The ABINIT Users Mailing List ( CLOSED )

["Aaron Zhou" <fishzhou@gmail.com>]

Hi Matteo

Thank you very much for your super fast reply. the keywords you told
me does help me reduce the size of kss, and calculations much faster.

But i still have a few more questions here:

1. How do i make sure the diagnolization is paralleled in the KSS
step? I didn't see anything like: "P newkpt: treating     17 bands
with npw=    4356 for ikpt=   1 by node    0"

2. How do I guarantee that i don't lose too much accuracy when I
specify the value of nbandkss?

3. continuous of last question, What factors decide the values of
nband and nbandkss? I do know that the easiest and the slowest way is
to increase the value little by little and see when it converges, but
it's not physics, would you mind tell me how I should do that?

say, there are 6 Si and 6 H atoms in my supercell, and there are
6*2+6=18 occupied bands, so what kind of fomula or physics you are
based to decide these two parameters?

Thanks a lot for taking your time

Aaron



On Tue, Nov 11, 2008 at 2:49 PM, Matteo Giantomassi
<Matteo.Giantomassi@uclouvain.be> wrote:
>> Hi Dear Fabien,
>>
>> I just saw this email and felt quite curious about how you did that.
>> because I
>> was trying to run GW calculation in parallel using both Abinit 5.4.3 and
>> 5.6.2.
>> There are only 12 atoms in my system. feels like it takes forever to
>> generate
>> the KSS.
>>
>> I remembered in one of your email you replied to the forum saying that you
>> have
>> implemented the parallelisation in Abinit 5.6.x, but I didn't feel too
>> much
>> improvement. I only see "treating xxx bands with npw=xxx for ikpt=1 by
>> node 0"
>> or 1... in DFT part, but I didn't see the same output in the KSS step.
>> Could
>> please you tell me what I have missed?
>> (the input file is at the end of this email: a 2D silicon surface band gap
>> calculation)
>
>  Dear Zhiyong
>
>  The problem is that you are using kssform=1 (default option) and asking
> for the
>  maximum allowable number of bands (nbandkss -1).
>  Therefore the code is performing a full diagonalization of the
>  Hamiltonian matrix and I guess the number of plane waves you are using is
> quite large
>  due to the size of the unit cell. Actually the inversion of the Hamiltonian
>  is parallelized over k-points and spin but the CPU effort required by the
> diagonalization
>  scales as N_G**3.
>  That's the reason why your calculation seems to be blocked during the KSS
> generation.
>
>  Possible solutions: either reduce the number of bands to be saved on the
> KSS file
>  by specifying a positive value for nbandkss or use kssform=3 to produce
> the wave functions and eigenvalues
>  through the standard conjugate gradient approach.
>
>  In the former case, the code will perform a partial diagonalization of
> the Hamiltonian calculating only the first
>  nbandkss states. This allows to save CPU time since the computational
> effort now scales as N_bands * N_G**2.
>
>  In the later case you have to perform an explicit NSCF calculation for
> all the
>  states you want to store in the KSS file, but you can avoid the direct
> inversion of the Hamiltonian
>  Example:
>
> ####################################
> #  DATASET 1: DFT SCF
> ####################################
> iscf1         7
> tolvrs1       1.d-10
> prtden1       1
> ####################################
> #  DATASET 2: KSS creation
> # Read the previous density, perform a standard
> # band structure calculation dumping the KSS file at the end
> ####################################
> getden2      -1
> kssform2      3
> nbandkss2    12
> nband2       15
> nbdbuf2       3
> iscf2        -3
> tolwfr2       1.d-10
> ####################################
>
> Hope it helps
> Best regards,
> Matteo Giantomassi
>
>
>
>