These html pages are based on the
PhD thesis "Cluster-Based Parallelization of Simulations on Dynamically Adaptive Grids and Dynamic Resource Management" by Martin Schreiber.
There is also
more information and a PDF version available.
4.11 Contributions
So far, we presented a single-threaded framework for DG simulations based on the Sierpiński SFC with
stack- and stream-based communication via edges and vertices. These communication schemes are
based on previous research for vertex- [BSVB08] and edge-based [BBSV10] simulations. Our new
contributions are summarized here:
- We gave a formal introduction of the Sierpiński SFC including a formal proof of the correct
stack-communication system.
- We developed a framework for such communication schemes and introduced clear
interfaces offered to the application developer to hide the stack-based communication
complexity.
- The code generator does not only lead to an efficient method to generate tailored and
thus optimized code based on the user requirements, but also leads to optimizations such
as parameter unrolling to avoid most if-branching mispredictions.
- We separated the stacks into their functional utilization, see Section 4.5. This avoids
obsolete access of memory e.g. by separation of structure, cell-data and adaptivity state
stacks.
- An automaton table considers the propagation direction of adaptivity information.
- SIMD optimizations allow vectorized computation of fluxes for finite volume simulations.
- A prospective stack reallocation for a good balance between memory requirements and
frequent stack reallocation is derived.