[Pw_forum] Nscf with large number of bands for GW

[Paolo Giannozzi]

Using 'cg' may sometimes work, but the machinery of iterative
diagonalization is aimed towards finding the lowest M eigenvectors and
eigenvalues when M << N = dimension of the basis set (number of plane
waves). It's neither efficient nor numerically stable if M is a
sizable fraction of N.

Paolo

On Wed, Apr 5, 2017 at 6:43 PM, Kanak Datta <kanak at umich.edu> wrote:
> Dear Thierry
>
> I faced the same Cholesky  problem in my calculation. Although, I calculated
> 2000 bands. The problem was gone when we used 'cg' diagonalization. I also
> reduced the convergence threshold to 1d-8. Although, the simulation became
> very slow, possibly due to 'cg' diagonalization.
>
> Thanks
> Kanak
>
> _______________________________________
> Kanak Datta
> Graduate Student
> Electrical  Engineering and Computer Science
> University of Michigan, Ann Arbor
>
> On Wed, Apr 5, 2017 at 11:22 AM, Thierry Clette
> <thierry.clette at uclouvain.be> wrote:
>>
>> Hello,
>>
>> I want to do some nscf calculations in order to compute GW corrections
>> afterwards. I know that I need a lot of bands to converge, around 3000 (I
>> already did some convergence studies in another code).
>> A example of an input is attached.
>> This seems to be a problem for the code, as when I go higher than about
>> 500 bands, I get the ¨error computing cholesky¨, which means the
>> diagonalization fails.
>>
>> I read a lot of different topics in the mailing list reporting this
>> problem, with various solutions such as :
>> - parallelization on bands
>> - CG diagonalization
>> - bigger ¨diago_thr_init¨ (even if I need to be quite precise in the end)
>> - beginning from a previous calculation with ~500 bands.
>>
>> I tried different things and it still didn't work. BTW, I know the
>> calculation is possible, because I did it without problems in Abinit with
>> the same parameters (Abinit seems to work differently).
>>
>> At this stage, I really don't know what I should do. Has someone already
>> done successful calculations with so many bands ?
>>
>> Thanks in advance,
>>
>> Thierry Clette
>>
>>
>>
>>
>> _______________________________________________
>> Pw_forum mailing list
>> Pw_forum at pwscf.org
>> http://pwscf.org/mailman/listinfo/pw_forum
>
>
>
> _______________________________________________
> Pw_forum mailing list
> Pw_forum at pwscf.org
> http://pwscf.org/mailman/listinfo/pw_forum



-- 
Paolo Giannozzi, Dip. Scienze Matematiche Informatiche e Fisiche,
Univ. Udine, via delle Scienze 208, 33100 Udine, Italy
Phone +39-0432-558216, fax +39-0432-558222