Understanding Dielectric calculations for Si

[KWH]

I am trying to reproduce the Si dielectric property example and I am a bit confused. I started with just a normal optimization of the structure [Si_only.rar]. I then used the INCAR POSCAR (contcar) KPOINTS WAVECAR POTCAR INCAR vasprun.xml IBZKPT DOSCAR and EIGENVAL files to my scratch directory and ran the batch file v6.sh. [Si_static.rar] As in the example, I only included the occupied bands (4) in the second calculation. In the first, VASP defaulted to 8 band (4 occupied and 4 unoccupied).

The second run was full of warnings and errors (see the out file) but the thing that stood out for me was "This is a result of running the parallel version.". While this is a small enough example to run on one cpu, the problems I want to do have more bands that the number of cpu I own, so parallel processes are essential.

Can you please clarify for me exactly what is my error, and how I should proceed if I had 2 nodes with 64 cpu each and of the order of 300 bands?

Best wishes

KWH

[andreas.singraber]

Hello!

Sorry for this late reply! I assume you got the example from this page on our Wiki: https://vasp.at/wiki/Dielectric_properties_of_Si, correct? While the example certainly contains the essential INCAR tags to try it out there is little context about the calculation. May I suggest to have a look at the more recent tutorials, specifically for linear response here: https://www.vasp.at/tutorials/latest/re ... sponse-e01. There the difference between LCALCEPS and LEPSILON is discussed (see also the Wiki page here). In the data you attached you did perform both calculations at the same time. Maybe this is technically possible but I am pretty sure this was not intended in the example. Rather you should either use the INCAR tags in this section

Code: Select all

## Static dielectric properties by means of DFPT
#NBANDS = 4
#EDIFF = 1E-6
#LEPSILON = .TRUE.
## try to add this to the DFPT calculation
#LPEAD = .TRUE.
## to get the ionic contributions to the
## static dielectric properties from
## perturbation theory
#IBRION = 8

or in the other section:

Code: Select all

## Static dielectric properties by means of PEAD
#EDIFF = 1E-8      # finite field requires very tight convergence
#LCALCEPS = .TRUE.
#NELM = 100

Generally, the examples and tutorials are set up in such way that they require a minimum amount of computational resources. Hence, VASP will not perform well if many MPI ranks are used here. Even worse, since VASP requires at least one band for each MPI rank, the number of bands in your second run was automatically increased to 40 just because that is also the number of MPI ranks used. Because this technical requirement contradicted the input given in the INCAR file, VASP issues the warning message you mentioned. For more details about the parallelization have a look at our corresponding Wiki page and NBANDS. There is also a video online from our team which explains a lot about parallelization in VASP : https://www.youtube.com/watch?v=KzIuL_e0zz8.

Since you mentioned specific numbers (order of 300 bands, 2 nodes with 64-cores each): whether the calculation you have in mind will actually saturate the CPU cores of the given machines needs to be tested carefully. I would recommend to benchmark with different number of cores before starting into actual production runs. Some hints on this topic are given on our Wiki here.

All the best,
Andreas Singraber