Modeling the Performance of Fast Mulipole Method on HPC platforms

• Main Author: Ibeid, Huda
• Other Authors: Keyes, David E.
• Language: en
• Published: 2012
• Online Access: http://hdl.handle.net/10754/267452; http://repository.kaust.edu.sa/kaust/handle/10754/267452

The current trend in high performance computing is pushing towards exascale computing. To achieve this exascale performance, future systems will have between 100 million and 1 billion cores assuming gigahertz cores. Currently, there are many efforts studying the hardware and software bottlenecks for building an exascale system. It is important to understand and meet these bottlenecks in order to attain 10 PFLOPS performance. On applications side, there is an urgent need to model application performance and to understand what changes need to be made to ensure continued scalability at this scale. Fast multipole methods (FMM) were originally developed for accelerating N-body problems for particle based methods. Nowadays, FMM is more than an N-body solver, recent trends in HPC have been to use FMMs in unconventional application areas. FMM is likely to be a main player in exascale due to its hierarchical nature and the techniques used to access the data via a tree structure which allow many operations to happen simultaneously at each level of the hierarchy. In this thesis , we discuss the challenges for FMM on current parallel computers and future exasclae architecture. Furthermore, we develop a novel performance model for FMM. Our ultimate aim of this thesis is to ensure the scalability of FMM on the future exascale machines.