HSE speed

[koehenri]

Hi,

I'm performing convergence calculations of a CsCl lattice (actual species are Fe-O) with 2 atoms in the unit cell. For each pair of (number of k points, encut) I'm doing first a PBE geometry relaxation and then an energy evaluation only using HSE. I'm running on 12 threads (6 physical cores, 2 threads/core, Xeon X5650). The INCAR file for the HSE step is:

Code: Select all

LASPH?=?.TRUE.
LMIXTAU?=?.TRUE.

PREC?=?Accurate??????????????#??precision?normal

LREAL?=?.FALSE.????????????#??real?space?projection?yes?/?no

ISMEAR?=?-5?????????????????#??method?to?determine?partial?occupancies
SIGMA?=?0.1

ISIF?=?2
NSW??=?0

NELMIN?=?10
EDIFF??=?1e-5

ISPIN?=?2

NPAR?=?12
ENCUT?=?300?????????????#??cutoff?used?throughout?all?calculations
IBRION?=??2
MAGMOM?=?1?1

ISTART?=?1
LHFCALC?=?.TRUE.??;?HFSCREEN?=?0.2
ALGO?=?Damped?????;?TIME?=?0.4
PRECFOCK??=?Fast??!?used?PRECFOCK?=?Normal?for?high?quality?calculations

The "problem" is that a single run is very slow, taking several days and not even finishing during that time. I'm afraid I might be using some stupid settings in the INCAR file? Is there any way --- besides using more cores --- to speed up the calculations?

-Krister

[kambiz]

Hi,

The HSE calculations are very slow. You are doing the right thing. Just consider that using PREC=Accurate will increase your computational time. Besides, you should also see how many kpoints you are using.

Cheers

[WolverBean]

To amplify what kambiz said:

I recently conducted both a PBE and an HSE calculation on MoO2. Using the same number of kpoints, bands, precision, etc (prec=accurate for both, and precfock=accurate for the HSE calculation), I found that the HSE calculation took 3251 times longer than the corresponding PBE calculation. So yeah... HSE calculations are VERY slow. Your mileage may vary, and using NKRED (for non-metallic systems) can help. But short of using more cores, there's not much you can do.

If it fits what you're doing, you can try using the new HSE12s functional proposed by Moussa et al (http://jcp.aip.org/resource/1/jcpsa6/v1 ... p204117_s1), which is claimed to be as accurate as HSE but has a shorter screening length, and therefore converges more rapidly with respect to kpoints (i.e. requires fewer k-points for the same accuracy, which would then be faster overall). To implement HSE12s in VASP, one would set
ISYM=3
LHFCALC=.TRUE.
HFSCREEN=0.408
AEXX=0.425
AGGAX=0.575
AGGAC=1.00
ALDAC=1.00

Good luck!

[koehenri]

Thanks! I'll check out the paper and try those settings. I haven't had any success in using NKRED=2 as suggested also in the manual. I get an error in VASP stdout about k-pointsm (some internal error). I'm using odd-number k-point setups.

[HJ.Kim]

Dear koehenri,

I think you should set NKRED = 3 or other number so that NKx/NKRED and NKy/NKRED and NKz/NKRED to be integer.