Inconsistance in estimated and used memory
I see inconsistent set of numbers for estimated and used memory.
Here is an example for exactly the same model with variable parallelization numbers of kpts, domain, and band.
You can run the script as:
gpaw python input.py K_B_D_KPTS_VAC_CUTOFF
where K, B, and D are parallelization numbers of kpts, domain, and band; KPTS, VAC, and CUTOFF are parameters of the model – k_x,y-points, vacuum, and PW cutoff.
For example:
gpaw python input.py 2_4_2_2_4_340
I also attach a slurm submission script.
The estimated memory is from "Memory estimate: Calculator: VALUE" and the used memory is from the "Memory usage: VALUE" of the standard output file.
The minor issue is the inconstancy in units – it is better to give both memory values in Mb.
The major issue is the inverse correlation of estimated memory vs used memory (15 points for various combination of K×D×B = 16):
Either I am doing something very wrong – please correct me – or there is a bug in memory estimation. In both cases, I am eager to help.
P.S. I am using GPAW 22.8 from conda-forge. All other parameters are in the files.