ABINIT Discussion Forums

Hi,

I'm trying to study a large system using NCPP, with the goal ultimately of computing rstrs, rfphon at gamma point, and rfelfd. I can compute the ground state and also do geometry optimization using paral_kgb, but I cannot ever get the response function calculations to read the output. For example, I run on 36 processors with the following options:

paral_kgb 1
npkpt 9
npspinor 1
npband 4
npfft 1
bandpp 4

localrdwf 0
accesswff 0

kptopt 2
iscf 7
tolvrs 1.0D-18


where the variables were tuned by a run with paral_kgb -36. It writes the output wavefunction to gso_WFK which I then re-name to gsi_WFK. Then, I run with exactly the same setup except replacing iscf7, tolvrs 1.0D-18 by

iscf -3
tolwfr 1.0D-20
irdwfk 1
rfddk 1
rfdir 1 1 1

and it immediately fails at the point of reading the GS wavefunction. If I try to run without paral_kgb in the second step it fails because the npw values are different than what is expected. Can someone suggest what I am doing wrong?

thanks,
Joe

Well, it appears from the log file (but not the on-line documentation) that paral_kgb is not compatible with accesswff 0, only with accesswff 1. I've re-run the job with ndtset 2, to compute both gs and ddk as a chain of calculations, using accesswff 1 and localrdwr 1, but it still crashes after the gs finishes. Any ideas?

Joe

Well, one problem I found was that I had been leaving dilatmx 1.05 in the input file for response calculation, left over from the geometry optimization, and that seems to be in conflict with response function calculations. Can someone confirm or deny? I can certainly check for it in chkinp.F90.

Nonetheless, I simply cannot get paral_kgb 1 to work when I am doing a response function calculation, only parallel over k points. Can someone comment on this?

Hello Joe

I think phonon calculations (with PP or PAW) are not parallelized on the kgb levels. I think you can only use it with the "regular" kpoint parallelization.

That's a problem for large supercells.

Boris

That is my impression, after some trial and error. I did eventually succeed in computing efield, strain, and phonons at gamma for a 68 atom cell (272 bands), primitive orthorhombic, with excellent agreement with heat capacity data, so I'm pleased for that--