NETCDF error

[Daniele Varsano]

Dear Martin,

Is there any remedy to circumvent negative eigenvalues in BS matrix, except switching back to TD approximation ?

What do you mean by circumvent? When diagonalizing full BSE matrix (not TD) resonant part valence-->conduction transitions and antiresonant conduction-> valence are considered and mixed. It is then correct to have positive and negative eigenvalues in the same number. For absorption spectra only positive eigenvalues are meaningful.

Is the truncation of empty bands in Yambo still experimental ?

The truncation of empty bands in the evaluation of the Self Energy in the Bruneval-Gonze (GTermKind= "BG") should be there fully working.
The truncation for the response function is still experimental.

About the diagonalization should be not slower in 4.1, I do not have idea on why this is happening.
Best,

Daniele

[martinspenke]

Dear Daniele,

I fully understand your point, however i obtain some thing like this when running BSE with coupling using for instance Haydock :

Code: Select all

 <06m-14s> P0001: [Haydock] Iteration 6 Accuracy : 5.495872|0.1000E-4
P0001: [ERROR] STOP signal received while in :[09.01] Haydock solver
P0001: [ERROR] Bf=NaN likely because some eigenvalue of BSE is negative.

Maybe maybe this is the reason why diagonalization is so slow. I actually could n't diagonalize a FULL BSE 4000x4000 matrix after 12 hours, which is very strange.
Without coupling i do not have this problem, however i like to study possible coupling effects.
I have to add, i am studying a difficult 2d-system with SOC and with coulomb truncation using box z and rim.

What is the keyword of truncation of empty bands for the response ?

Is the new yambo version (4.1) mainly a bug fix or there are further new features, like vertices or self consistency, etc ?

many thanks and best wishes,
Martin

[Daniele Varsano]

Dear Martin,

Probably this is a problem of the Haydock algorithm when coupling is included. May be some other can give you details.
A BSE 4k x 4k is not that big, this sound strange, anyway when coupling is considered the matrix is 2 times bigger with respect the resonant only.

The keyword is XTermKind, but if I'm not wrong, this is working in the 4.1 only.
The 4.1 essentially enhance the parallelization and overall performance of the code.

Best,

Daniele

[Davide Sangalli]

Dear Martin

I would guess the reason why the diagonalization is so slow maybe indeed related to the message you get in the Haydock.
Indeed the diagonalization of the BSE matrix is always done with a lapack subroutine, both in 3.4 and in 4.0.
(We have scalapack in 4.1, but it is also experimental.)

Thus the only reason why 4.0 maybe slower is that, for some reason, the matrix is different.
Are you using the same input parameters in between 3.4 and 4.0 ?

Best,
D.

[martinspenke]

Dear Davide,

Yes, i used the same parameters for both.

Now i like to try out the version 4.1 but i can not extract the tar ball from the link you gave me.
http://www.yambo-code.org/testing-robot ... 109.tar.gz

I obtain this error :

Code: Select all

gzip: stdin: not in gzip format
tar: Child returned status 1
tar: Error is not recoverable: exiting now

Bests
Martin

[Davide Sangalli]

Ciao,
I just tried "tar -xzf yambo-4.1.0-rev.109.tar.gz" and it worked in my case.
Maybe there was a problem with the download ?

Best,
D.

[martinspenke]

Ciao Davide,

finally managed to open the tar ball.

Many thanks and best wishes
Martin

[martinspenke]

Dear Davide,

I obtain this error when compiling the newest yambo version :

Code: Select all

/home/martin/yambo-4.1.0-rev.109/lib/liblinear_algebra.a(SERIAL_lin_system.o): In function `serial_lin_system_':
SERIAL_lin_system.f90:(.text+0x168): undefined reference to `cgesv_'

I never had such a linking issue in the previous yambo versions.

Bests
Martin

[Davide Sangalli]

Yes, it's a problem we know.
It will be fixed in 4.1.1 which will be the official svn release

Meanwhile, to avoid that, you have to link some external blas and lapack library.

D.