Issues during optical properties calculations

[IBRAHIM]

Dear All,

During calculation of the optical properties using HSE functional in VASP, I noticed a strange issue. During run the same job from scratch several times I noticed changes in N E dE d eps ncg rms OR ort in OSZICAR file.
Even this difference can be negligible, but I think we must find no difference. Why this is happened?

The first steps are posted below in the two runs:

N E dE d eps ncg rms ort
SDA: 1 -0.108403354357E+04 -0.10840E+04 0.00000E+00 2160 0.655E-02 0.000E+00
CGA: 2 -0.108403417504E+04 -0.63147E-03 -0.61711E-03 2160 0.649E-02 0.326E-02
CGA: 3 -0.108403458221E+04 -0.40718E-03 -0.73332E-03 2160 0.170E-02-0.135E-02
CGA: 4 -0.108403471428E+04 -0.13207E-03 -0.11697E-03 2160 0.107E-02 0.234E-03
CGA: 5 -0.108403479044E+04 -0.76158E-04 -0.87642E-04 2160 0.651E-03 0.282E-03


N E dE d eps ncg rms ort
SDA: 1 -0.108403354357E+04 -0.10840E+04 0.00000E+00 2160 0.661E-02 0.000E+00
CGA: 2 -0.108403416401E+04 -0.62044E-03 -0.60640E-03 2160 0.663E-02 0.318E-02
CGA: 3 -0.108403458665E+04 -0.42264E-03 -0.74763E-03 2160 0.165E-02-0.933E-03
CGA: 4 -0.108403471646E+04 -0.12981E-03 -0.12301E-03 2160 0.112E-02 0.241E-03
CGA: 5 -0.108403478619E+04 -0.69730E-04 -0.82160E-04 2160 0.849E-03 0.258E-03

Many thanks in advance,
Ibrahim

[martin.schlipf]

When running calculations in parallel minor differences between the energies are expected, because the order of operations is no longer guaranteed. In your case there seems to be a 0.004meV difference, which is reasonable. It may even get a bit smaller if you run both calculations until convergence.

[IBRAHIM]

Thanks for reply.

I would like to mention that in the two runs everything is the same; cpu and input files. So we should expect no difference. I found that happened only in optical calculations with hybrid functional. Note in other normal calculations no difference found.

Best regards,
Ibrahim

[martin.schlipf]

Sorry, but did you run in parallel (either MPI or OpenMP) or in serial?

Parallel calculations are not deterministic, typically the changes are small, but in general hybrid functional with their HF operator are "rougher" and more difficult to get accurate than local DFT.

[IBRAHIM]

OK, thank you very much.
Yes, I did parallel calculations.

So as I understood that difference is not a bug, right? I was just curious that it can be a bug.

Best regards,
Ibrahim

[ferenc_karsai]

It's not a bug. Deviations depend also on your hardware, compiler versions and on how many cores you use in your calculations.

Nevertheless the deviations should be only in the last few digits.

[IBRAHIM]

OK, thank you very much.

In my calculations, everything is the same, cpu, compilar, input, etc. However, as you explained, that difference can be normal.

Thanks again for your explanations.

Best regards,
Ibrahim