bse calculation stop in the step of BSE Diagonalization

[ljzhou86]

Dear sir
I try to do a GW+bse's calculation monolayer MoS2 by "yambo -b -o b -k sex -y d -V qp ". It always run to the step of [06] BSE solver and stop here for hours without the ability to continue, showing the information as follows
...
<25m-19s> P001: BSK |################### | [095%] 10m-47s(E) 11m-21s(X)
<26m-14s> P001: BSK |####################| [100%] 11m-42s(E) 11m-42s(X)
<26m-14s> [06] BSE solver(s)
<26m-14s> [06.01] Diagonalization solver
<26m-14s> [M 2.083 Gb] Alloc BS_mat ( 2.027)


Note that: I also try to run yambo -M -S with 448 cpus paralleled on 8 nodes to allocate memory and use the fragments approach, and it also makes nonsense. The version is yambo-3.4.1-rev51. One thing has to be mentioned that I can quickly get my final results by using (h)aydock BSE solver. Why is it not available for (d)iagonalization to get the BSE results so that I can obtain excitons' wave. Thanks in advance.

[Daniele Varsano]

Dera Zhou,
from what I can see from the report file, you are handling a very big matrix:

Code: Select all

 |Dimension               : 11664

in the input file you also included the coupling (which doubles the matrix size and makes it non hermitian).
My impression is that it just takes a lot of time to diagonalize it. Full diagonalization is very time consuming for
big matrixes. Please also note that the diagonalization task
it is not very well parallelized in this version of the code, so adding cpu's does not help. Also the memory issue does not seem
a problem as you have around 2Gb allocated.
If you have chance to enter in the nodes, you can check if at least one cpu is running 100% to verify it.
Or you have some error message? does not look the case looking at your standard output.
My suggestion if you do not want to wait a lot of time, and you need it for the exciton wave function only is to reduce
the number of bands (hence the dimension of the matrix), and I guess that you still can have qualitatively right exciton wave functions.
Hope it helps,
Daniele

[ljzhou86]

from what I can see from the report file, you are handling a very big matrix:

Code: Select all

 |Dimension               : 11664

in the input file you also included the coupling (which doubles the matrix size and makes it non hermitian).

My suggestion if you do not want to wait a lot of time, and you need it for the exciton wave function only is to reduce
the number of bands (hence the dimension of the matrix), and I guess that you still can have qualitatively right exciton wave functions.
Hope it helps,

The studied system is very small and "bands" I set is only 100, why the yambo has the dimension be so high and no ability to handle it? The parameters set in some references (such as PHYSICAL REVIEW B 88, 045412 (2013), in which the bands are beyond 200 and can get the exciton wave normally as shown in Fig3) are usually larger than I set. Thus, I guess it is not normal, how about your opinion?
To reduce the No. of bands, do you mean the "BSEBands"? I have limited it into a small range of "4 | 15 | "

[Daniele Varsano]

Yes,
it has nothing to do with the Xbands.
I refer to BSEbands.
The dimension of the excitonic matrix is given by: Nc x Nv x Nkpoints x Nsymm
So, even if the Bsebands it is not a big number, if your Kpoint number is large the matrix dimension is large as it is the case shown in the report: 11664

Best,
Daniele