Crash when running GW for a large system

[stefan19rkc]

Dear all,

I am trying to run Yambo on a large system (832 electrons) with low accuracy to examine the scaling and convergence. However, my runs often crash with the following message in some of the log files:

<02s> P75: [05] Coloumb potential CutOffslab
P75: [ERROR] STOP signal received while in[05] Coloumb potential CutOffslab
P75: [ERROR] Reading File ./gw_rimw_ppa_3Ry_gap//ndb.cutoff; Variable NOT DEFINED; NetCDF: Unknown file format

I tried examining what causes the crash, but I can't figure it out as it sometimes works and sometimes not (e.g. the same code with NGsBlkXp=2Ry instead of 3Ry worked fine). Please find the relevant files attached.

Kind regards,
Stefan Velja
PhD Student
Carl von Ossietzky University of Oldenburg

[claudio]

Dear Stefan

this crash can be due to a memory problem, I propose you some possible solutions:

1) compile Yambo with the openMP support using --enable-open-mp and then run calculations
using 2 or 4 threads, typically setting OMP_NUM_THREADS=2 or 4, this will double or quadruple the available memory

2) parallelize your calculation on bands, add the flag "-V par" in the input generation and set all parallelization
on valence and conduction bands

3) increase the number of G-vectors much slowly, for example use:
NGsBlkXp=2000 mRy, then NGsBlkXp=2200 Ry and so on

4) decrease the total numer of G-vectors in the calculation, add the flag "-V RL" in the input generation
and decreasi FFTGvecs to 1/2 or 1/3 of its value. You can check the effect recalculating optics "yambo -o c -V RL" with different number of G-vectors,
and see it the dipoles (and the spectrum) change.

best
Claudio

[Daniele Varsano]

Dear Stefan,
I'm not sure you this is a memory problem,
it seems you have just 1 k point and I suspect that the rimW can have problem in this situation (e.g. problem in interpolating W).

What's the reason behind using rimW and slab-z geometry with a gamma only calculation?

Best,
Daniele