Parallelization on Heterogeneous Multicore and Multi-GPU Systems of the Fast Multipole Method for the Helmholtz Equation Using a Runtime System

International audienceThe Fast Multipole Method (FMM) is considered as one of the top ten algorithms of the 20th century. The FMM can speed up solving of electromagnetic scattering problems. With N being the number of unknowns, the complexity usually O(N 2) becomes O(N log N ) allowing a problem with hundreds of millions of complex unknowns to be solved. The FMM applied in our context has a serious drawback: the parallel version is not very scalable. In this paper, we present a new approach in order to overcome this limit. We use StarPU, a runtime system for heterogeneous multicore architectures. Thus, our aim is to have good efficiency on a cluster with hundreds of CPUs, and GPUs. Much work have been done on parallelization with advanced distribution techniques but never with such a runtime system. StarPU is very useful, especially for the multi-level algorithm on a hybrid machine. At present, we have developed a multi-core and a GPU version. The techniques for distributing and grouping the data are detailed in this paper. The first results of the strategy used are promising