GW correction on large isolated system

You can post here problems arising when using the last release of Yambo. Issues as parallelization strategy, performance issues and other technical aspects realted to the new release.

Moderators: Davide Sangalli, Daniele Varsano, andrea.ferretti, andrea marini, Conor Hogan, myrta gruning

Post Reply
el16yz
Posts: 23
Joined: Mon Sep 23, 2019 4:34 pm
Location: United Kingdom

GW correction on large isolated system

Post by el16yz » Mon Aug 10, 2020 3:46 am

Dear Yambo users,

I am using Yambo 4.5.1 to perform GW correction on a large isolated system. The parallel calculation is applied for speeding up the process: 64 cores, 25G/core. The calculation keeps running 48 hs but doesn't finish. And there is not any new update in log files since 2 hours from calculation begin. From the report file, it seems that the process is trapped in the calculation of Dynamic Dielectric Matrix. The input, report, and log files are submitted as attachments. It will be very appreciated if you can check and solve my problem. Thank you.

Best,
Yang
You do not have the required permissions to view the files attached to this post.
Yang Zhou
PhD student
University of Leeds
United Kingdom

User avatar
claudio
Posts: 265
Joined: Tue Mar 31, 2009 11:33 pm
Contact:

Re: GW correction on large isolated system

Post by claudio » Wed Aug 12, 2020 1:15 pm

Dear Yang

I don't know why your calculation got stuck, anyway I give
some suggestions.

1) Use fewer processors for linear algebra, for example
X_nCPU_LinAlg_INV= 16

2) try to use a lower cutoff for the dielectric constant
NGsBlkXp= 5000 mHa # [Xp] Response block size

3) in the polarization you should put all the bands
1 | 400 | # [Xp] Polarization function bands

4) try to parallelization mode "memory"
PAR_def_mode= "memory"

and check if it works
best
Claudio
Claudio Attaccalite
[CNRS/ Aix-Marseille Université/ CINaM laborarory / TSN department
Campus de Luminy – Case 913
13288 MARSEILLE Cedex 09
web site: http://www.attaccalite.com
Freely download scientific books from: http://www.freescience.info

el16yz
Posts: 23
Joined: Mon Sep 23, 2019 4:34 pm
Location: United Kingdom

Re: GW correction on large isolated system

Post by el16yz » Wed Aug 12, 2020 4:41 pm

Dear Claudio,

Thanks for your suggestions.
2) try to use a lower cutoff for the dielectric constant
NGsBlkXp= 5000 mHa # [Xp] Response block size
I have tried to set NGsBlkXp= 1 Ry before. It can produce a result in a few minutes. But the result is an unexpected one, which may be due to convergence issues. Hence, I need to keep increasing NGsBlkXp until the result catches convergence. But the problem occurs when I increase it to 4 Ry. What is the possible reason for that?

Best,
Yamh
Yang Zhou
PhD student
University of Leeds
United Kingdom

el16yz
Posts: 23
Joined: Mon Sep 23, 2019 4:34 pm
Location: United Kingdom

Re: GW correction on large isolated system

Post by el16yz » Wed Aug 12, 2020 4:42 pm

Dear Claudio,

Thanks for your suggestions.
2) try to use a lower cutoff for the dielectric constant
NGsBlkXp= 5000 mHa # [Xp] Response block size
I have tried to set NGsBlkXp= 1 Ry before. It can produce a result in a few minutes. But the result is an unexpected one, which may be due to convergence issues. Hence, I need to keep increasing NGsBlkXp until the result catches convergence. But the problem occurs when I increase it to 4 Ry. What is the possible reason for that?

Best,
Yang
Yang Zhou
PhD student
University of Leeds
United Kingdom

User avatar
claudio
Posts: 265
Joined: Tue Mar 31, 2009 11:33 pm
Contact:

Re: GW correction on large isolated system

Post by claudio » Thu Aug 13, 2020 8:25 am

Dear Yang

GW convergence of isolated systems can be quite complicated.
First of all keep NGsBlkXp= 5000 mHa, this value is usually enough large
to have converged results.

Then you have to number of bands in the G and W. I think 400 is a small number,
you should have something of the order of 1000 or 2000 bands.

Then you have to consider the size of the supercell, that should be
large at least double of your cluster or molecule,
and then you should use a cutoff to reduce the interaction with the neighbor replica,
you can activate it with the flag "-r".


Have a look here:
http://www.yambo-code.org/wiki/index.ph ... rial:_h-BN
http://www.attaccalite.com/yambo-input- ... explained/
http://www.attaccalite.com/reasonable-p ... culations/

best
Claudio
Claudio Attaccalite
[CNRS/ Aix-Marseille Université/ CINaM laborarory / TSN department
Campus de Luminy – Case 913
13288 MARSEILLE Cedex 09
web site: http://www.attaccalite.com
Freely download scientific books from: http://www.freescience.info

Post Reply