Too slow for 1008 Ni atoms, 20 or more minutes per step

[Marcella Iannuzzi]

Dear  Huang,

As already pointed out, you have to improve the computational settings and 
provide enough computational resources.
If the system is metallic, you need to use a standard diagonalization 
as optimisation scheme, including the smearing of the occupation numbers 
and a proper mixing. 
The ELPA library for the diagonalisation, which replaces the corresponding 
function of the ScaLapack, is to be preferred because of the significantly 
better performance. 
The orbital transformation method is not going to work for metallic 
systems. In any case, when the SCF is not well converged, the electronic 
structure and the forces are going to be wrong. What happens then is 
totally out of control. 
Kind regards
Marcella

On Monday, April 10, 2017 at 10:45:34 AM UTC+2, yel... at gmail.com wrote:
>
> Dear  cp2k  users and developers, 
>
> I am trying to run ab initio molecular dynamics on the system including 
> 1008 Ni atoms with PBE functional. The 1*24 mpi task is too slow, around 80 
> steps per day on average; the 3*24 mpi task is slower, 38 steps per day on 
> average.
>
> Any help is highly appreciated.
>
>  
>
> Best regards,
>
>  Huang
>
>
> CP2K version 4.1
>
>  SVN source code revision svn:17462
>
>  cp2kflags: fftw3 parallel mpi2 scalapack
>
> composer_xe_2015.2.164
>
> Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz Haswell
>
>
> # Step Nr. Time[fs] Kin.[a.u.] Temp[K] Pot.[a.u.] Cons Qty[a.u.] 
> UsedTime[s]
>
>  0 0.000000 9.566949152 2000.000000000 -170544.785566084 -170535.208832257 
> 0.000000000
>
>  1 1.000000 9.646234174 2016.574776405 -170554.174550500 -170544.519479220 
> 13708.945765018
>
>  2 2.000000 9.903391164 2070.334232377 -170556.710812560 -170546.798982374 
> 1826.637025118
>
>  3 3.000000 10.258759808 2144.625134870 -170557.585906167 
> -170547.318873932 748.965154886
>
>  4 4.000000 10.622230399 2220.609774417 -170557.531726730 
> -170546.901213604 856.725823164
>
>
> &GLOBAL
>
>   ! limit the runs to 5min
>
>   ! WALLTIME 1000
>
>   ! reduce the amount of IO
>
>   IOLEVEL  LOW
>
>   ! the project name is made part of most output files... useful to keep 
> order 
>
>   PROJECT Ni-1008-2k
>
>   ! various runtypes (energy, geo_opt, etc.) available.
>
>   RUN_TYPE MD             
>
> &END GLOBAL
>
>
> &FORCE_EVAL
>
>   STRESS_TENSOR ANALYTICAL
>
>   ! the electronic structure part of CP2K is named Quickstep
>
>   METHOD Quickstep
>
>   &DFT
>
>     ! basis sets and pseudopotential files can be found in cp2k/data
>
>     BASIS_SET_FILE_NAME BASIS_SET
>
>     POTENTIAL_FILE_NAME GTH_POTENTIALS            
>
>
>     ! Charge and multiplicity
>
>     CHARGE 0
>
>     MULTIPLICITY 1
>
>
>     &MGRID
>
>        ! PW cutoff ... depends on the element (basis) too small cutoffs 
> lead to the eggbox effect.
>
>        ! certain calculations (e.g. geometry optimization, vibrational 
> frequencies,
>
>        ! NPT and cell optimizations, need higher cutoffs)
>
>        CUTOFF [Ry] 300 !500 
>
>     &END
>
>
>     &QS
>
>        ! use the GPW method (i.e. pseudopotential based calculations with 
> the Gaussian and Plane Waves scheme).
>
>        METHOD GPW 
>
>        ! default threshold for numerics ~ roughly numerical accuracy of 
> the total energy per electron,
>
>        ! sets reasonable values for all other thresholds.
>
>        EPS_DEFAULT 1.0E-7 !10 
>
>        ! used for MD, the method used to generate the initial guess.
>
>        EXTRAPOLATION ASPC 
>
>     &END
>
>
>     &POISSON
>
>        PERIODIC XYZ ! the default, gas phase systems should have 'NONE' 
> and a wavelet solver
>
>     &END
>
>
> !    &PRINT
>
> !       ! at the end of the SCF procedure generate cube files of the 
> density
>
> !       &E_DENSITY_CUBE OFF
>
> !       &END E_DENSITY_CUBE
>
> !       ! compute eigenvalues and homo-lumo gap each 20nd MD step
>
> !       &MO_CUBES
>
> !          ! compute 4 unoccupied orbital energies
>
> !          NLUMO 4
>
> !          NHOMO 4
>
> !          ! but don't write the cube files
>
> !          WRITE_CUBE .FALSE.
>
> !          ! do this every 10th MD step.
>
> !          &EACH
>
> !            MD 20
>
> !          &END
>
> !       &END
>
> !    &END
>
>
>     ! use the OT METHOD for robust and efficient SCF, suitable for all 
> non-metallic systems.
>
>     &SCF                              
>
>       SCF_GUESS ATOMIC ! can be used to RESTART an interrupted calculation
>
>       MAX_SCF 50
>
>       EPS_SCF 1.0E-4 ! accuracy of the SCF procedure typically 1.0E-6 - 
> 1.0E-7
>
>       ! do not store the wfn during MD
>
>       &PRINT
>
>         &RESTART OFF
>
>         &END
>
>       &END
>
>       
>
>       &OT
>
>         ! an accurate preconditioner suitable also for larger systems
>
>         PRECONDITIONER FULL_SINGLE_INVERSE
>
>         ! the most robust choice (DIIS might sometimes be faster, but not 
> as stable).
>
>         MINIMIZER DIIS
>
>       &END OT
>
>       &OUTER_SCF ! repeat the inner SCF cycle 10 times
>
>         MAX_SCF 20
>
>         EPS_SCF 1.0E-4 ! must match the above
>
>       &END
>
>     &END SCF
>
>
>     ! specify the exchange and correlation treatment
>
>     &XC
>
>       ! use a PBE functional 
>
>       &XC_FUNCTIONAL 
>
>          &PBE
>
>          &END
>
>       &END XC_FUNCTIONAL
>
>     &END XC
>
>   &END DFT
>
>  
>
>   ! description of the system
>
>   &SUBSYS
>
>     &CELL 
>
>       ! unit cells that are orthorhombic are more efficient with CP2K
>
>       ABC [angstrom] 21.98167992 21.98167992 25.64529419
>
>     &END CELL
>
>
>     ! atom coordinates can be in the &COORD section,
>
>     ! or provided as an external file.
>
>     &TOPOLOGY
>
>       COORD_FILE_NAME Ni-1008-2k.xyz
>
>       COORD_FILE_FORMAT XYZ
>
>     &END
>
>
>     ! MOLOPT basis sets are fairly costly,
>
>     ! but in the 'DZVP-MOLOPT-SR-GTH' available for all elements
>
>     ! their contracted nature makes them suitable
>
>     ! for condensed and gas phase systems alike.
>
>     &KIND Ni                              
>
>       BASIS_SET DZV-GTH-PADE        
>
>       POTENTIAL GTH-PBE-q18             
>
>     &END KIND
>
>   &END SUBSYS
>
> &END FORCE_EVAL
>
>
> ! how to propagate the system, selection via RUN_TYPE in the &GLOBAL 
> section
>
> &MOTION
>
> ! &GEO_OPT
>
> !   OPTIMIZER LBFGS ! Good choice for 'small' systems (use LBFGS for large 
> systems)
>
> !   MAX_ITER  100
>
> !   MAX_DR    [bohr] 0.003 ! adjust target as needed
>
> !   &BFGS
>
> !   &END
>
> !  &END
>
>  &MD
>
>    ENSEMBLE NPT_I  ! sampling the canonical ensemble, accurate properties 
> might need NVE
>
>    TEMPERATURE [K] 2000
>
>    TIMESTEP [fs] 1
>
>    STEPS 1000000
>
>    # GLE thermostat as generated at http://epfl-cosmo.github.io/gle4md 
>
>    # GLE provides an effective NVT sampling.
>
>    &BAROSTAT
>
>        PRESSURE 1.0
>
>    &END BAROSTAT
>
>    &THERMOSTAT
>
>       &NOSE
>
>       &END NOSE
>
>    &END THERMOSTAT
>
>  &END MD
>
>  &PRINT
>
>    &TRAJECTORY
>
>      &EACH
>
>        MD 100
>
>      &END EACH
>
>    &END TRAJECTORY
>
>    &VELOCITIES OFF
>
>    &END VELOCITIES
>
>    &FORCES OFF
>
>    &END FORCES
>
>    &RESTART_HISTORY
>
>      &EACH
>
>        MD 200
>
>      &END EACH
>
>    &END RESTART_HISTORY
>
>    &RESTART
>
>      BACKUP_COPIES 3
>
>      &EACH
>
>        MD 200
>
>      &END EACH
>
>    &END RESTART
>
>   &END PRINT
>
> &END
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20170411/3488fd32/attachment.htm>