Efficiency and scalability of the parallel Barnes-Hut tree code PEPC
Robert Speck, P. Gibbon, Martin W. G. Hoffmann · Advances in parallel computing · 2010
We present scaling results of the parallel tree code PEPC on an IBM BlueGene/P and identify performance bottlenecks and intrinsic algorithmic issues. Using more than 8192 processors the tree code is capable of simulating more than 100 million particles, but as our analysis shows, fundamental changes will be necessary for porting this code to petaflop systems. However, an efficiency examination reveals a very good ratio between communication and computation in the traversal process. Furthermore, we present a library version of the code, which may act as a ‘black box’ for front-end users.