As my calculation of paramagnetic susceptibility of transition metals does not come out ideally (for some metals, the predicted susceptibility is half while for others is double the experimental data), I am thinking of whether OEP-EXX will be better. Using a negative value of xctype (-3 or -20, for example), I found the program runs extremely slowly even when I set all other parameters small (such as the number of k-points and the gmax). I would like to know whether this slowdown is normal and whether setting the xctype negative is the only parameter related to oep-exx. Also the program prints something like info(oepvnl):xx of xx. Is this a warning message?
with best regard!
zhangyg
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I read carefully the previous forum messages. It seems that my question has already been answered clearly. I will give a try using the method you gave in the previous message.
Thanks.
I also hope that someone who is an expert on the first-principles calculations of paramagnetic susceptibility of nonmagnetic transition metals can give me a guidance or hint on how to predict the property accurately while I am studying very hard by myself.
zhangyg
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Could you kindly have a look at my input file below for an EXX-OEP calculation on hcp Titanium. The calculation is TOO slow. The program did not finish even ONE step within two days! For my computer, with much more K-points, and without EXX-OEP, the program can normally finish 50 steps for the same period. Did I do something wrong?
OEP-EXX is still an experimental feature, and so should not be used for production as yet.
It is a very slow technique because the exchange term in the Kohn-Sham equations is treated exactly. This requires the non-local Coulomb matrix elements for all k-point combinations, and so scales as the number of k-points squared. The most we've ever used is a 5x5x5 k-point set with one atom.
As we refine the technique it may get a bit faster, but not by much. It is parallel however, and the speed scales with the number of processors.
Cheers,
Kay.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I would like to tell you gladly that I have made EXCITING running under cluster-openmp using the compiler provided by Intel. I hope this would benefit you if you have a cluster of PCs. There is probably just one thing that need to be changed in the code. It is that cluster openmp doesn't support nested parallelism so that parallelism on nst, natoms, etc need to be commented out if there is a parallelism on nkpt already. Others things may be taken care of using simply compiler options.
I have checked results between cluster openmp and pure openmp using exciting.in in the subdirectory /Fe. The results are very close (last digit for total energy and last three digits for total moment). I am not sure if the calculated resuts are correct for OEP-EXX because I have never obtained a result using pure openmp and my test run for cluster openmp is still running. I will give you the result if the latter can be finished.
I am tired and going to sleep now. Hope you can give some comments.
with best wishes
zhangyg
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I played a little bit with OEP-EXX on MgO. I just want to see if OEP-EXX can give a more accurate EOS prediction than LDA and GGA. To my surprise, in the energy vs volume curve the energy increases continuously from small to large volume (fcc cell length from 7.25 to 9 Hartree) instead of a valley in between as seen in GGA. Could you kindly give me a hint on what is going wrong here.
Because of the small volume at high pressure, I used small muffin-tin radius (1.55 for Mg and 1.45 for O). The number of k-points is also small (3x3x3=6 points), but GGA can give correct results even with this number of k-points. other parameters are as follows:
Following my last message, I also noticed that EXCITING prints always exchange energy of 0.0000000, which is not zero only at the last step. Why is this the case?
zhangyg
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
The exact exchange energy is calculated only at the last iteration because it is fairly time-consuming. This is the energy you should use in EXX-OEP calculations. The previous energies are without exchange and should not be used.
Does this fix the MgO problem?
If not, try silicon and see if you get a parabola.
Regards,
Kay.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Dear Kay:
As my calculation of paramagnetic susceptibility of transition metals does not come out ideally (for some metals, the predicted susceptibility is half while for others is double the experimental data), I am thinking of whether OEP-EXX will be better. Using a negative value of xctype (-3 or -20, for example), I found the program runs extremely slowly even when I set all other parameters small (such as the number of k-points and the gmax). I would like to know whether this slowdown is normal and whether setting the xctype negative is the only parameter related to oep-exx. Also the program prints something like info(oepvnl):xx of xx. Is this a warning message?
with best regard!
zhangyg
Dear Kay:
I read carefully the previous forum messages. It seems that my question has already been answered clearly. I will give a try using the method you gave in the previous message.
Thanks.
I also hope that someone who is an expert on the first-principles calculations of paramagnetic susceptibility of nonmagnetic transition metals can give me a guidance or hint on how to predict the property accurately while I am studying very hard by myself.
zhangyg
Dear Kay:
Could you kindly have a look at my input file below for an EXX-OEP calculation on hcp Titanium. The calculation is TOO slow. The program did not finish even ONE step within two days! For my computer, with much more K-points, and without EXX-OEP, the program can normally finish 50 steps for the same period. Did I do something wrong?
tasks
0
maxscl
50
spinpol
.true.
spinorb
.true.
bfieldc
0.0 0.0 23.292e-4
autormt
.true.
rgkmax
7.0
gmaxvr
12.0
beta0
0.01
betamax
0.1
nempty
10
epspot
1.e-6
xctype
-20
stype
3
swidth
0.01
avec
0.5 -0.86602540378444 0.0
0.5 0.86602540378444 0.0
0.0 0.00000000000000 1.0
scale1
5.57620
scale2
5.57620
scale3
8.85431
sppath
'/home/zhang/program/exciting/species/'
atoms
1 : nspecies
'Ti.in' : spfname
2 : natoms
0.333333333333 0.666666666667 0.25 0.0 0.0 0.0 : atposl, bfcmt
0.666666666667 0.333333333333 0.75 0.0 0.0 0.0 : atposl, bfcmt
ngridk
8 8 6
vkloff
0.0 0.0 0.5
Thanks a lot.
with best wishes
zhangyg
Dear zhangyg,
OEP-EXX is still an experimental feature, and so should not be used for production as yet.
It is a very slow technique because the exchange term in the Kohn-Sham equations is treated exactly. This requires the non-local Coulomb matrix elements for all k-point combinations, and so scales as the number of k-points squared. The most we've ever used is a 5x5x5 k-point set with one atom.
As we refine the technique it may get a bit faster, but not by much. It is parallel however, and the speed scales with the number of processors.
Cheers,
Kay.
Dear Kay:
I would like to tell you gladly that I have made EXCITING running under cluster-openmp using the compiler provided by Intel. I hope this would benefit you if you have a cluster of PCs. There is probably just one thing that need to be changed in the code. It is that cluster openmp doesn't support nested parallelism so that parallelism on nst, natoms, etc need to be commented out if there is a parallelism on nkpt already. Others things may be taken care of using simply compiler options.
I have checked results between cluster openmp and pure openmp using exciting.in in the subdirectory /Fe. The results are very close (last digit for total energy and last three digits for total moment). I am not sure if the calculated resuts are correct for OEP-EXX because I have never obtained a result using pure openmp and my test run for cluster openmp is still running. I will give you the result if the latter can be finished.
I am tired and going to sleep now. Hope you can give some comments.
with best wishes
zhangyg
Dear Kay:
I played a little bit with OEP-EXX on MgO. I just want to see if OEP-EXX can give a more accurate EOS prediction than LDA and GGA. To my surprise, in the energy vs volume curve the energy increases continuously from small to large volume (fcc cell length from 7.25 to 9 Hartree) instead of a valley in between as seen in GGA. Could you kindly give me a hint on what is going wrong here.
Because of the small volume at high pressure, I used small muffin-tin radius (1.55 for Mg and 1.45 for O). The number of k-points is also small (3x3x3=6 points), but GGA can give correct results even with this number of k-points. other parameters are as follows:
tasks
0
rgkmax
8.0
gmaxvr
16.0
beta0
0.1
betamax
1.0
nempty
10
epspot
1.e-6
xctype
-20
stype
1
swidth
0.01
avec
0.0 0.5 0.5
0.5 0.0 0.5
0.5 0.5 0.0
scale
7.25
sppath
'/home/zhang/program/exciting/species/'
atoms
2 : nspecies
'Mg.in' : spfname
1 : natoms
0.0 0.0 0.0 0.0 0.0 0.0 : atposl, bfcmt
'O.in' : spfname
1 : natoms
0.5 0.5 0.5 0.0 0.0 0.0 : atposl, bfcmt
ngridk
3 3 3
vkloff
0.5 0.5 0.5
With best regards!
zhangyg
Dear Kay:
Following my last message, I also noticed that EXCITING prints always exchange energy of 0.0000000, which is not zero only at the last step. Why is this the case?
zhangyg
Dear Zhangyg,
The exact exchange energy is calculated only at the last iteration because it is fairly time-consuming. This is the energy you should use in EXX-OEP calculations. The previous energies are without exchange and should not be used.
Does this fix the MgO problem?
If not, try silicon and see if you get a parabola.
Regards,
Kay.