Next: Computational Costs
Up: Castep
Previous: Castep
Contents
Parallelisation
Since the eigenstates at different k-points are almost independent of
each other, interacting only via the density, it is natural to
distribute the data over the k-points. Unfortunately the number of
k-points required decreases with increasing system size, and for large
systems
is O(1). This means that on HPC machines it is rarely
possible to distribute the data and workload efficiently using k-point
parallelism alone.
A further distribution strategy used by Castep is to distribute the
data by the Fourier components, i.e. the `G-vectors'. Since the number
of plane-waves is often large, and grows with simulation cell size,
this enables efficient data distribution and load-balancing.
By combining both k- and G-parallelism Castep has demonstrated
excellent scaling properties from 1 to O(100) nodes across many
different computer architectures.
Sarfraz A Nadeem
2008-09-01