QuinnKaleCSSI_PI_Poster.pdf

Tree-based algorithms for spatial domains scale poorly without extensive experimentation, especially when the application exhibits dynamic behavior (e.g. when particles migrate in particle-based simulations for astronomy, with high density variations). The ParaTreeT project has developed a highly reusable library, with parallel abstractions supported by efficient parallel algorithms, that facilitates creation of efficient tree-based codes that scale to large parallel machines, for a variety of different algorithms. ParaTreeT was developed by collaboration between two research groups, one with expertise in computational astronomy and the other with expertise in parallel programming and adaptive runtime systems. The framework features built-in tree types and traversals, a novel partition-subTree dual decomposition, and pointer-based tree-structured shared-memory particle caches. It has been demonstrated for gravity, SPH, and planetary disk formation, on 20,000+ cores. As an example of productivity boost, the gravity algorithm required only 150 lines of code when using the framework.