Achieving step by step convergence by setting NELM
Posted: Tue Apr 29, 2014 9:32 am
I guess this is a bit of a weird question, and a strange thing to do, but it actually makes sense in the context of my situation. Basically I need to run hybrid functional calculations that typically take very long. The batch system of the cluster I am using favors short jobs and it is very difficult to get the whole job I need to do allocated.
The strategy I am using is to set NELM to some small value so that VASP has the time to finish each subcalculation within the time frame allowed by the batch system. Schematically, the logical sequence would be something like this:
What are the possible issues and performance losses associated with this approach, besides the obvious read/write of files?
<span class='smallblacktext'>[ Edited ]</span>
The strategy I am using is to set NELM to some small value so that VASP has the time to finish each subcalculation within the time frame allowed by the batch system. Schematically, the logical sequence would be something like this:
Code: Select all
cat >INCAR<<eof
General options for geometry etc.
Hybrid functional options
ISTART = 1 ! Restart from previous calculation
NSW = 0 ! Don't update ionic positions
NELM = 3 ! Write WF after 3 iterations
eof
for i in 1 2 3 4...
do
# Each of these is allocated by the queueing system
srun vasp
# This allows to monitor convergence
cp OSZICAR OSZICAR_step_${i}
done
<span class='smallblacktext'>[ Edited ]</span>