Page 1 of 1

Convergence problems on HCP and BCC primitive unit cells when number of cores > 1

Posted: Tue Dec 20, 2011 9:31 pm
by rbgetman
Hello,

I have recently started using VASP again after ~ a 2.5 year hiatus. As a first task, I am converging some bulk metal structures, e.g., for Ru, which is HCP. I am using VASP 5.2.

I set up a POSCAR according to experimental crystal data (Ru-Ru distances ~ 2.7 Angstroms). The primitive unit cell has lattice vectors of (0.5, -sqrt(3)a/2, 0); (0.5, sqrt(3)a/2, 0); (0,0,c), where a ~ 2.7, and c ~ 1.6a. It has 2 Ru atoms at fractional coordinates of (1/3,2/3,1/4) & (2/3,1/3,3/4). I used a gamma centered KPOINT mesh with > 160 irreducible k-points, and the following flags (amongst others, which I will elaborate on if necessary) in the INCAR:

IALGO = 48 -or- default value
ALGO = Normal -or- fast
LREAL = .FALSE.
IBRION = 1
NFREE = 10
POTIM = 0.1
NPAR = 1
LPLANE = .TRUE.
NSIM = 10 -or- 8

Oddly, if I run this simulation on a single core, the simulation finishes without errors. If I run across all 24 cores on the same node, I get errors, e.g., such as the following:

WARNING: Sub-Space-Matrix is not hermitian in DAV

I am running this simulation on Linux nodes with quad hexacore Intel Xeon processors. My vasp executable uses the MIT FFTW and MKL BLAS and LAPACK libraries, but I have tried executables that use the default LAPACK libraries and gotten the same errors. (I can share more details about my compilation if it will be useful.)

Interestingly, I do not get the same errors when I run simulations on FCC primitive unit cells.

I found several hits on this forum containing the same error message that I'm getting, but none of them seems to be solved by running on 1 core vs. multiple. I did find one other result (http://cst-www.nrl.navy.mil/users/erwin ... s/#notherm) that may be related.

Any advice on why I am getting this error?

Thank you very much.

Convergence problems on HCP and BCC primitive unit cells when number of cores > 1

Posted: Wed Dec 12, 2012 1:05 am
by dEnvar
Have you found the solution?

Convergence problems on HCP and BCC primitive unit cells when number of cores > 1

Posted: Thu May 02, 2013 4:07 pm
by juhL
I'm also having this with any kind of IBRON, IALGO, and ISYM settings. Besides that it isn't necessary to calculate small systems on more than one node, does anyone have an educated explanation for this?