Yambo Community Forum

Dear Daniele,

I'm doing two calculations one without Spin-Orbit Coupling (SOC) and one without SOC. Actually my material contains heavy elements, that's why I guess, SOC will be vital.

Q1: I want to ask that what parametes do I need to have for SOC calculation, I have done the scf and nscf from Quantum ESPRESSO with the following settings:

noncolin =.TRUE.
lspinorb =.TRUE.

Apart from it do I need to redefine my k-points settings? And will Yambo automatically consider it to SOC calculation and will it break the time-reversal symmetry, for SOC brakes this symmetry! Correct?

Q2: Since my material is a transition metal oxide and I am finding optical properties using Yambo, I want to apply different levels of theory unless I get close to the experimental curve. But I don't want to use LDA for oxide, which means I will not be able to use ALDA and LRC kernal options! Am I right?

Q3: Are both LRC and ALDA kernel falls under the umbrella of TDDFT?


Regards,
Haseeb Ahmad,
LUMS - Pakistan.

Dear Haseeb,
A1: yes these are the settings to include SOC, provided you have full relativistic pseudopotentials. You do not need to redefine your k point grid as the integration of BZ is independent of the fact you have spinors or not. What you need to take care is the number of bands as now you have spinors so the number of bands is double with respect to the non-spin resolved case. SOC does not break the TR.

A2: If you do not want to use LDA you should not use ALDA. LRC is a different story as contrary to ALDA it behaves as 1/q^2 and it is a good option for solids but it is not parameter-free.

A3: ALDA can be a reasonable option for finite systems (molecules) but it has rather poor effects in bulk. LRC can provide good results if the parameters are set properly, but these are system-dependent (see Botti Phys. Rev. B 69, 155112 2014).

As a general comment, it is important you discuss these issues with your colleagues and your supervisor as they are general questions about the physics of your system and not strictly related to the Yambo code.

Best,
Daniele

Dear Yambo Developers,

The following question is explicitly related to GS calculations from QE; however, I am having some issues before starting to deploy the Yambo. I want to test PBE0 functional for GS calculations (before going to Yambo).

In PBE0, I have the issue of memory. One node has 128 GB, but the calculation size can't fit in 128 GBs. Can I parallelize over memory (just like we do in Yambo)? If yes, then what would be the best way to do that, please?

I tried using multiple nodes for memory handling, but the calculation became too slow because CPU usages were mere 10% per node! So, I thought, I'm not exploiting the parallel structure properly.

I am not using any "npool, ntg" things:
In Slurm, I'm invoking pw.x in the batch script via:

mpirun -np $ncpu $bindir/pw.x -input scf-PBE0_666-222.in > scf-PBE0_666-222.out

The input and output files (initial part of out only) are attached herein.

Thanks in advance,
Haseeb

Dear Haseeb,

You should probably write to the QE forum in order to have useful advises.
Are you sure you are going out of memory?
I any case my advise is to use more nodes and use the npool option. npools must be a divisor of the total number of k-points.

Not related with your problem but not specifying exxdiv_treatment it is not a good idea.

Best,

Daniele

Dear Daniele,

Thanks for your response!

Not related with your problem but not specifying exxdiv_treatment it is not a good idea.

Yes, I will learn more about it and apply ws_cut, as you suggested in the other thread too!

Thanks again,

Dear Haseeb,

additionally to Daniele's suggestions, you may want to

(1) explicitly set diago_david_ndim=2 : this is now the default for davidson diagonalization, while it used to be 4 in previous versions (you are using qe-6.5
and I don't remember whether the default was already changed or not); in any case, probably a good idea to switch to newer versions of qe.

(2) your plane-wave parallelisation is boarder line (see the warning: "some processors have no G-vectors for symmetrization"), you may want to use 2 task groups
( -ntg 2 ) or to increase the number of pools. Using pools would be ideal, though more memory is typically required.

(3) asking for 500 bands with davidson, with ndim=2 you are diagonalising 1000x1000 dense matrices, 2000x2000 with ndim=4, perhaps you want to set -ndiag=4,16,25,... or so (run some tests, results may be very diverse and counterintuitive)

(4) diagonalization "cg" would take less memory but is usually very slow

(5) Given your input file, I am not sure to properly understand why you have the message

"EXX: setup a grid of 160 q-points centered on each k-point
(set verbosity='high' to see the list)",

perhaps better to set verbosity="high"

take care
Andrea

Thanks, dear Andrea for these useful comments.

Regards,

Haseeb