[Pw_forum] big calculations II

[Ricardo Faccio]

Dear nicola
Thanks gr your email, i would like to take the opportunity and ask you  
a further question. I have a similar situation than José. I'm working  
with a cobalt porphyrin deposited over a metallic substrate (similar  
to your recent PRL). I optimized the structure with gamma point with  
25 and 250 ecut and rhocut respectively. It takes almost the 100% of  
my 8Gb RAM quad core computer. Now i want to get a reasonably DOS fir  
this system, but i need more k-points, and thus I will need more RAM  
memory. Would you recommend anything else for getting a good DOS for  
my quad core 8 GB RAM machine?
Thank you in advance.
Ricardo

--------------------------------------------
Dr. Ricardo Faccio
Prof. Adj. de Física

Av. Gral. Flores 2124. CC 1157. CP 11800.
Phone: + 598 2 924 9859
Fax: + 598 2 924 1906
Web: http://cryssmat.fq.edu.uy/ricardo/ricardo.htm
---------------------------------------------

El 19/01/2011, a las 21:54, Nicola Marzari <nicola.marzari at materials.ox.ac.uk 
 > escribió:

>
>
> Given your system size, it's likely that gamma point
> is good enough - a full relaxation with 18 Ry cutoff and
> the gamma point would bring you very very close to the correct
> result.
>
> Then, try 30 Ry and Gamma point - do the results change a lot ?
> Probably note, and maybe 25 Ry is enough (there are tests of
> the C ultrasoft on the pseudopotential page). I'd assume a dual of
> 8, i.e. a cutoff for the charge density of 8 times the cutoff on the
> wfcs.
>
> The 2 2 2 1 1 1 mesh will be almost certainly more than enough
> (i.e. a 2x2x2 shifted) -  you won't need more than these, and probably
> gamma is enough.
>
> Note that if you use the keyword gamma for the k-point mesh
> pw takes into account that at gamma wavefunctions are real,
> halving the memory and speed.
>
> Last, my favourite trick is to use 1/4 1/4 1/4 as a k-point
> (in relative coordinate), and stop pw from replicating that
> with nosym=true . In this way, you use only one k-point
> (albeit with complex wavefunctions) and the results are almost
> indistinguishable from the 2 2 2 1 1 1 mesh.
>
> For even larger systems, it might be worth becoming familiar with
> cp, and use a combined damped dynamics on electrons and ions
> (actually, only on the electrons tend to be enough). This can
> be quite efficient for huge systems, and faster/smaller calculations
> than pw.
>
> At a certain point we shoudl write a small tutorial about that...
>
>          nicola
>
>
> On 1/19/11 8:58 PM, José Alberto Pires Fernandes wrote:
>> Hello
>>
>> I'm a newby in calculations with Quantum espresso and I
>> think I'm trying to do an impossible task.
>>
>> I want to do a structural optimization to a cell with 338
>> atoms (C, H, N and O) and approximate dimensions of
>> 15*15*15  angstroms. If I choose Ecut of 18 Ry and a
>> kpoint mesh of 1*1*1, with pz-rrkjus pseudopotential the
>> optimisation takes 13 hours. With a 2*2*2
>> kpoint mesh is taking more than 4 days, and hasn't
>> finished yet. I believe that for reliable results I need a
>> larger Ecut (at least 30 Ry, I suppose) and a number of
>> kpoints in the order of thousands, which would imply in a
>> calculation of several years.
>>
>> The cell is a P1 triclinic cell but about 300 atoms are
>> fixed, because I only need to do particial optimisation.
>> So, what can I do to have reliable results in useful time,
>> or Quantum Espresso is not suitable for my task?
>>
>>
>>
>>
>> José Fernandes
>> Departamento de Quimica
>> Universidade de Aveiro
>> 3810-193 Aveiro
>> Portugal
>> Tel: +351234370720
>> Fax: +351234370084
>> _______________________________________________
>> Pw_forum mailing list
>> Pw_forum at pwscf.org
>> http://www.democritos.it/mailman/listinfo/pw_forum
>
>
> -- 
> ----------------------------------------------------------------------
> Prof Nicola Marzari    Department of Materials    University of Oxford
> Chair of Materials Modelling  Director, Materials Modelling Laboratory
> nicola.marzari at materials.ox.ac.uk     http://mml.materials.ox.ac.uk/NM
> _______________________________________________
> Pw_forum mailing list
> Pw_forum at pwscf.org
> http://www.democritos.it/mailman/listinfo/pw_forum