Page 1 of 1

Want to reproduce exact same broken SCF calculation

Posted: Thu Sep 16, 2021 11:41 pm
by ErnekMB
I would like to start a job with the same initial conditions as a previous job that crashed due to cluster time constraints.
When ISTART=0 , INIWAV is set to 1 by default, so the "wavefunction arrays are filled with random numbers". Relaunching the calculation with the same number of resources and same initial input files gives different
'entropy T*S EENTRO' and 'eigenvalues EBANDS = ' from the very first Iteration step.

Questions:
1) I am assuming this is happening due a different initialization of those random numbers. Is that correct or is there something else happening that I might be missing from the OUTCAR file when NWRITE = 2 ?

2) Is there a way to relaunch the same calculation that crashed and get the exact same initial conditions and subsequent steps?

Thank you

Re: Want to reproduce exact same broken SCF calculation

Posted: Mon Sep 20, 2021 7:37 am
by ferenc_karsai
The random number generators use different seeds and if you run on multiple cores your results are never the same even if you don't use random numbers, due to finite numerical accuracies.

So getting out exactly the same is impossible, but the results should be fairly reproducable within the few last digits if you use the same plane wave coefficients as staring basis (so no random numbers).