BSE: Screened interaction matrix calculation

[Evgeniia]

Dear Daniele,

please find my files attached. There are inputs and outputs again in txt formats. I tried to ran my bse_br.in again, redirecting it to the log - file, as you suggested, it is currently in process. However for some reason log is still empty. Will it contain information after the calculation stops? I am not sure how it should work. Sorry about these stupid things..

Best Regards,
Evgeniia

[Daniele Varsano]

Dear Eugenvia,

Looking at your input there is this namelist wich is not closed, there is a missin % symbol.
This can make confusion:

Code: Select all

%KfnQP_E
0.000000|1.000000|1.000000|             #       (EXTQP)(BSK)(BSS)       E parameters (c/v) eV|adim|adim

Also note that with your intput parameters you have a very large BSE matrix:

Code: Select all

 |Dimension               :  66096

It is not the best example to test.


Best,
Daniele

[Evgeniia]

Dear Daniele,

the missing % however doesn't stop the calculation for 10 K-points in scf( But I will try tomorrow morning. I will compare everything with the tutorial.

Can this large BSE matrix cause the netCDF problem?
Will it be possible to ask you again if netCDF remains?

Best Regards,
Evgeniia

[Daniele Varsano]

Dear Evgeniia,
the missing % make the input file incorrect and it is possible that it is not read correctly by the parser.
The large BSE matrix it is not the problem, but it makes the calculation very large, and time and memory consuming and depending on your machine
you can run out-of-memory. In general, you do not need to include all the occupied bands but only a set around the Fermi energy checking the convergence.
Finally please note that the report and the input you attached does not come from the same calculation (BSE bands 1-35 in the report, 14-23 in the input).
I think it is a good idea to do the step by step tutorials in order to get practice with the code.
Best,
Daniele

[Evgeniia]

Dear Daniele,

That was actually my second question. They DO come from the same calculation, I am absolutely sure, that I do everything in the same folder, so BSE report is correct. But for some reason BSE takes ALL bands, instead of what I ask in the input. Previously I thought that this works similarly with the calculation of screened interaction matrix, where it changes the input to the numbers it wants to use. However you say it is not. Today we double checked, that I ask to calculate one range of bands, and it takes another. This is true also for calculation the smaller number of K-points, where BSE worked well.
Is it a yambo bug?
I repeated LiF and NaCl, everything is fine. I may try another ionic crystal, but I think it is not necessary...
Best Regards,
Evgeniia

[Daniele Varsano]

Dear Evgeenia,
you had a broken input and this is the possible source of error, there is a namelist which is not closed and this can cause unexpected problem.
My suggestion is to start from scratch: delete all the ndb.* file from your SAVE directory and generate new inputs.

Previously I thought that this works similarly with the calculation of screened interaction matrix, where it changes the input to the numbers it wants to use.

No, yambo does not change the number of bands, yambo takes the bands you indicate in input also for the screened interaction matrix, what does change is the dimension of the G vector matrix in order to close G vector shell, taking the number closest to the one you indicate in input.
Best,

Daniele

[Evgeniia]

Dear Daniele,
I entered the missing % to the input, and then I decided firstly to check the divergence for the scf wave-vectors net before writing you the result. Again the same netcdf error appears for the 12 12 12 and 15 15 15 net.
My experience says that that may mean, that all previous calculations (10 10 10 and lower) may be incorrect then. I double checked everything with the tutorial inputs. It seems for me, that there are no missing signs now. The error is the same - netCDF:one or more variable sizes violate format constrains. I guess that now the problem might be in the choice of parameters...
Could you please help me again?

Best Regards,
Evgeniia

[Daniele Varsano]

Dear Evgeniia,
some comments after looking at your reports:
1) I can see you are using a quite old version of Yambo (GPL Version 4.0.4 Revision 107). Moreover, Yambo is compiled in Serial and this poses a limitation on the dimension of the systems you can treat in terms also of pushing convergence parameters. I strongly suggest you update to a newer version of the code as many bugs have been solved since then.

2) My impression is that the problem you face with this k point grid is most probably due to memory issue:
The BSE matrix has a huge dimension (50625):

Code: Select all

BSK|Identifier              : 1926
       |Dimension               :  50625
       |Bands                   : 16  23

This means that you will need to allocate the matrix:
50625* 50625*16/(1024^3)=38Gb

You can check that the machine you are running has this resources. You can try to reduce the number of BSEBands and hence the memory needed to see if this solves the issue,
but of course, then you could be out of convergence with respect the bands' variables. Note that the memory needed for the BSE increase quadratically with respect the number of k points. In any case note that beside the K point grid convergence, you are running with very few BSENGBlk and BSENGexx. Once assessed these parameters to converged values I doubt you can successfully run a calculation using a serial compilation of the code.
Note that in the most recent release of the code also the parallelization strategy has been optimized allowing memory distribution among cpu's.


Best,
Daniele