CALC_WANNIER_TRAFO_PROJ error upon wannierization

[armin_eghdami]

Hello!
I've been trying for weeks to get a wannierization to run using VASP (I previously had only done it in QE), but I keep encountering this error message:

Code: Select all

CALC_WANNIER_TRAFO_PROJ: xGESDD returned with error code 3 

I can successfully run a wannierization on simple silicone structures so I think I got the workflow correct (first do an scf calculation, then add the LWANNIER90 = .T. label, WANNIER90_WIN tags and then rerun using the scf-WAVECAR). I also already really pushed down my ECUT as well as stopped using HSE, went down to a half sized version of my original structure and went down to only only using the gamma k-point, so at this point I don't think I can have any memory issues anymore. I also erased any wannier inputs that might lead to errors and only use the essential ones.

Googling this error message also gives me no results.

My goal is to wannierize the four bands around the fermi level (2 above, 2 below), which stem from my antimony atoms.

Any help is greatly appreciated! Thank you!

wannier_problem.zip

[fabien_tran1]

Hi,

I got help from a colleague (H. Miranda). The problem should be due to a problem in VASP during a SVD decomposition of a matrix when bands are excluded. This problem can be avoided by adding LWANNIER90_RUN=.TRUE. in INCAR so that Wannier is called in library mode. Please, try to see if it works.

[armin_eghdami]

Thank you for the quick reply! I tried with that setting and I do not get the error anymore, but I still get the same

forrtl: severe (174): SIGSEGV, segmentation fault occurred

error. I saw online that this can sometimes be due to a lack of memory, but I'm running this on A LOT of nodes and the scf steps all finish within a few seconds, so I'm not sure if that's really the cause of the issue.

[fabien_tran1]

Hi,

It seems that the crash is due to a bug that was fixed in VASP.6.3.2 (you are using VASP.6.3.0). The solution is either to upgrade VASP (which is recommended) or to fix the bug in your VASP.6.3.0 version.

If you choose to fix the bug, then you need to replace (at lines 1654 and 1655 in mlwf.F)

Code: Select all

     CALLMPI( M_sum_z(MLWF%COMM,MLWF%U_matrix_opt,MLWF%NB_TOT*MLWF%NUM_WANN*MLWF%NKPTS*MLWF%ISPIN) )
     CALLMPI( M_bcast_l(MLWF%COMM,MLWF%lwindow,MLWF%NB_TOT*MLWF%NKPTS*MLWF%ISPIN) )

by

Code: Select all

    CALLMPI( M_sum_z(MLWF%COMM,MLWF%U_matrix_opt,(MLWF%NB_TOT-MLWF%NEXCLB)*MLWF%NUM_WANN*MLWF%NKPTS*MLWF%ISPIN) )
    CALLMPI( M_bcast_l(MLWF%COMM,MLWF%lwindow,(MLWF%NB_TOT-MLWF%NEXCLB)*MLWF%NKPTS*MLWF%ISPIN) )

and recompile the code. Please, report us if this is still not working.

Not related to the above bug, note that:
-It is still recommended to have "NBANDS = 192" in INCAR.
-In case of Wannier90 3.x, you should compile a serial version by removing COMMS=mpi in the make.inc of Wannier90 (see wiki/index.php/Makefile.include).

[armin_eghdami]

Thank you for the help! With the newer VASP version the vasp_gam version is working now. I then tried to apply it to my actual system which uses a 2x2x1 kgrid, but the wannier90.out seems to get stuck right before the wannierization starts, but the run doesn't fail and I also get no error code.
I then cut everything down to my test version again which ran fine with the 1x1x1 kgrid and the vasp_gam executive. Then I tried doing a 2x2x2 kgrid and the same thing happened, with the run getting stuck right before the wannierisation. With a 1x1x1 kgrid using the vasp_std instead of the vasp_gam exe, it also got stuck.
So the solution you provided seems to work fine for the vasp_gam exe, but not for the vasp_std exe. I attached the outputs of all 3 cases (1x1x1 kgrid with vasp_gam; 1x1x1 with vasp_std; 2x2x2 with vasp_std).

Maybe you have an idea of what else I could do to get it to run beyond a gamma-point-only calculation.
Thank you!

[fabien_tran1]

I can not reproduce your problem with the installation of VASP (v6.3.2) and Wannier90 (v3.1.0) that I am using. The three calculations finished properly.

Could you:
-Specify which compiler was used to compile VASP and Wannier90.
-Which types of machines/processors you are using.
-Run with different number of cores. In particular, run 1x1x1-std without parallelism (it should take less than two minutes).

Besides, as a general comment, for such small systems it is not necessary to use so many cores.

[armin_eghdami]

Thank you, by reducing the cores, it worked!