[Wannier] wannier90 + vasp initial projections parallelism
Aron Szabo
szaboa at iis.ee.ethz.ch
Tue Jan 13 13:21:58 CET 2015
Dear Giovanni,
thank you very much for your help.
Best,
Aron
On 01/13/2015 01:12 PM, Giovanni Pizzi wrote:
> Dear Aron,
> indeed, the calculation of overlaps and projections is done by the
> code interface, so for direct help of the VASP interface, you should
> address directly to the VASP developers (disabling the calculation of
> one part, or adding parallelism on the calculation of initial
> projections, should be easy, I believe).
>
> If you want to use Quantum ESPRESSO, as Nicola suggested, the
> executable to run is pw2wannier90.x (and since all tutorials of
> Wannier90 are written with this interface in mind, you will already
> find instructions and example input files to pw2wannier90.x in the
> Wannier90 tutorial/examples).
> You will want to explicitly define write_amn and write_mmn in the
> input file to pw2wannier90.x to select whether to calculate these
> matrices (other options exist to choose whether to write the u_nk for
> Wannier function plotting [write_unk], etc.; see the source file
> PP/src/pw2wannier90.f90 inside the QE distribution, for instance).
>
> Best,
> Giovanni
>
>
> On 01/13/2015 12:35 PM, Aron Szabo wrote:
>> Dear Nicola,
>>
>> I meant to do separately the calculation of the overlaps (with the
>> unmodified code. This part seemed to scale with the cpu-s, and it
>> doesn't depend on the projections), and the calculations of the
>> projections that could be "parallelized" by running vasp many times
>> with selecting just a few projections in the input for each, and just
>> concatenate the results (and change the band indices). That's how I
>> do it now, and it works, but the overlaps are calculated in each
>> case, while they are identical, and it takes a lot of time
>> unnecessarily, making it impossible to go above a certain size (300
>> atoms approx.).
>>
>> Anyway, my next question would have been that how is it done in
>> quantum espresso. So thank you for your answer, I'm going to check
>> whether I could do all my tasks with quantum espresso. It seems to be
>> more practical in this respect then.
>>
>> And thank you for your offer to meet in person. I will probably use
>> the opportunity, and contact you in private to arrange the details.
>>
>> Best regards,
>> Aron
>>
>> On 01/13/2015 12:00 PM, Nicola Marzari wrote:
>>>
>>> Hi Aron,
>>>
>>>
>>> but without overlaps you won't have MLWFs? I guess you are just trying
>>> to do "poor man's Wannier" via projections?
>>>
>>> Still, might be worth trying quantum-espresso, since those calculations
>>> are parallelized. All plane wave codes are fairly similar, at the
>>> end...
>>>
>>> I'll be in ETHZ next week (wed/thu), so we could always catch up there.
>>>
>>> nicola
>>>
>>>
>>> On 13/01/2015 11:55, Aron Szabo wrote:
>>>> Dear Wannier90 developers,
>>>>
>>>> I'm using vasp 5.3 and wannier90 2.0. I've noticed that the
>>>> calculation
>>>> of the initial projections for the wannierization in vasp is
>>>> sequential,
>>>> it doesn't speed up with the increasing number of cpu-s. For large
>>>> systems I'm calculating separately smaller sets of projections to
>>>> emulate some parallelism, but there is a huge overhead from
>>>> calculating
>>>> the overlaps first in each case. It can take 1-2 days just to start
>>>> calculating the projections when I have 2-300 atoms (1-2000 bands). So
>>>> generating the projections can take more time than the actual
>>>> wannierization. If the calculation of the overlaps could be somehow
>>>> switched off, it would already solve this problem practically, because
>>>> then the calculation of the projections would scale nicely, and the
>>>> overlaps could be calculated in a separate run. I'm not good in
>>>> fortran,
>>>> so I'm asking you whether it could be done maybe by just commenting
>>>> out
>>>> a few lines somewhere either in vasp's or wannier90's code, and if so,
>>>> where is it exactly? I'm also interested if it is planned at all to
>>>> parallelize the wannierization sometimes in the future, or is the
>>>> problem itself something that's not possible to parallelize
>>>> efficiently?
>>>>
>>>> Thanks in advance,
>>>> Aron Szabo
>>>> PhD student at IIS, ETH Zurich
>>>> _______________________________________________
>>>> Wannier mailing list
>>>> Wannier at quantum-espresso.org
>>>> http://mailman.qe-forge.org/cgi-bin/mailman/listinfo/wannier
>>>
>>
>> _______________________________________________
>> Wannier mailing list
>> Wannier at quantum-espresso.org
>> http://mailman.qe-forge.org/cgi-bin/mailman/listinfo/wannier
>
>
More information about the Wannier
mailing list