Page 1 of 1

ZGHEV errors vs number of cores

Posted: Fri Jun 20, 2014 9:57 pm
by t.ossowski
We compiled parallel version of VASP 5 using Intel fortran compiler (and mkls). We have Intel processors 4x10 cores. For some systems we have problems with messages like "matrix not hermitian". It is surprising that for example there are no problems with calculations of clean surface, but when we add adsorbate atoms, problems appear (message: "matrix not hermitian" and calculations carshed).
Another example - clean surface and one-side relaxations with dipol correction. Even for pure surface calculations crashed.

But ..... when we run calculations on 8 cores problems disappeared! We tested calculations using different numbers of cores and calculations crashed when are running on 5,7,9,10,11 cores. For 1,2,3,4,6,8,12,16 cores there are no problems.
Maybe someone knows why???

ZGHEV errors vs number of cores

Posted: Mon Aug 18, 2014 7:07 am
by t.ossowski
Does anyone know the solution of our problem? We have still problems with calculations. Last time with Fe(110) surface and (1x2) primitive cell. When we run it on 16 cores we have message "Matrix not hermitian ..." and caclulations crash at the beginning of the first step. The same job on 8 cores run without problems!

ZGHEV errors vs number of cores

Posted: Tue Aug 19, 2014 7:12 am
by admin
Upon changing the number of cores vasp changes the number of bands (NBANDS must be divisable by NPAR). This can lead to differences in the course of the calculation. Try to fix the number of bands.
[ Edited Tue Aug 19 2014, 09:26AM ]

Re: ZGHEV errors vs number of cores

Posted: Fri Sep 05, 2014 6:32 pm
by cchang
admin: When you say number of cores, do you mean the number of MPI ranks generally, or the divisor of the band number after accounting for NPAR/NCORE? So, if I have 10 MPI ranks and NPAR=2, NCORE=5. If I understand http://cms.mpi.univie.ac.at/vasp/vasp/P ... R_tag.html, that would divide each band 5 ways, and the number of bands 2 ways. Would it then be good enough to have NBANDS divisible by 2, or should it be divisible by 10?

Thanks

Re: ZGHEV errors vs number of cores

Posted: Tue Sep 09, 2014 11:48 am
by admin
NBANDS depends exclusively on your system (atoms, electrons). Before choosing parameters for
parallel calculations one should know NBANDS of the system. NBANDS must be divisable by NPAR.
When NBANDS = 28 and NPAR=4 nothing is changed by VASP.
At NPAR=6 VASP adjusts NBANDS to 30.