Fast Multipole Method on the Cell Broadband Engine: the Near Field Part
Résumé
The implementation of the near field part of the Fast Multipole Method, which solves hierarchically N-body problems, is presented for the Cell Broadband Engine. Algorithmic and hardware issues, as well as code specific optimizations, are detailed for this first step towards the first full implementation of a hierarchical N-body method on the Cell processor. We compute up to more than 8.5 billion interactions per second (115.8 Gflop/s) on the Cell Processor, and up to more than 17 billion interactions per second on the IBM QS20 blade (230.4 Gflop/s), for both uniform and non-uniform distributions of particles.