Reduce memory demand in rf

Total energy, geometry optimization, DFT+U, spin....

Moderator: bguster

Locked
roginovicci
Posts: 75
Joined: Thu Dec 02, 2010 10:36 pm

Reduce memory demand in rf

Post by roginovicci » Wed Nov 21, 2018 11:03 am

Hi,
To perform optical calculations one needs to perform ddk response function calculations ( rfelfd > 0 ) using a big set of empty bands ( nband > 300), which leads to huge amount of memory demand. For crystal with tetragonal symmetry and 10 atoms in unit cell the required memory (RAM) is more than 250Gb. The amount of memory is independed if I use 12 or 6 or 1 processors for calculations. Is there any possibility to reduce memory demand (except using lowes number of unocc. bands - nband)?

Many thx in advance.

ebousquet
Posts: 469
Joined: Tue Apr 19, 2011 11:13 am
Location: University of Liege, Belgium

Re: Reduce memory demand in rf

Post by ebousquet » Tue Nov 27, 2018 9:53 pm

Dear roginovicci,
How much ecut and k-point grid did you use, because the number of bands and atoms is indeed not super large?
Best wishes,
Eric

roginovicci
Posts: 75
Joined: Thu Dec 02, 2010 10:36 pm

Re: Reduce memory demand in rf

Post by roginovicci » Thu Nov 29, 2018 6:52 am

Dear Eric,
The nonlinear optical calculations by "optic" utility (SHG) require rather hi k-mesh and ecut to converge (which is also explained in tutorial https://docs.abinit.org/tutorial/optic/index.html). I've found the convergence by ecut is 30Ha and k-mesh is 12x12x8 for my system. I know that the optical properties calculation (in different codes) demand a lot of memory, so I'd like to know how to reduce it in case response function ddk calculations (rfelfd=2), maybe this would be fair to use kptopt=2 instead of 3 since calculations of derivative in Gamma point (qpt = 0 0 0)?

Thank you for interest to my problem,
Eugene

ebousquet
Posts: 469
Joined: Tue Apr 19, 2011 11:13 am
Location: University of Liege, Belgium

Re: Reduce memory demand in rf

Post by ebousquet » Wed Dec 05, 2018 10:11 am

Dear Eugene,
You can indeed put kptopt=2 for the ddk and E-field at Gamma point, you will save half of the k-point number and thus the memory. Beside that, nothing can be done at the moment, there is a project running to find a way to reduce the memory consumption of DFPT but it is under way...
Best wishes,
Eric

Locked