Probleme with Npar in vasp5.2.11

[jsfilhol]

I am getting different energies for single point calculation using Npar=1/2/..32 and Npar=64 when using 64 cores. From my previous run,all the values but Npar=number of core are correct. I have tried various setting changing the parallization options but NPAR=nb of core is always giving uncoherent results preventing me to run lepsilon calculations. Has someone any idea to solve this problem ?

Thanks
JSF

PS my INCAR is

PREC = Normal
ALGO = FAST
EDIFF = 1e-6
# LEPSILON =.TRUE.
ISPIN=1
NELM = 60
ENCUT = 550
LREAL= .FALSE.
Ionic relaxation
EDIFFG = 1e-4
NSW = 0
IBRION = 2
LVTOT=.TRUE.
LPLANE = .TRUE.
NPAR = 64
LSCALU = .FALSE.
NSIM = 4

[forsdan]

Please first check if the number of bands (grep for NBANDS in the OUTCAR file) is the same. The number of bands will be modified so that NBANDS divided by NPAR is an integer. If NBANDS differ between the runs so can the energy to a small extent.

Cheers,
/Dan

[jsfilhol]

Thanks for this answer, but NBANDS does not do the trick even setting it at multiple of my 32 core
Running with NPAR=nb of cores is giving positive non physical energies whereas the other NPAR gives physical negative energies. Even staring with a converge WAVECAR for NPAR=nb of core/2 giving the right energy and restaring to a NPAR=nb of core still gives positive unphysical energy.
It looks like a parallelism problem or a library trick
we are using vasp 5.2.11 compiled with scalapack, using ifort/mkl 11.1 +intel mpi

Thanks

JSF

[jsfilhol]

Finally we have solved the problem:
-the bug is surprisingly coming from the mkl 11.1 --> upgrading to mkl 12.0 seems to restore the correct energy behavior.
There is then some incompatibility between mkl 11.1 and vasp 5.2.11

JSF