Parallel Unstructured Grid Adaption and its Use for 3D Problems
Stefan Lang (Uni Heidelberg)
Simulations of porous media or other challenging problems are
often restricted by computer resources, even when powerful
HPC machines are used.
Such soluton adaptive methods are getting more and more importance, when
realistic simulations of complex models and domains are performed.
Since geological formations or industrial cad designs may have arbitrary
computational meshes have to be very flexible to resolve the domain of concern.
Thus unstructured meshes are advantageous.
In this talk we present a grid adaption subsystem for unstructured meshes.
It allows to handle computational meshes in two and three space dimensions
out of various element types. Meshes can be adapted to solution specific
phenomena to concentrate the unknowns on those areas. Furthermore refinements
of former adaptions can be removed, which allows efficient front tracking of
instationary processes. To combine the advantages of mesh adaption and HPC,
the adaption process is parallelized for MIMD machines, this involves the
need of dynamic load balancing and migration.
Some of these functionality aspects are discussed in
the context of software design and implementation.
The previous described subsystem is integrated into the large
simulation software system UG. This allows to present the application of grid adaption to porous media
and elastoplasticity examples,
showing speedups up to 300 on 512 processors in the solution process of