One of the most important computational challenges in the context of the
numerical treatment of Partial Differential Equations is the generation,
management, and dynamic adaptivity of grids. Dynamic adaptivity is extremely
important in applications that require frequent changes of the grid pattern during a
simulation run. One such application example is Tsunami simulation, where waves
have to be tracked with highly resolved local grids. Arbitrary unstructured grids
that can handle dynamic adaptivity have a considerable memory and computing
time overhead. Therefore, the focus of this work is on the Sierpinski space-filling
curve-based, recursively structured and dynamically adaptive triangular grid
management system.
Space trees recursively split the geometrical domain into smaller sub-domains
according to certain predefined sub-division rules. In this thesis we concentrate on
the recursive splitting of triangles. The depth-first traversal of the binary refinement
tree inherently orders the leaf triangles according to the Sierpinski space-filling
curve. We address the challenges of traversal and management of the dynamically
adaptive triangular grid, in serial and parallel computing environment. The target
application is the parallel simulation of a simplifies version of the shallow water
equations. The application settings are simplistic in order to demonstrate the
excellent performance of our Sierpinski grid management system.
While unstructured grids may require more than 1000 bytes per triangle cell for grid
maintenance purposes, our approach uses less than 50 bytes. Due to linearization
of the refinement tree, and to the sophisticated data storage and access scheme,
the grid traversals exhibit excellent cache hit-rate, and the numerical computation
achieves very good MFLOP/sec rates. The re-meshing is about 3.5 times more
expensive than one Euler time step in terms of execution time. This ratio will
decrease significantly when higher-order discretization schemes will be applied.
The full SWE simulation with dynamic adaptivity in each time step has 80-90%
strong speed-up efficiency. Fast re-meshing and sustainable parallel scaling
capabilities will make it possible to run sub-realtime Tsunami simulations with
increased number of unknowns, resulting in much higher accuracy than possible
before
«
One of the most important computational challenges in the context of the
numerical treatment of Partial Differential Equations is the generation,
management, and dynamic adaptivity of grids. Dynamic adaptivity is extremely
important in applications that require frequent changes of the grid pattern during a
simulation run. One such application example is Tsunami simulation, where waves
have to be tracked with highly resolved local grids. Arbitrary unstructured grids
that can handle dynamic...
»