Convergence and disk space of BSE calculation

[Flex]

Hello,

I am working at the moment on BSE calculations of the optical properties of monolayer black phosphorus, based on pwscf calculations. I did the scf and nscf needed (with a grid of 20*20), converted with p2y without problems. That gave me about 600,000 Gvecs that I reduced to 10,000 when I did the initialization. I have 121 Q-points.

Attached (BSE.in/txt) is the input I obtained with the command : yambo -o b -k sex -y h. I reduced some parameters (BSENGexx, BSENGBlk, and block size) in order to do a first run, not precise but fast.

This is where I ran into the first problems. Even with these very reduced parameters, the calculation seems to need huge computing power and RAM. I had to run it on 16 nodes, 16 cores each and 16G of ram per core. If I used less, it would take an entire week. And this is only the first non precise run. Max job time on my cluster is about 1 week.

(BTW, I also did GW calculations earlier, and they needed much less resources.)

Is it normal ? Do I have to explicitely add parameters to optimize parallelization ? Are my parameters still too big ?

So, I launched the calculation with 16*16 and it ran quite well until about halfway, when I saw the save took about 600 GB of disk space. I had to abort it.
This is a problem for me as I work on a cluster I share with other people and I only have about 1 TB of disk space I can use. I already use about 300 GB for other pwscf and GW.

Is there a way to reduce the disk space used ? How much disk space does a full precision BSE calculation need ? (by full precision I mean good enough to have reliable results)

I can attach other logs and files if needed

Thanks in advance

[amolina]

Dear Flex,

looking at the input file I see you have set BSEBands = [1,30]. This is a huge number of bands for a Bethe-Salpeter calculation. You should set this number to the conduction and valence bands close to the gap, with importance for the calculation of the absorption. A good test is to set just one valence and one conduction band to make some convergence tests.

Cheers,
Alejandro.

[Flex]

Oh, so that was the problem.

I guess the spectrum can be already very precise with about 5-10 bands around the gap, then.

Thank you very much for the quick answer.

[Daniele Varsano]

Dear Flex,

.first please note that you are using a very old version of yambo. I strongly suggest to update to a newer version as you will have more flexibility int he parallelization strategy.

.Next, as suggested by Alejandro, in BSE most probably you do not need deep bands: start with few bands across Fermi and enlarge the windows till convergence. I can's see the report file, but there you can find the dimensione the BSE, which is Nc*Nv*NK, where Nk is the entire BZ so it can explode easily.

. in the 4.x of the code there is also a flag to avoid to write the BS matrix on disk.

Best,

Daniele