Hierarchical methods like the Barnes-Hut (BH) tree code and the Fast Multipole Method (FMM) are important tools to decrease the algorithmic complexity of solving the N-body problem. We report on recent efforts to achieve massive scalability on IBM BlueGene/Q and other highly concurrent supercomputers with our BH tree code PEPC by replacing a thread synchronization strategy based on traditional lock data structures with atomic operations. We also describe the integration of the Dual Tree Traversal a more recent algorithm that combines advantages of both BH tree code and FMM into PEPC. We explain how the scalability of this algorithm is influenced by the existing communication scheme and we propose a modification to achieve better distributed memory scalability.
展开▼