[DFTB-Plus-User] DFTB+ Scaling
aradi at uni-bremen.de
Wed May 3 15:19:56 CEST 2017
> In the DFTB+ paper (2007 JPCA), it is mentioned that the bottleneck is
> the diagonalization and it's scaling is O(N^3) [due to sparse to square
> conversion and using LAPACK routines]. May I know, whether this is the
> scaling of the current version of DFTB+ or any improvements are there?
The bottleneck is indeed the diagonalization (and not the sparse->dense
conversion). N is the number of basis functions, yes. Currently, we have
a developer version, where we use PEXSI instead of diagonalization,
there your scaling should vary between O(N) and O(N^2) depending on the
dimensionality of your system. That won't give you eigenvalues and
eigenvectors, though, only densities. If you want to have eigenvalues as
well, you will have to face O(N^3). But, the parallel version at least
helps you to distribute it among many computers to get reasonable wall
clock times for fairly large systems.
Dr. Bálint Aradi
Bremen Center for Computational Materials Science, University of Bremen
-------------- next part --------------
A non-text attachment was scrubbed...
Size: 836 bytes
Desc: OpenPGP digital signature
More information about the DFTB-Plus-User