SCF, MD run-time verses atomic species.

Simiam Ghan

unread,

Jul 1, 2016, 7:38:47 AM7/1/16

to cp2k

Dear all,

I am running NVT MD with Quickstep on a box of water with 64 molecules. If I replace a water molecule with a KCl ion pair, i observe that the MD and SCF step times more than double on my setup. (MD from ~20 sec to ~46 sec). SCF iterations are converging (except the very first MD step) in around 30 steps in both cases but each SCF now takes over twice as long as before. Is there an explanation of why such a 'small' change in system could double the run time? Is KCl really so heavy to calculate compared to H2O? The number of electrons in the system increases by 8, from 512 to 520. Below my KCl input file.

Greetings,

Simiam

&GLOBAL

PROJECT H2O_KCl

RUN_TYPE MD

PRINT_LEVEL MEDIUM

&END GLOBAL

&FORCE_EVAL

METHOD Quickstep ! GPW method.

&SUBSYS ! A subsystem: coordinates, topology, molecules and cell.

&CELL ! Supercell setup.

ABC [angstrom] 12.414 12.414 12.414 ! Using 64 H2O molecules, we thus get a density of 1g/cm^3.

PERIODIC XYZ ! Use PBC in all dimensions.

&END CELL

&COORD

UNIT angstrom

H -0.567712 -0.469646 -0.645913

H 0.626116 -0.687796 0.308193

O 0 0 0

(...)

K 2.1035 2.1035 4.2735

Cl 4.1035 4.1035 2.6035 ###

H 3.73881 3.10388 5.46104

H 3.65742 2.89924 6.989

O 3.1035 3.1035 6.207

(...)

&END COORD

&KIND O

BASIS_SET DZVP-MOLOPT-GTH-q6

POTENTIAL GTH-PBE-q6

&END KIND

&KIND H

BASIS_SET DZVP-MOLOPT-GTH-q1

POTENTIAL GTH-PBE-q1

&END KIND

&KIND K

BASIS_SET DZVP-MOLOPT-SR-GTH-q9

POTENTIAL GTH-PBE-q9

&END KIND

&KIND Cl

BASIS_SET DZVP-MOLOPT-GTH-q7

POTENTIAL GTH-PBE-q7

&END KIND

&END SUBSYS

&DFT

BASIS_SET_FILE_NAME BASIS_MOLOPT

POTENTIAL_FILE_NAME GTH_POTENTIALS

! SPIN_POLARIZED ! Do spin-polarized calculation

&POISSON

PERIODIC XYZ

&END POISSON

&QS

METHOD GPW

EPS_DEFAULT 1.0E-10 ! Set various epsilons for QS to values that will lead

! to energy correct up to 1e-10.

&END QS

&MGRID

CUTOFF 400 ! This is Ecut of eq. 39 in VandeVondele (2005), i.e., plane-wave cutoff

! that determines size of finest grid (see caption of Fig. 1). Cutoffs for

! the subsequent, coarser grid levels are given by eq. 39.

NGRIDS 4 ! This is N of eq. 39 in VandeVondele (2005), i.e., number of grids used.

REL_CUTOFF 40 ! This controls the grid level onto which Gaussians will be mapped.

&END MGRID

&XC

&XC_FUNCTIONAL

&PBE

PARAMETRIZATION ORIG

&END PBE

&END XC_FUNCTIONAL

&VDW_POTENTIAL

POTENTIAL_TYPE PAIR_POTENTIAL

&PAIR_POTENTIAL

TYPE DFTD3

REFERENCE_FUNCTIONAL PBE

CALCULATE_C9_TERM .TRUE.

PARAMETER_FILE_NAME dftd3.dat

R_CUTOFF 15.0

&END PAIR_POTENTIAL

&END VDW_POTENTIAL

&END XC

&SCF

SCF_GUESS RESTART ! Use data from previous run as initial guess for wavefunction.

EPS_SCF 1.0E-6 ! Threshold for converged total energy.

MAX_SCF 300 ! Maximum number of SCF iterations performed.

&OT

PRECONDITIONER NONE ! This should be stable with respect to the "Cholesky errors"

&END OT

&PRINT

&RESTART ON

BACKUP_COPIES 1

&EACH

MD 1

&END EACH

ADD_LAST NUMERIC

&END RESTART

&END PRINT

&END SCF

&PRINT

&E_DENSITY_CUBE

STRIDE 1 1 1

&EACH

MD 99999999

&END EACH

ADD_LAST NUMERIC

&END E_DENSITY_CUBE

&PDOS

COMPONENTS .FALSE.

NLUMO = -1

FILENAME dosfile

LOG_PRINT_KEY TRUE

&EACH

MD 99999999

&END EACH

ADD_LAST NUMERIC

&END PDOS

&END PRINT

&END DFT

&END FORCE_EVAL

&MOTION

&MD

ENSEMBLE NVT

STEPS 10000

TEMPERATURE 300.0 ! K

TIMESTEP 0.5 ! fs

&THERMOSTAT

REGION GLOBAL

TYPE NOSE

&NOSE

LENGTH 3 ! Length of Nose-Hoover chain

TIMECON 20.0 ! Period of typical vibrational motion in system in fs

&END NOSE

&END THERMOSTAT

&END MD

&PRINT

&RESTART

&EACH

MD 1

&END EACH

ADD_LAST NUMERIC

&END RESTART

&TRAJECTORY ON

ADD_LAST NUMERIC

FILENAME trajectory

&END TRAJECTORY

&END PRINT

&END MOTION

hut...@chem.uzh.ch

unread,

Jul 1, 2016, 8:10:52 AM7/1/16

to cp...@googlegroups.com

Hi

no this shouldn't be, but without more information I will have to guess.
You could also have a look at the timings at the end of the output to
see if some routines got slower or if all parts of the run were affected.

Two things to consider:

1) Use REFERENCE_C9_TERM TRUE in order to reduce the time for vdW in MD.
2) 30 SCF iterations in MD for such a simple system is pointing to a problem
with your setup.

regards

Juerg
--------------------------------------------------------------
Juerg Hutter                         Phone : ++41 44 635 4491
Institut für Chemie C                FAX   : ++41 44 635 6838
Universität Zürich                   E-mail: hut...@chem.uzh.ch
Winterthurerstrasse 190
CH-8057 Zürich, Switzerland
---------------------------------------------------------------

-----cp...@googlegroups.com wrote: -----To: cp2k <cp...@googlegroups.com>
From: Simiam Ghan
Sent by: cp...@googlegroups.com
Date: 07/01/2016 01:38PM
Subject: [CP2K:7883] SCF, MD run-time verses atomic species.

Dear all,I am running NVT MD with Quickstep on a box of water with 64 molecules. If I replace a water molecule with a KCl ion pair, i observe that the MD and SCF step times more than double on my setup. (MD from ~20 sec to ~46 sec). SCF iterations are converging (except the very first MD step) in around 30 steps in both cases but each SCF now takes over twice as long as before. Is there an explanation of why such a 'small' change in system could double the run time? Is KCl really so heavy to calculate compared to H2O? The number of electrons in the system increases by 8, from 512 to 520. Below my KCl input file.

Greetings,Simiam

&GLOBAL
PROJECT H2O_KCl RUN_TYPE MD PRINT_LEVEL MEDIUM
&END GLOBAL
&FORCE_EVAL
METHOD Quickstep ! GPW method.
&SUBSYS ! A subsystem: coordinates, topology, molecules and cell.
&CELL ! Supercell setup. ABC [angstrom] 12.414 12.414 12.414 ! Using 64 H2O molecules, we thus get a density of 1g/cm^3. PERIODIC XYZ ! Use PBC in all dimensions. &END CELL

&COORD UNIT angstromH -0.567712 -0.469646 -0.645913H 0.626116 -0.687796 0.308193O 0 0 0(...)K 2.1035 2.1035 4.2735Cl 4.1035 4.1035 2.6035 ###H 3.73881 3.10388 5.46104
H 3.65742 2.89924 6.989O 3.1035 3.1035 6.207(...) &END COORD

--

You received this message because you are subscribed to the Google Groups "cp2k" group.

To unsubscribe from this group and stop receiving emails from it, send an email to cp2k+uns...@googlegroups.com.

To post to this group, send email to cp...@googlegroups.com.

Visit this group at https://groups.google.com/group/cp2k.

For more options, visit https://groups.google.com/d/optout.

Samuel Andermatt

unread,

Jul 4, 2016, 5:08:52 AM7/4/16

to cp2k

Could you provide the timing sections at the end of your calculations?

Simiam Ghan

unread,

Aug 17, 2016, 7:08:12 PM8/17/16

to cp2k

Hello Juerg,

Thank you for your reply. I added REFERENCE_C9_TERM .TRUE. to the vdw section as you suggested and found that my MD simulation became twice (!) as fast. That is, time per MD step went from ~60 sec to ~30 sec for box of water with 64 molecules and K,Cl ions. I made a quick comparison of Energy_Force results with and without this flag and indeed forces and energies do not change significantly. This is great news. Why is this not a default? What's the catch? Also, would you say this is now a correct way to declare a PBE-D3 setup?

&VDW_POTENTIAL

POTENTIAL_TYPE PAIR_POTENTIAL

&PAIR_POTENTIAL

R_CUTOFF 1.5000000000000005E+01

TYPE DFTD3

PARAMETER_FILE_NAME dftd3.dat

REFERENCE_FUNCTIONAL PBE

CALCULATE_C9_TERM T

REFERENCE_C9_TERM T

&END PAIR_POTENTIAL

&END VDW_POTENTIAL

Finally, I mentioned earlier that 30 scf steps were needed to converge the same box of water with ions. I was then using EPS_SCF = 1E-06. Reducing this to EPS_SCF = 1E-05 brought the number down to 10-20 scf necessary. Does that still sound high? Is the definition of EPS_SCF documented somewhere, as it is apparently not just Energy convergence.

Cheers,

Simiam Ghan

hut...@chem.uzh.ch

unread,

Aug 18, 2016, 8:05:57 AM8/18/16

to cp...@googlegroups.com

Hi

The PBE-D3 setup looks fine.

Yes, 10-20 SCF iterations during an MD are on the very high end for
such a well behaved system. I would try the following setup (in case you
don't use it already)

&QS
....
EXTRAPOLATION ASPC
EXTRAPOLATION_ORDER 4
&END QS

&OT
PRECONDITIONER FULL_SINGLE_INVERSE
MINIMIZER DIIS
&END OT

regards

Juerg
--------------------------------------------------------------
Juerg Hutter                         Phone : ++41 44 635 4491
Institut für Chemie C                FAX   : ++41 44 635 6838
Universität Zürich                   E-mail: hut...@chem.uzh.ch
Winterthurerstrasse 190
CH-8057 Zürich, Switzerland
---------------------------------------------------------------

-----cp...@googlegroups.com wrote: -----To: cp2k <cp...@googlegroups.com>
From: Simiam Ghan
Sent by: cp...@googlegroups.com

Date: 08/18/2016 01:08AM
Subject: Re: [CP2K:8075] SCF, MD run-time verses atomic species.

Hello Juerg,Thank you for your reply. I added REFERENCE_C9_TERM .TRUE. to the vdw section as you suggested and found that my MD simulation became twice (!) as fast. That is, time per MD step went from ~60 sec to ~30 sec for box of water with 64 molecules and K,Cl ions. I made a quick comparison of Energy_Force results with and without this flag and indeed forces and energies do not change significantly. This is great news. Why is this not a default? What's the catch? Also, would you say this is now a correct way to declare a PBE-D3 setup?

&VDW_POTENTIAL POTENTIAL_TYPE PAIR_POTENTIAL &PAIR_POTENTIAL R_CUTOFF 1.5000000000000005E+01 TYPE DFTD3 PARAMETER_FILE_NAME dftd3.dat REFERENCE_FUNCTIONAL PBE CALCULATE_C9_TERM T REFERENCE_C9_TERM T &END PAIR_POTENTIAL &END VDW_POTENTIAL
Finally, I mentioned earlier that 30 scf steps were needed to converge the same box of water with ions. I was then using EPS_SCF = 1E-06. Reducing this to EPS_SCF = 1E-05 brought the number down to 10-20 scf necessary. Does that still sound high? Is the definition of EPS_SCF documented somewhere, as it is apparently not just Energy convergence.

Cheers,Simiam Ghan

Simiam Ghan

unread,

Aug 18, 2016, 1:30:55 PM8/18/16

to cp2k

Hi,

Thank you for the tips! I am not using EXTRAPOLATION or PRECONDITIONER currently. If I include them now will they affect the accuracy significantly? I am in the middle of long MD trajectories of TiO2 crystal with water and KCl ions. Also running the water box with ions. My current setup is below (same for both systems). The TiO2 system uses 1-80 scf steps for each MD step. Would you recommend these changes for that system also? Best regards.

&GLOBAL

PRINT_LEVEL MEDIUM

PROJECT_NAME brookite

RUN_TYPE MD

&END GLOBAL

&MOTION

&MD

ENSEMBLE NVT

STEPS 250

TIMESTEP 4.9999999999999989E-01

STEP_START_VAL 3751

TIME_START_VAL 4.3234999999995443E+03

ECONS_START_VAL -1.5206763646667368E+04

TEMPERATURE 3.0000000000000000E+02

&THERMOSTAT

TYPE NOSE

REGION GLOBAL

&NOSE

LENGTH 3

TIMECON 1.9999999999999993E+01

&COORD

4.2157834584794962E-01 -3.8977051434234506E+00 3.0959934319656378E+02

&END COORD

&VELOCITY

5.7927878733942134E-05 7.8527005854180164E-04 5.0048257140974906E-04

&END VELOCITY

&MASS

1.4749962622867215E+06 6.4949196930282767E+02 6.4949196930282767E+02

&END MASS

&FORCE

4.2152207720701857E-08 6.1579061284564603E-06 -8.4610128141038230E-07

&END FORCE

&END NOSE

&END THERMOSTAT

&AVERAGES T

&RESTART_AVERAGES

ITIMES_START 1

AVECPU 9.1875550803485453E+01

AVEHUGONIOT 0.0000000000000000E+00

AVETEMP_BARO 0.0000000000000000E+00

AVEPOT -1.5208725197223592E+04

AVEKIN 1.0826866556597368E+00

AVETEMP 3.0108762437707566E+02

AVEKIN_QM 0.0000000000000000E+00

AVETEMP_QM 0.0000000000000000E+00

AVEVOL 5.5116376409126919E+04

AVECELL_A 2.7174261790895553E+01

AVECELL_B 6.9659084710430690E+01

AVECELL_C 2.9116900255502028E+01

AVEALPHA 9.0000000000000071E+01

AVEBETA 9.0000000000000071E+01

AVEGAMMA 9.0000000000000071E+01

AVE_ECONS 1.6039009887889247E+02

AVE_PRESS 0.0000000000000000E+00

AVE_PXX 0.0000000000000000E+00

&END RESTART_AVERAGES

&END AVERAGES

&END MD

&CONSTRAINT

&FIXED_ATOMS

COMPONENTS_TO_FIX XYZ

LIST 148 172 196 150 174 198 64 88 112 \

259 270 281 221 234 247 149 173 197 289 \

313 337 260 271 282 225 238 251 155 179 \

203 163 187 211 164 188 212 290 314 338 \

71 95 119 266 277 288 265 276 287

&END FIXED_ATOMS

&END CONSTRAINT

&PRINT

&TRAJECTORY ON

ADD_LAST NUMERIC

FILENAME trajectory

&END TRAJECTORY

&RESTART SILENT

ADD_LAST NUMERIC

&EACH

MD 1

&END EACH

&END RESTART

&END PRINT

&END MOTION

&FORCE_EVAL

METHOD QS

&DFT

BASIS_SET_FILE_NAME BASIS_MOLOPT

POTENTIAL_FILE_NAME GTH_POTENTIALS

&SCF

MAX_SCF 300

EPS_SCF 1.0000000000000001E-05

SCF_GUESS RESTART

&OT T

PRECONDITIONER NONE

&END OT

&PRINT

&RESTART ON

ADD_LAST NUMERIC

BACKUP_COPIES 1

&EACH

MD 1

&END EACH

&END RESTART

&END PRINT

&END SCF

&QS

EPS_DEFAULT 1.0000000000000000E-10

METHOD GPW

&END QS

&MGRID

NGRIDS 4

CUTOFF 4.0000000000000000E+02

REL_CUTOFF 4.0000000000000000E+01

&END MGRID

&XC

DENSITY_CUTOFF 1.0000000000000000E-10

GRADIENT_CUTOFF 1.0000000000000000E-10

TAU_CUTOFF 1.0000000000000000E-10

&XC_FUNCTIONAL NO_SHORTCUT

&PBE T

PARAMETRIZATION ORIG

&END PBE

&END XC_FUNCTIONAL

&VDW_POTENTIAL

POTENTIAL_TYPE PAIR_POTENTIAL

&PAIR_POTENTIAL

R_CUTOFF 1.5000000000000005E+01

TYPE DFTD3

PARAMETER_FILE_NAME dftd3.dat

REFERENCE_FUNCTIONAL PBE

CALCULATE_C9_TERM T

REFERENCE_C9_TERM T

&END PAIR_POTENTIAL

&END VDW_POTENTIAL

&END XC

&POISSON

PERIODIC XYZ

&END POISSON

&PRINT

&E_DENSITY_CUBE SILENT

ADD_LAST NUMERIC

STRIDE 1 1 1

&EACH

MD 99999999

&END EACH

&END E_DENSITY_CUBE

&PDOS SILENT

ADD_LAST NUMERIC

FILENAME dosfile

LOG_PRINT_KEY T

COMPONENTS F

NLUMO -1

&EACH

MD 99999999

&END EACH

hut...@chem.uzh.ch

unread,

Aug 19, 2016, 3:11:48 AM8/19/16

to cp...@googlegroups.com

Hi

Setting an apropriate preconditioner and the DIIS minimizer will
speed up convergence and will not affect accuracy.
The ASPC order=4 setting gives you better initial wavefunctions
(=faster convergence) and better energy conservation (=stability of MD).
Again accuracy will be the same or better.

regards

Juerg
--------------------------------------------------------------
Juerg Hutter                         Phone : ++41 44 635 4491
Institut für Chemie C                FAX   : ++41 44 635 6838
Universität Zürich                   E-mail: hut...@chem.uzh.ch
Winterthurerstrasse 190
CH-8057 Zürich, Switzerland
---------------------------------------------------------------

-----cp...@googlegroups.com wrote: -----To: cp2k <cp...@googlegroups.com>
From: Simiam Ghan
Sent by: cp...@googlegroups.com

Date: 08/18/2016 07:31PM
Subject: Re: [CP2K:8079] SCF, MD run-time verses atomic species.

Hi,Thank you for the tips! I am not using EXTRAPOLATION or PRECONDITIONER currently. If I include them now will they affect the accuracy significantly? I am in the middle of long MD trajectories of TiO2 crystal with water and KCl ions. Also running the water box with ions. My current setup is below (same for both systems). The TiO2 system uses 1-80 scf steps for each MD step. Would you recommend these changes for that system also? Best regards.

Simiam Ghan

unread,

Aug 21, 2016, 1:36:13 PM8/21/16

to cp2k

Greetings Juerg,

I attached a picture showing the effect of REFERENCE_C9_TERM and EXTRAPOLATION and PRECONDITIONER on my MD simulation. Together they have reduced time per MD step from 75-200sec to 36sec and SCF count from 1-80 to 6. Now, as this seems too good to be true I did an ENERGY_FORCE calculation on a snapshot of the system with and without REFERENCE_C9_TERM: the difference in total energy is 0.02meV/atom. With/without the EXTRAPOLATION and PRECONDITIONER which you gave me the total energy difference is 1meV/atom. This difference is reduced if i tighten EPS_SCF from 1E-5 to 1E-6. I have not checked if these small changes translate to better/worse accuracy.

Could you elaborate on what REFERENCE_C9_TERM does: how does it double speed while maintaining accuracy? Why is it not a default?

Also , could you explain the definition of EPS_SCF, for example what are its units?

Many thanks,

Simiam

P.S. EXTRAPOLATION and PRECONDITIONER also put a stop to drifting in the Cons.Quantity of the .ener file, thank you.

SCf_performance.pdf

hut...@chem.uzh.ch

unread,

Aug 23, 2016, 3:43:14 AM8/23/16

to cp...@googlegroups.com

Hi

REFERENCE_C9_TERM True

keeps the C6 terms used for the calculation of the C9 term constant.
This considerably simplifies the calculation of forces.

EPS_SCF in OT is MAXVAL(dE/dCai - lambda Sab*Cbi) [the 'gradient'] in
atomic units.

regards

Juerg
--------------------------------------------------------------
Juerg Hutter                         Phone : ++41 44 635 4491
Institut für Chemie C                FAX   : ++41 44 635 6838
Universität Zürich                   E-mail: hut...@chem.uzh.ch
Winterthurerstrasse 190
CH-8057 Zürich, Switzerland
---------------------------------------------------------------

-----cp...@googlegroups.com wrote: -----To: cp2k <cp...@googlegroups.com>
From: Simiam Ghan
Sent by: cp...@googlegroups.com

Date: 08/21/2016 07:36PM
Subject: Re: [CP2K:8085] SCF, MD run-time verses atomic species.

[attachment "SCf_performance.pdf" removed by Jürg Hutter/at/UZH]

Reply all

Reply to author

Forward