this is for holding javascript data
Joseba Alberdi-Rodriguez edited Parallelization.tex
over 9 years ago
Commit id: 0701a45be8f0eebef2be5e151dbf2e97ecd03882
deletions | additions
diff --git a/Parallelization.tex b/Parallelization.tex
index 497628b..2d51575 100644
--- a/Parallelization.tex
+++ b/Parallelization.tex
...
In order to numerically calculate physical observables we need to discretise
the problem. To this end, we use a mesh representation (e.g. which changes the charge
density and the electrostatic potential to discrete functions), with values defined
over a finite number of points distributed over a mesh. As shown in the Figure \ref{fig:partitioning}, two parameters define the mesh: the length and the spacing. Once constructed, it is partitioned over the processes (6 in the example). Real-space mesh approach is really efficient in terms of performance. {\sc Octopus} has shown a high performance in current architectures \cite{Alberdi_2014}, not only because of the mesh-partitioning but also because of other kind of parallelizations. One of those natural way of parallelization is the state parallelization. Indeed, every state (or wavefunction) of the system is represented over a discrete mesh, being each state independent of the others.
The shape of the mesh can be parallelepiped, sphere or adaptive (as shown in Figure \ref{fig:partitioning}. In general, besides the general mesh, another one is used for the Poisson solver. This latter one has to be necessarily parallelepiped.