[Wannier] Plans for parallel version?

[Nick Papior Andersen]

Dear Jonathan,

I have only recently used the wannier program, so have not really
delved into the code structure. I was merely interested in the
development plans and progress made.

On a note of parallalism.
My typical experiences with parallel programming is the following:
It seems *almost* obvious to utilize a shared memory model for almost
any program with large matrices (whence I think this would also be
applicable for the Wannier90 code), and ok, this is a large statement,
but small changes often leads to better shared parallelisation. It
typically requires limited effort and will yield a nice speed up.
MPI is *much* more difficult to get a nice scalability with, and often
a restructure of the serial program is required to obtain a reasonable
speed up.
Hybrids (are for me) the best way to go. MPI on the *bigger* picture,
OpenMP on the *local* picture.

It seems to me that the disentanglement is quite an entangled
computation :). In which case I fully agree with you, the OpenMP
should be the way to go.

Do you have a public site for following the development version? Or is
that off limits to the public?

However, it is nice to know that you have the parallelisation under
consideration (albeit the time issue). I look forward to see your
implementation details of Wannier90 in parallel.

Kind regards Nick




2013/4/8 Jonathan Yates <jonathan.yates at materials.ox.ac.uk>
>
>
> On 8 Apr 2013, at 10:47, Nick Papior Andersen wrote:
>
> >
> > Dear,
> >
> > I have tried searching for whether it is envisioned to do a parallel version (not just library openmp parallelization) of the code in the archive, to no avail.
> >
> > So, are there any plans of implementing an MPI, MPI/OpenMP hybrid or full OpenMP, parallel version of the program?
>
> Dear Nick,
>
>  The 2.0 version of Wannier90 is in its final stages before release - this will offer MPI parallelism for the calculation of properties ie distributing energies or kpoints for spectral properties. However, we have not attempted to parallelise the core disentanglement and wannierisation routines.
>
>  Personally, I've always felt that some sort of shared memory parallelisation would make more sense than a full distributed memory MPI implementation. Even for a big systems the memory requirements are rather small and there is not an obvious way to distribute the data. It would be an interesting problem to address. The main reason we haven't looked at this more seriously is simply lack of time. If you, or anyone else, have ideas about this, Arash and I would be happy to discuss further.
>
>  Yours
>   Jonathan
>
>
>
>
>
> --
> Department of Materials, University of Oxford, Parks Road, Oxford, OX1 3PH, UK
> tel: +44 (0)1865 612797                http://users.ox.ac.uk/~oums0549/
>
> _______________________________________________
> Wannier mailing list
> Wannier at quantum-espresso.org
> http://www.democritos.it/mailman/listinfo/wannier