Ralf Deiterding
Construction and Application of an AMR Algorithm for Distributed Memory Computers
The blockstructured adaptive mesh refinement (AMR) technique of Berger
and Collela is tailored especially for the efficient simulation of
instationary supersonic flow phenomena on high-performance
computers. While the parallelization of the method on shared memory
architectures is relatively straight-forward, appropriate distribution
strategies for the emerging generation of distributed memory machines
are a topic of on-going research. In this talk, we propose a
locality-preserving domain decomposition that partitions the entire
AMR hierarchy from the base level on. We demonstrate that our approach
reduces the communication costs and simplifies the implementation. A
precise topological notation is employed to introduce the sequential
AMR method and to explain the necessary extensions in
parallel. Emphasis is put on the effective parallelization of the flux
correction procedure, which is indispensable for conservative finite
volume schemes. Briefly, we sketch the design of our public-domain
framework AMROC (Adaptive Mesh Refinement in Object-oriented C++) that
implements the proposed algorithms generically. Beside an easily
reproducible standard benchmark, we show AMROC simulations of
transient hydrogen-oxygen detonations with detailed non-equilibrium
chemistry for nine gaseous species that impressively demonstrate the
advantages in employing a massively parallel AMR algorithm. In
particular, the first successful three-dimensional regular cellular
structure simulation with detailed chemistry and a two-dimensional
detonation diffraction, which is in perfect agreement with
experimental results, are presented. These simulations were run
effectively on a Linux-Beowulf-cluster of 48 CPUs and spent 1500-2000h
CPU time in the numerical update routines.