Hello, I'm trying to set up a VASP calculation for GW+BSE to obtain the dielectric function of a metal-oxo cluster structure. During the additional DFT "virtual" orbitals (empty states) calculation after the groundstate calculation, I encountered the following error, and the calculation terminates immediately without starting.
BAD TERMINATION OF ONE OF YOUR APPLICATION PROCESSES
RANK 0 PID 3296766 RUNNING AT n009
KILLED BY SIGNAL: 9 (Killed)
Does anyone know how to solve this? I've searched and found that it might be a memory issue, so I increased the number of cores (32->64->96), but it still didn't solve the problem. My INCAR file is as follows.
SYSTEM = metal-oxo cluster ALGO = Exact LREAL = Auto NELM = 1 ENCUT = 450 PREC = Normal ISMEAR = 0 SIGMA = 0.1 LOPTICS = .TRUE. OMEGAMAX = 100 NBANDS = 300 NCORE = 4
Thank you in advance.
Heey Woongpyo Jeon
There are a number of possibilities causing this error outside of your calculation, such as limitations in the environment where you’re running your calc etc. Approaching what we can easily access, based on your INCAR file and considering that it could potentially be a memory issue, you can reduce some computational demand by changing ALGO=Exact, which can be computationally very expensive, to something lighter. Additionally, NBANDS set to 300 might be too high or too low, depending on the size of your system. I understand that for GW calculations, you’ll typically need more bands than for a simple ground-state DFT calc. However, as a test, try reducing it to, let’s say, 250 to see if it helps. NCORE set to 4 increases performance but might also increase memory demand. Try NCORE=2 if memory issues persist. You can try these steps one at a time, following the aforementioned suggestions.
Extra thought: I didn’t get why you set NELM=1, which seems unusually low. Usually, people use the default value or set it to higher, not lower values.
Best,
Gabriel Vinicius
Gabriel Vinicius Thank you for your kind response!
I followed your advice and adjusted the NBANDS and NCORE values, but the issue still remains unresolved. However, when I changed the ALGO tag from Exact to Normal, the calculation proceeded.
The reason I set ALGO=Exact and NELM=1 was based on the information found here:
https://www.vasp.at/wiki/index.php/Dielectric_properties_of_Si_using_BSE
I'm now wondering if changing ALGO to Normal and increasing NELM value will cause any issues for subsequent GW+BSE calculations compared to the reference.
Thank you in advance.
Woongpyo Jeon
Woongpyo Jeon ALGO=Exact carries an exact diagonalization of the Hamiltonian, which is immensely memory (and to a lesser extent cpu) consuming for larger systems so you're almost certainly facing hardware (mostly memory) related issues.
Sadly I don't think there's a workaround as you need a large number of accurate converged unoccupied orbitals for GW/BSE (In fact I am not sure 300 will be enough and this needs to be tested), and usual iterative methods such as ALGO=normal converges extremely slowly for the top of the spectrum so that you actually need to increase NBANDS even further to be able to use them.
- Badges
- Science topic