Page 1 of 1

insufficient virtual memory

Posted: Wed Jun 08, 2011 6:10 pm
by mantega
Hi All,

I'm trying to calculate the bands structure of a big system with 440 atoms (only Si and H), but the job crashes within few minutes with the following error code:

forrtl: severe (41): insufficient virtual memory
Image PC Routine Line Source
vasp 0000000000C9636E Unknown Unknown Unknown
vasp 0000000000C95300 Unknown Unknown Unknown
vasp 0000000000C411B2 Unknown Unknown Unknown
vasp 0000000000BFC7BF Unknown Unknown Unknown
vasp 0000000000C265DB Unknown Unknown Unknown
vasp 000000000052461E Unknown Unknown Unknown
vasp 0000000000446E4E Unknown Unknown Unknown
vasp 000000000042FD82 Unknown Unknown Unknown
libc.so.6 0000003C81C1D994 Unknown Unknown Unknown
vasp 000000000042FCA9 Unknown Unknown Unknown

I tried to run it in a parallel cluster, using 32 nodes with 8cpus/node with different configurations from one task per cpu till one task per node (memory saving). I always get the same error. I don't have any problem performing a structural relaxation for the same system.
This is the my INCAR file:

SYSTEM = Si-1H_b_SP

Startparameter for this Run:
ISTART = 0
NWRITE = 2
LPETIM = F write-flag & timer

Electronic Relaxation 1
PREC = High
ENMAX = 500 eV
NELM = 80
EDIFF = 5E-04 stopping-criterion for ELM
INIWAV = 1 random initial wavefunctions
MAXMIX = 40
NELMDL = -7

NELECT = 1522

Ionic Relaxation
EDIFFG = -0.02 stopping-criterion for IOM
NSW = 0 number of steps for IOM
NBLOCK = 1; KBLOCK = 100 inner block; outer block
IBRION = -1 ionic relax: 0-MD 1-quasi-New 2-CG
ISIF = 2 stress and relaxation
IWAVPR = 11 prediction: 0-non 1-charg 2-wave 3-comb
ISYM = 1 0-nonsym 1-usesym
LCORR = F Harris-correction to forces

# ISMEAR = -5
ISMEAR = 0
SIGMA = 0.05
LREAL = Auto
LMAXMIX = 4

ICHARG = 11

ISPIN=2
MAGMOM=1*1.0 360*0.0 80*0.0


Any guess?
Is there anything I can optimize in my INCAR file?

PS: I have the same problem for non spin-polarized, non electron doped systems.

Thanks in advance

Mauro

insufficient virtual memory

Posted: Thu Jun 09, 2011 3:05 pm
by admin
I suppose the major difference between the scf-relaxation run and the band structure run is the number of k-points. therefore, the first choice would be to reduce this parameter, eg by taking a coarser grid along the lines, or by splitting the calculation of the bands along the lines of the BZ (as ICHARG=11, you have the same potential for all of them). if there are any constant shifts in the KS-spectra for the runs, please adjust them at the high-symmetry points where the lines match.
you could of course reduce PREC and ENMAX alternatively, if you want to keep the k-points as they are.

insufficient virtual memory

Posted: Mon Jun 13, 2011 5:23 pm
by mantega
I tried with only 5 kpoints with the following KPOINTS file but it crashed again. (The SCF-relaxation runs properly with 5 kpoints).
I'm now trying reducing ENMAX.

k-points along high symmetry lines
5
Line-mode
rec
0.0 0.0 0.0 ! Gamma
0.5 0.0 0.0 ! J

insufficient virtual memory

Posted: Thu Jun 16, 2011 7:16 pm
by mantega
I managed to run it with 10 k-points per job. Still I can't understand why it crashes using the same parameters as the SCF-relaxation run (same number ok k-points, cutoff and prec).