Page 1 of 1

BUG in MPI version of prtdos=3 in spin-polarized PAW

Posted: Wed Jul 30, 2014 8:08 pm
by Apophis
Hello,

I have been calculating PDOS of Ni3C using PAWs in abinit. The carbide is known to be paramagnetic and abinit also gives this result when calculating the total DOS. However, when I perform a spin-polarized calculation (abinit correctly reports no magnetization) with prtdos=3 the spin channels differ in magnitude (see page 3 of attachment). This does not look like a problem with normalization, since they do not differ by a constant factor over the whole energy range, there is even a cross-over point (page 4, around -10eV). With prtdensph=1 abinit does not report any magnetization in the atom spheres. Also, all eigenvalues are identical for both channels.

For comparison a colleague ran a calculation using vasp. For the total DOS the abinit and vasp agree well (page 1). But vasp does also not show any difference between the spin channels for PDOS (page 2 vs. 3).

Running with pawprtdos=1 shows that the discrepancy orginates solely in the PW component of the DOS, all-electron and pseudo contribution show no difference between the spin channels.

My input file looks like this:

Code: Select all

ntypat 2
znucl  28 6
natom 8
typat 2 2 6*1
xred
 -3.94426681151606E-18 -2.32635776975131E-17  7.71373150378205E-19
 -1.31961253500590E-17  5.00000000000000E-01  5.00000000000000E-01
  1.70593337472857E-01  2.50000000000000E-01  2.50000000000000E-01
  5.00000000000000E-01  9.20593337472857E-01  2.50000000000000E-01
  1.70593337472857E-01  4.20593337472857E-01  7.50000000000000E-01
  8.29406662527143E-01  5.79406662527143E-01  2.50000000000000E-01
  5.00000000000000E-01  7.94066625271432E-02  7.50000000000000E-01
  8.29406662527143E-01  7.50000000000000E-01  7.50000000000000E-01

acell 8.68823391483284E+00  8.68823391473050E+00  9.60347684124753E+00
rprim
  1.0 -4.38964554255e-29 -8.43617061338e-57
  -0.500000000002 0.866025403783 1.0184651346e-27
  0.452348355634 0.261163444886 0.852745343119

toldfe   1.0d-9

ecut   36.0
pawecutdg 52
ngkpt   3*12

nshiftk   1
shiftk   0.0 0.0 0.

nband   44
tsmear 1e-3
occopt 3 #fermi

prtdos 3

prtden 1

nsppol 2


I saw this using several versions of abinit, including 7.8.1

[Edit] This seems to be a bug in the MPI parallel code, the results are fine when abinit is run on a single CPU core.

Thank you.

Re: BUG in MPI version of prtdos=3 in spin-polarized PAW

Posted: Fri Feb 20, 2015 5:52 pm
by bmatt
Yes, I came across the same problem using MPI. If you set the mpi core value to 1 on your submission script, the pdos turns out fine. It obviously takes longer, but it is the only way to do it until they fix the issue
If I remember, it takes about 2-4 hours to do 1 atom. If you run multiple files at the same time, you can do the entire unit cell, just converge the WFK file before you run in serial and just have it as file_io_WFK