How to set MAXMEM for GW calculations?

[pnikacevic]

I'm trying to run single-shot GW calculations and they keep crashing, with the following message near the end of the OUTCAR:

This job will probably crash, due to insufficient avaiable memory.
Available memory per mpi rank: 15046 MB, required memory: 3600 MB.
Reducing NTAUPAR or using more computing nodes might solve this
problem.

I set MAXMEM=3600 (as before it was 6800 by default and crashing), but it crashed again. What I don't understand is this warning message that I got; why does the job crash when required memory is much lower than the available memory per mpi rank? How can I set it so it doesn't crash?

I attach the OUTCAR, POSCAR, INCAR, job script, slurm output file.

[fabien_tran1]

Hi,

Before I look into detail at your case, I was wondering if the problem reported at forum/viewtopic.php?f=4&t=18449&p=21697 ... mem#p21697 may be related to your case. Did you have a look?

[merzuk.kaltak]

You study a reasonably large system with the quartic scaling GW algorithm.
I strongly suggest to switch to the low-scaling GW implementation. Also, please update your vasp version from 6.1.0 to 6.3.2. There is a bug in 6.2.1 concerning the memory prediction for GW jobs as explained here.

In general, we propose following steps when encountering memory issues running GW jobs:

• reduce PREC from Accurate -> Normal (or even Single)

• for large cells (like yours) set LREAL=Auto in the INCAR. Note, the vasp code prints a corresponding warning in the stdout.

• reduce the number of k-points (you have already done that)

• use NTAUPAR=1 for low-scaling GW jobs (unnecessary with vasp-6.3.1 and newer, since NTAUPAR is set internally based on available memory).

Running the job of following INCAR file with the gamma-only version of vasp-6.3.2 requires around 400 GB of memory.

Code: Select all

ISMEAR  = 0
SIGMA   = 0.05
LREAL = A 
EDIFF   = 1E-8
ISPIN = 2 
LASPH=.TRUE.
NEDOS=5000

ALGO = G0W0R 
NBANDS=1152