[Wannier] wannier90 + vasp parallelism

Aron Szabo szaboa at iis.ee.ethz.ch
Tue Jul 22 16:36:24 CEST 2014


Dear Wannier90 users,

I have noticed that calculating the initial projections for wannier90 
with VASP doesn't speed up with the number of CPU-s used. Is it an 
installation error, or is it only implemented in sequential, like 
wannier90.x itself? And is it the same with other dft-tools (quantum 
espresso, abinit)?

With large supercells it's a very time consuming part, taking days for a 
system with about 500 Wannier functions. I've somewhat circumvented the 
problem by calculating separately the projections on the different atom 
and orbital types, then concatenating (with proper indexing - I can 
share my Matlab script if anyone is interested) the files, but it's 
still very inefficient, because the same overlaps have to be calculated 
first every time (I can't tell VASP that I only need the projections), 
and on supercomputers, where only whole nodes can be reserved, still 
only one cpu will work per node.

So I would appreciate any hints how to deal with large systems 
efficiently, if anyone has any experience with it.

Thanks in advance!

Aron Szabo
(PhD student at ETH Zurich)


More information about the Wannier mailing list