Resource reduction for calculation of electronic band structure (PROCAR)

[jeffrey_zom]

Hi,

I'd like to calculate the electronic band structure of a supercell (80 atoms, 640 electrons) that contains a small distortion.
However, this calculation takes a significant amount of time (on the order of a day), and appearently requires a lot of memory.

I've only been able to not run into memory issues when combining the following approaches:
- use of a high amount of cores (64 over 4 nodes)
- removing any parallelisation (KPAR = 1, NPAR = 1)
- reducing the amount of k-points to 10 points between high symmetry points in the IBZ (90 total).

Is there a way to reduce the amount of resources?
I'd like to increase the amount of k-points, but I can not do that without running into memory issues.
For instance, Is it possible to calculate the PROCAR for a certain energy interval of interest, similar to EINT used for PARCHG?

Kind regards,

Jeffrey

[jonathan_lahnsteiner2]

Dear jeffrey_zom,

It seems a little odd, that a band structure calculation will take 1-day or more. Does this
Does the time interval also contain the SCF calculation?
What would also be interesting to know is which functional you are using.

Could you please supply your input files according to the vasp forum guidelines ?
Then I will take a look to see what is going wrong.


All the best Jonathan

[jeffrey_zom]

Hi Jonathan,

- I've used ICHARG=11 and provided the CHGCAR from a SCF calculation.
So this time interval should not contain the SCF calculation.
- I am doing GGA (Perdew-Burke-Ernzerhof) + U calculations.

I've provided the input files (except for the CHGCAR) and the outcar in the attachment.

Kind regards,

Jeffrey

[jonathan_lahnsteiner2]

Dear Jeffrey,

I tested your calculation now. First I did a SCF calculation with a 2x2x2 kpoints grid. After that I used the input files you sent and computed the band structure.
It took 1.95h to compute the band structure on 64 cores. The memory consumption per core was roughly 1GB which makes a total memory consumption of around 64GB.
You can check the memory requirements in the OUTCAR file by

Code: Select all

grep -i memory OUTCAR

which will give you the line

Code: Select all

total amount of memory used by VASP MPI-rank0   959317. kBytes

This number multiplied by the number of cores on which you are running will give you the total amount of used memory.

To reduce the memory consumption for your calculation you could chop your KPOINTS file into several smaller paths.
For your example I tried to only compute the band structure along the points Γ-X-S-Y. This reduced the compute time to 0.9h
and the memory consumption per core was reduced to 0.35GB. This makes a total memory consumption on 64 cores of 22.6GB.
So I would suggest to use smaller windows of the total kpoints path you want to analyse and then add the results together during post processing.
There is currently no EINT option for the PROCAR file as you mentioned in your post.

I hope this is of help.
All the best Jonathan

[jeffrey_zom]

Hi Jonathan,

Thank you very much for doing the testing for me.
I wonder why calculation on the cluster I use takes significantly longer despite the same settings.
I like the idea of splitting up the calculation, although it will make the postprocessing take some additional time.
I will use your advice.

Kind regards,

Jeffrey