Scalable Fast Multipole Method for Large-Scale Electromagnetic Scattering Problems on Heterogeneous CPU-GPU Clusters

This letter investigates a hybrid framework for the solution of large-scale electromagnetic scattering problems by using the fast multipole method (FMM) on a heterogeneous CPU-GPU system. Enabling the use of both CPU and GPU resources available in the cluster allows for solving significantly larger problem sizes than using only CPU or GPU resources. The performance is evaluated on a 13-node cluster equipped with NVIDIA Tesla M2090 GPUs. The experimental results demonstrate that our FMM implementation on CPU-GPU is up to 72.3× faster than that of the 12-core eight-node CPU implementation. The scalability of the CPU-GPU implementation is very close to the theoretical expectations.