Memory RAM usage

[Fabiof]

Dear all,


Im doing same test calculations with BN.

My PWscf file has

Code: Select all

 ecut = 50Ry,  and 12x12x1  K points

I am just wondering if it is normal, for a system like BN, the memory RAM usage at the beginning (i think at the point it is calculating the HF energies) using like 30GB of memory RAM?

after HF energies it drops to 2GB. Here is the yambo file:

Code: Select all

em1s                         # [R Xs] Static Inverse Dielectric Matrix
gw0                          # [R GW] GoWo Quasiparticle energy levels
cohsex                       # [R Xp] COlumb Hole Screened EXchange
HF_and_locXC                 # [R XX] Hartree-Fock Self-energy and Vxc
ElecTemp=  0.02585     eV    # Electronic Temperature
EXXRLvcs=  20595       RL    # [XX] Exchange RL components
Chimod= "hartree"            # [X] IP/Hartree/ALDA/LRC/BSfxc
% BndsRnXs
   1 | 100 |                 # [Xs] Polarization function bands
%
NGsBlkXs= yy            RL    # [Xs] Response block size
% LongDrXs
 1.000000 | 0.000000 | 0.000000 |        # [Xs] [cc] Electric Field
%
% GbndRnge
   1 | 100 |                 # [GW] G[W] bands range
%
%QPkrange                    # [GW] QP generalized Kpoint/Band indices
  1| 19|  1|100|
%
%QPerange                    # [GW] QP generalized Kpoint/Energy indices
  1| 19| 0.0|-1.0|

And if increase bands of Polarization function bands and G[W] bands range by 50, the memory RAM increases a lot. But the thing is that this happens only at the beginning, because when it is computing the qp energies the memory ram usage is so low.

[Daniele Varsano]

Dear Fabio,
first of all, I do not understand what does it means:

Code: Select all

NGsBlkXs= yy  

Next please note that you are calculating:

Code: Select all

%QPkrange                    # [GW] QP generalized Kpoint/Band indices
  1| 19|  1|100|
%

QP corrections for 1900 states. Do you need all of them?
Memory usage is essentially needed for the wfs, you can lower the amount of memory by reducing the G vector component using the variable
FFTGvecs in your input file (usually this can be reduced a bit without loosing accuracy, but not too much as then orthonormality conditions are not verified anymore).
If you still have problem, please post report and standard output file in order to have a look exactly where you are experiencing problem.

Best,
Daniele

[Fabiof]

Dear Daniele,

thanks for the quick answer.

Code: Select all

NGsBlkXs= yy   

yy is just a variable to study the convergence.

I changed

Code: Select all

%QPkrange                    # [GW] QP generalized Kpoint/Band indices
  1| 19|  1|100|
%

to

Code: Select all

%QPkrange                    # [GW] QP generalized Kpoint/Band indices
  1| 19|  1|20|
%

and it reduces the memory usage significantly. I didnt know that i could study the convergence of BndsRnXs and GbndRnge increasing it but computing only the QP sates that i need.

Thanks for the help!

[Fabiof]

Dear Daniele,

one question:

when i decrease the QPkrange, am i decreasing GbndRnge?

For example, in my input files i have

Code: Select all

% GbndRnge
   1 | 50 |                 # [GW] G[W] bands range
%

but in my report file i have

Code: Select all

% GbndRnge
 |   1 | 18 |                   # [GW] G[W] bands range
 | %

[Daniele Varsano]

Dear Fabio,
very strange. The two variable are totally independent.
Are you doing GW or Cohsex calculations? In the latter case the default algorithm needs occupied bands only, so you do not need to check
GbndRnge as convergence parameter, so in the report you will find as GbndRnge the number of occupied variable, independently on what you set in input.
You can also perform cohsex using empty bands (different algorithm): this is activated by unchecking the UseEbands flag, but there is not reason to do that.
If you are doing GW the only reason it can happen is that you have stored 18 bands only in your databases, ie you calculated 18 bands only in the ground state calculations.

Best,
Daniele

[Fabiof]

Thanks for the answer, im doing Cohsex calculations.