Results
We decide to use p4est as basis for a scalable grid library for
dynamically-adaptive tree-structured grids.
We optimized p4est to facilitate minimal-invasive integration by providing
random-access to arbitrary direct cell neighbors. Additionally, we implemented
"virtual cells" which are cells embedded at refinement boundaries. These cells
do not hold numerical degrees of freedom but are able to store data persistently
and include this data in the ghost exchange.
They allow writing numerical algorithms which only interact with cells of the
same size.
This enhanced version of p4est has been integrated into ESPResSo in multiple
steps such that different physical subsystems can now be used on p4est-based
grids:
1. Short-range molecular dynamics for modelling a molecular ensemble.
2. The lattice-Boltzmann method to subject this ensemble to a background flow.
3. Long-range molecular dynamics to model electrostatic interactions of
charged particles
4. The continuum model of the electrokinetic to model ionic solvents.
Coupling those subsystems in a physically correct way is work-in-progress
together with subproject C.5.
To reduce the overhead necessary for coupling all necessary information has to
be locally available at each process. To this end, we have developed distributed
algorithms for dynamically changing the grid resolution and/or changing the
partitioning between processors together with subproject D.9.
Additionally, we have performed first scalability tests for the
lattice-Boltzmann method and for short-range MD at the Hazel Hen machine at
HLRS. In both cases we obtained promissing results.
BU:
Dynamic-adaptive simulation of a lid-driven cavity scenario.
The initial grid of discretization level 3 is shown on the [left|top].
This grid is statically refined up to level 5 for cells close to the
geometric boundaries.
After 2048 timesteps using dynamic-adaptivity based on fluid velocity and
vorticity we obtain the grid on the [right|bottom].