DENSITY_PROPAGATION with the AUXILIARY_DENSITY_MATRIX_METHOD

[Luca]

Dear Florian and dear Samuel

 

first of all thank you for your answer. The system on which I am focused is 
a coated metallic cluster of about one hundred atoms (600 electrons in 
total). I would like to obtain the electronic spectrum of such a system at 
b3lyp level and it is good to know that for such type of system the 
standard algorithm is faster than the linear scaling one.  

 

About my computational resources, at the moment, I have access to Dual 
Intel Haswell E5-2670 CPUs (2.3 Ghz) giving 24 cores per node with 56Gb/s 
FDR Infiniband interconnect. 

When running on 192 cores CP2K needs 560 sec for each real-time-propagation 
step (rtps), that is too high to obtain a spectrum in a reasonable time 
(for me reasonable would be 24-48 hours for each polarization direction of 
light). In this case I used the PADE approximant for the exponential but 
considering your answer I have planned new tests with the ARNOLDI 
algorithm. About the ADMM basis set I tested FIT11 for the metal and cpFIT3 
for the other species. Reducing the basis sets to FIT3 leads to a reduction 
of the time for rtps to 460 sec, that is still too high and probably will 
give a spectrum that would not be accurate. Is there the possibility to 
reduce that time without increasing the number of cores? Which are the 
keyword with which I can 'play' to reduce the time for rtps? 

 

Finally my latest question concerns the segmentation fault that I obtain 
when I use the  AUXILIARY_DENSITY_MATRIX_METHOD with the cam-b3lyp 
functional. In this case the problem comes out at the beginning of the scf 
procedure. Did you ever try that combination ?  Did you ever run up against 
a similar issue?

 

For Samuel: the INPUT that gave me the segmentation fault is :

 

&FORCE_EVAL

  METHOD QUICKSTEP

  &DFT

    &AUXILIARY_DENSITY_MATRIX_METHOD

      METHOD BASIS_PROJECTION

      ADMM_PURIFICATION_METHOD NONE

    &END

    &REAL_TIME_PROPAGATION

       MAX_ITER 7

       MAT_EXP BCH 

       PROPAGATOR ETRS

       EXP_ACCURACY 1.0E-11

       EPS_ITER 1.0E-2

       INITIAL_WFN SCF_WFN

       DENSITY_PROPAGATION .TRUE.

    &END

    BASIS_SET_FILE_NAME ../../../data/BASIS_MOLOPT

    BASIS_SET_FILE_NAME ../../../data/BASIS_ADMM

    POTENTIAL_FILE_NAME ../../../data/POTENTIAL

    &MGRID

      CUTOFF 100

    &END MGRID

    &QS

      EPS_DEFAULT 1.0E-10

      MAP_CONSISTENT

    &END QS

    &POISSON

      PERIODIC NONE

      PSOLVER MT

    &END

 

    &SCF

      MAX_SCF 5

      EPS_SCF 1.0E-6

      SCF_GUESS ATOMIC

      &OT ON

        ROTATION

      &END OT

    &END SCF

    &LS_SCF

      EPS_FILTER 1.0E-10

      EPS_LANCZOS 1.0E-8

      MAX_ITER_LANCZOS 30

      SIGN_SQRT_ORDER 3

    &END LS_SCF

    &XC

      &XC_FUNCTIONAL NONE

      &END XC_FUNCTIONAL

      &HF

        &SCREENING

          EPS_SCHWARZ 1.0E-10

          EPS_SCHWARZ_FORCES 1.0E-10

          SCREEN_ON_INITIAL_P FALSE

        &END

        &INTERACTION_POTENTIAL

          POTENTIAL_TYPE COULOMB

        &END

        &MEMORY

          MAX_MEMORY 50

          EPS_STORAGE_SCALING 0.1

          TREAT_FORCES_IN_CORE TRUE

        &END

      &END

    &END XC

  &END DFT

  &SUBSYS

 

    &CELL

      ABC 8.0 8.0 8.0

      PERIODIC NONE

    &END CELL

    &COORD

    O   0.000000    0.000000   -0.065587

    H   0.000000   -0.757136    0.520545

    H   0.000000    0.757136    0.520545

    &END COORD

    &KIND H

      BASIS_SET DZVP-MOLOPT-GTH

      POTENTIAL GTH-PBE-q1

      AUX_FIT_BASIS_SET cFIT3

    &END KIND

    &KIND O

      BASIS_SET DZVP-MOLOPT-GTH

      POTENTIAL GTH-PBE-q6

      AUX_FIT_BASIS_SET cFIT3

    &END KIND

  &END SUBSYS

&END FORCE_EVAL

&GLOBAL

  PROJECT H2O-hybrid-EMD

  RUN_TYPE EHRENFEST_DYN

  PRINT_LEVEL MEDIUM

&END GLOBAL

&MOTION

  &MD

    ENSEMBLE NVE

    STEPS 2

    TIMESTEP [au_t] 0.25

    TEMPERATURE 300.0

  &END MD

&END MOTION

On Tuesday, May 5, 2015 at 12:10:11 PM UTC+2, Florian Schiffmann wrote:
>
> Hi Luca,
>
> you are trying some serious stuff there ;-) I am afraid, the linear 
> scaling version doesn't support hybrids yet. It's related to the problem 
> that the sparsity of the exchange part does not match the sparsity of the 
> overlap matrix which is used throughout CP2K (that might have caused the 
> segfault).
> However, for medium sized systems, up to 1600 atoms the standard algorithm 
> is anyway faster if you use the Arnoldi method for the matrix exponential.
> Here, it is possible to use hybrid functionals. However, ADMM is somewhat 
> more tricky as the purified versions do not yield an analytic KS matrix 
> which is required by the RTP method. Therefore ADMM/RTP does only work with 
> PURIFICATION NONE. I only tested it on the spectrum of a water molecule and 
> it seemed to work fine, however there is no absolute warranty that it will 
> always work. So if you want to use it, handle with care.
>
> Cheers
> Flo
>
 
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20150506/5272bdf3/attachment.htm>