High-Performance Simulation of Electromagnetic Scattering by 3D Objects Using the GPU-accelerated Parallel MLFMA

This paper proposes a high-performance massively parallel multilevel fast multipole algorithm (MLFMA) on clusters with graphics processing units (GPUs) accelerators. ELLPACK format-based schemes are proposed to ensure a high computational throughput for GPU accelerated evaluation of interpolation and anterpolation operations in MLFMA. The program is further optimized using the mechanism of stream and the shared-memory/register hierarchical memory architecture of GPU. The implementation of all the proposed techniques make the aggregation/disaggregation calculations an order of magnitude faster than the conventional ones, and halved the iteration time for an aircraft model with 10 billion unknowns, which is a significant progress in GPU accelerated massively parallel MLFMA.