warning message during parabands

[Dani]

Dear BerkeleyGW community,

I am trying to use ParaBands to generate wavefunctions with larger number of empty bands. During execution, I am getting the following warning:

k-point pools are broken. Overwriting `number_pools` to 1

which I understand in the following way: I cannot parallelize over k-points. Thus, I am essentially running in one node one k-point after the other.

Has anyone seen this before? Can this happen due to bad compilation? Any hint on how to solve it would be very welcome.

Best,

Daniel

[Felipe H. da Jornada]

Dear Dani,

This message simply says that you cannot parallelize over multiple k-points. You can still use several MPI processors/nodes per k-point, you just can't solve more than one k-point pool at the same time. So, you must have `number_pools 1` in your input file.

K-point pools used to work and would provide a separate layer of parallelization. However, we recently noticed that the code would freeze, and hence we temporarily disabled the flag. This is not a critical feature, but if there is enough interest, maybe we can fix this for BGW 3.1...

All the best,

Felipe

--
You received this message because you are subscribed to the Google Groups "BerkeleyGW Help" group.
To unsubscribe from this group and stop receiving emails from it, send an email to help+uns...@berkeleygw.org.
To view this discussion on the web visit https://groups.google.com/a/berkeleygw.org/d/msgid/help/84ab1b5c-fff3-473a-b4be-a33eb70ed627n%40berkeleygw.org.

--

Felipe H. da Jornada

Assistant Professor

Department of Materials Science and Engineering

129 Durand Building

Stanford, CA 94305

Stanford University

https://jornada.stanford.edu/

[Chin En Hsu]

Dear Felipe,

I have noticed this discuss since I ran into same problem while using parabands.

I wonder is it safe using (or can we use) parabands in Berkeley 2.0 to solve WFN.h5 with SOC and use epsilon.x in Berkeley 3.0 ?

Regards,

Andy Hsu

 

Felipe H. da Jornada 在 2022年1月25日 星期二上午3:22:42 [UTC+8] 的信中寫道：

[Chin En Hsu]

Dear all,

Sorry I have tested it quickly, it cannot work.

ERROR: Cannot use spinor WFN file (nspin = 4).

So at least in 3.0 we can safely use parabands diagonalizes the DFT Hamiltonian with SOC, right?

Regards,

Andy Hsu

Chin En Hsu 在 2022年2月8日 星期二上午12:56:47 [UTC+8] 的信中寫道：

[Felipe Jornada]

Dear Andy,

 

You are right, the previous version does not support spinors. If you have access to the source code, you can try to comment the lines in MeanField/ParaBands/inread.f90 after the comment: “! FHJ: FIXME! Pools are not working!”

 

Use at your own peril, though. There is a high chance that the code will just freeze.

 

Out of curiosity, how many k-points do you have? It may be easier for the next minor release of BGW to include a restart feature in ParaBands, to continue from the last k-point that stopped, rather than fixing k-point parallelization.

 

Best,

Felipe

 

–

Felipe H. da Jornada

Assistant Professor

Department of Materials Science and Engineering

 

131 Durand Building

Stanford, CA 94305

Stanford University

https://jornada.stanford.edu/

 

Pronouns: he/him/his

[Chin En Hsu]

Dear Felipe,

Thanks for the information.

Using 12x12x1 coarse grid and symmetry I have 30 k-points in WFN file, and I want to produce as least 5000 bands for further converge test (it's 2D TMDs)

the number of G-vectors is about 80000 which was estimated from PW calculation info. And I think the restart feature will be handy !

Another question I wonder, In the parabands document it has tested the calculation time with respect to N (matrix of H size is NxN). I wonder how is the calculation time increase with different number of bands to generate, or should it not effect the CPU time since it  diagonalizes the whole DFT Hamiltonian?

and also will the number of bands in the input WFN (generate from pw2bgw.x) effect the parabands results?

Regards,

Andy Hsu 

Felipe Jornada 在 2022年2月8日 星期二上午3:53:25 [UTC+8] 的信中寫道：

[Felipe Jornada]

Hi Andy,

 

In the documentation, N is the number of rows (or columns) of the matrix, not the number of entries.

 

The time that parabands takes doesn’t depend much on the number of bands that you ask for, except for I/O, of course. You may get a factor of 2x speedup on a solver like ELPA, at most, which is particularly efficient if you ask for a smaller fraction of your eigenspace.

 

The number of bands in the input calculation doesn’t affect parabands at all. The code only reads the energies up to the Fermi energy + a few empty states for consistency checks. We have a beta version that no longer needs an input WFN.

 

Finally, you may benefit from using the ELPA solver, if that’s available in your platform.

 

All the best,

[Chin En Hsu]

Dear Felipe,

Thanks for the suggestion, I have already include ELPA solver for parabands!

I think the parabands code is running just fine. That's hope that I can get the good converge results after that.

Best wishes,

Andy Hsu

Felipe Jornada 在 2022年2月8日 星期二上午10:42:47 [UTC+8] 的信中寫道：

[Chin En Hsu]

Dear Felipe,

I have mange to do the parabandns calculation using ELPA.

And I have test several system.

However, some of the system will occur the problem that "WARNING: Difference between input and output MF energies larger than 1e-2."

I can see that the difference is quite large by grep the output

$ grep Max. PA.out

Max. number of G-vectors per k-point: 65785

Max. number of occ states: 16

Max. difference between input and output mean-field energies: 1.251E+00 Ry

Max. difference between input and output mean-field energies: 1.253E+00 Ry

Max. difference between input and output mean-field energies: 1.259E+00 Ry

Max. difference between input and output mean-field energies: 1.270E+00 Ry  

I think this is unusual and seems like there are some problem with my input file (WFN..VXC...etc), I wonder for such problem, how can one check the quailty of the initial WFN or VXC...etc.

I thought it would be affect by the energy cutoff In pw.x since the G-vecs number is directly affect by the energy cutoff in plane wave. However, increasing the cutoff seems not helping

I have check my pw.x calculation, seems like everything is just fine. (without paraband, I can still able t generate the WFN for about 1000 bands)

How should I deal for such problem?

Regards,

Andy Hsu

Chin En Hsu 在 2022年2月8日 星期二下午7:39:21 [UTC+8] 的信中寫道：

[Felipe Jornada]

Dear Andy,

 

I’m sorry I missed your previous message. Did you figure out the issue? If not, can you provide more details? Are you running parabands with the real or complex version? And are you sure that the code is reading both VSC and VKB?

[Chin En Hsu]

Dear Felipe, 

I have solved the problem, seems like I link the wrong file in the parabands.inp (I use VXC replaced VSC file)

Everything is just working fine now, and I also mange to do the parabands.cplx.x calculation using ELPA + OMP +  MPI. 

This allows me to generate lots empty states for several systems. 

Again thanks for all the helping and suggestion!

Best wishes,

Andy Hsu

Felipe Jornada 在 2022年3月8日 星期二下午3:04:37 [UTC+8] 的信中寫道：

[Felipe Jornada]

Dear Andy,

 

That’s great, I’m happy that you sorted this out and that the code was helpful.