[CP2K-user] [CP2K:13957] Re: Kpoints, Diagonalization algorithm and Mixing

[Lucas Lodeiro]

Dear Marcella,

As you mention, with OT method I tried with different preconditioners.
FULL_KINETIC is 20% faster than FULL_SINGLE_INVERSE, the other
preconditioners are slower. Thanks for the advice, I taked it into account,
thinking in order to reduce the CUTOFF and/or increase the EPS_SCF (Which
another could be a useful change to speed up?), but I preferred to search
another method first, to maintain the calculation tightness.

In the case of Diagonalization method with Kpoints, the extrapolation
methods based on wavefunction are not available... only LINEAR_P,
USE_PREV_P and USE_PREV_RHO, and their performance is not very good for an
MD calculation. :(

Regards

El mié., 23 sept. 2020 a las 5:40, Marcella Iannuzzi (<marci... at gmail.com>)
escribió:

> Dear Lucas,
>
> With OT you can try other preconditioners.
> The best would be FULL_ALL, but is also the slowest. FULL_SINGLE_INVERSE
> should be a good compromise.
> With the diagonalisation the extrapolation is probably better using the
> previous wavefunction.
> To speed up MD you can also consider to make the convergence criteria less
> tight, but keep an eye on the conservation of the total energy.
>
> Kind regards
> Marcella
>
> On Wednesday, September 23, 2020 at 7:40:34 AM UTC+2 Lucas Lodeiro wrote:
>
>> Hello all,
>>
>> I am trying to improve the performance of MD steps for a big slab system
>> which contains Cu atoms... the systems has ~1500 electrons and needs 2x2x1
>> kpoints to compute their electronic structure (as we explore in a PW based
>> program, Quantum Espresso).
>>
>> The first approach was to use multiple_cell 2 2 1 with just a Gamma point
>> and &OT method (with FULL_KINETIC preconditioner). The system is well
>> behaved and works fine, but the time needed for each frame is a little high
>> (87s per frame) to compute long MD.
>>
>> As a second approach, to explore if there are a method to speed up the
>> MD, we use multiple_cell 1 1 1 and 2x2x1 kpoints, with standart
>> diagonalization algorithm with broyden mixing (attached file), but the
>> performance is not so good, mainly in the extrapolation step, where ASPC
>> cannot be used. We explore LINEAR_P, and previous P and Rho, but this does
>> not work fine... and needs several SCF steps to converge.
>> The combinations for Diagonalization algorithm and mixing are various and
>> I prefer to ask if there is a way to improve diagonalization based
>> calculations.
>>
>> As the system is well behaved, is it an advantage to use another
>> algorithm and mixing method?
>> Could be a diagonalization-kpoint calculation faster than the analogous
>> one with OT?
>>
>> Another thing is based on the kpoint scheme... We want to use a 2x2x1
>> gamma centered grid... but I am not clear about MONKHORST-PACK is the Gamma
>> centered or displaced grid. MACDONALD and GENERAL, which are their
>> differences? I am familiar with the Quantum-Espresso kpoint scheme.
>>
>> There are a lot of questions, sorry, but I need some insight to not
>> search in the fog, one by one.
>>
>> Regards - Lucas Lodeiro
>>
>>
>> --
> 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 cp... at googlegroups.com.
> To view this discussion on the web visit
> https://groups.google.com/d/msgid/cp2k/1148bb83-54ef-4818-9bef-7b62f2edfa4bn%40googlegroups.com
> <https://groups.google.com/d/msgid/cp2k/1148bb83-54ef-4818-9bef-7b62f2edfa4bn%40googlegroups.com?utm_medium=email&utm_source=footer>
> .
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20200923/45a76e0d/attachment.htm>