time (sec): 193.947
VASP5.4.4/with_SOC/OUTCAR: Elapsed time (sec): 1128.975
VASP6.3.2/without_SOC/OUTCAR: Elapsed time (sec): 315.258
VASP6.3.2/with_SOC/OUTCAR: Elapsed time (sec): 1978.097
Thus, with 6.3.2 the calculations are roughly 1.7 times slower (and not 4 times slower as you mentioned), and I could reproduce this trend myself by repeating your MBJ calculations. I was not aware of this difference in the timing with MBJ between the various versions of VASP. I will investigate to figure out why it is the case and if this is specific to MBJ or if the same behavior is obtained for other functionals.
Indeed, some bugs in LMBJ have been fixed. Starting from VASP.6.4.1, LMBJ should be bug free.
-
ralf_tonner1
- Newbie
- Posts: 17
- Joined: Mon Dec 21, 2020 2:19 pm
Re: Performance issue with MBJ functional in vasp6.3.2
Dear Fabien,
Thank you very much for pointing the time measurement discrepancy. I agree with your slow down measurement based on the total elapsed time data. However, I would like to point out that VASP5.4.4 took more iteration steps to reach the convergence than VASP6.3.2.
Based on the OSZICAR file:
VASP5.4.4/without_SOC took 42 iterations
VASP5.4.4/with_SOC took 47 iterations
VASP6.3.2/without_SOC took 36 iterations
VASP6.3.2/with_SOC took 32 iterations
The time measurement I gave was based on average time taken (approx) for RMM-DIIS algorithms only. ('grep RMM-DIIS OUTCAR')
I have now included exact numbers for without_SOC calculations. Here is the summary. I am also attaching the figures. In the figures, I only included the (real) time taken for the RMM-DIIS, CHARGE, EDDIAG, and LOOP in each iteration. Overall primary observation showed that the first three contributions changed the most between the two versions.
VASP6.3.2 RMM-DIIS is on an average 3.5 times slower than VASP5.4.4. ('grep RMM-DIIS OUTCAR')
VASP6.3.2 CHARGE is on an average 2.6 times faster than VASP5.4.4. ('grep CHARGE OUTCAR')
VASP6.3.2 EDDIAG is on an average 3.66 times slower than VASP5.4.4. ('grep EDDIAG OUTCAR')
VASP6.3.2 LOOP is on an average 2 times slower than VASP5.4.4. ('grep LOOP OUTCAR'; this includes the first 5 Davidson steps as well)
I hope this will clarify the time measurement comparison better.
Thank you
Thank you very much for pointing the time measurement discrepancy. I agree with your slow down measurement based on the total elapsed time data. However, I would like to point out that VASP5.4.4 took more iteration steps to reach the convergence than VASP6.3.2.
Based on the OSZICAR file:
VASP5.4.4/without_SOC took 42 iterations
VASP5.4.4/with_SOC took 47 iterations
VASP6.3.2/without_SOC took 36 iterations
VASP6.3.2/with_SOC took 32 iterations
The time measurement I gave was based on average time taken (approx) for RMM-DIIS algorithms only. ('grep RMM-DIIS OUTCAR')
I have now included exact numbers for without_SOC calculations. Here is the summary. I am also attaching the figures. In the figures, I only included the (real) time taken for the RMM-DIIS, CHARGE, EDDIAG, and LOOP in each iteration. Overall primary observation showed that the first three contributions changed the most between the two versions.
VASP6.3.2 RMM-DIIS is on an average 3.5 times slower than VASP5.4.4. ('grep RMM-DIIS OUTCAR')
VASP6.3.2 CHARGE is on an average 2.6 times faster than VASP5.4.4. ('grep CHARGE OUTCAR')
VASP6.3.2 EDDIAG is on an average 3.66 times slower than VASP5.4.4. ('grep EDDIAG OUTCAR')
VASP6.3.2 LOOP is on an average 2 times slower than VASP5.4.4. ('grep LOOP OUTCAR'; this includes the first 5 Davidson steps as well)
I hope this will clarify the time measurement comparison better.
Thank you
You do not have the required permissions to view the files attached to this post.
-
fabien_tran1
- Global Moderator
- Posts: 418
- Joined: Mon Sep 13, 2021 11:02 am
Re: Performance issue with MBJ functional in vasp6.3.2
Yes sorry you are right. I was too fast and forgot to check the number of iterations. I will try to figure out at which point in the VASP history MBJ calculations started to be slower.