GPU communication in multi-GPU parallel FDTD algorithm based on OpenMP and NCCL

This paper proposes a parallel finite difference time domain (FDTD) algorithm based on multi-GPU, which assigns a large number of threads from multiple GPUs to each Yee cell to perform parallel updates of various electromagnetic field components within the computational domain. To address the optimization of the subdomain communication, a comparison was conducted between a CPU indirect data communication based on OpenMP and a GPU direct data communication based on NCCL. The numerical results indicate that the proposed algorithm based on NCCL and OpenMP significantly enhances computational efficiency while maintaining algorithmic accuracy, and the NCCL achieves higher efficiency.