Collaborating CPU and GPU for the electromagnetic simulations with the FDTD algorithm

摘要

The co-processors with the capability of powerful floating-point operations have been used to study thernelectromagnetic (EM) simulations with the finite-difference time-domain (FDTD) method, which usuallyrnconsumes huge computational resources. This work studied the implementation and optimization of thernthree-dimensional FDTD algorithm with the uniaxial perfectly matched layer boundary conditions on GPUrnclusters. In this study, CPUs and GPUs collaborate to solve the different parts of the computation domainrnconcurrently, and a set of techniques are used to optimize the execution efficiency of GPUs, such as the usernof GPU texture memory, the pinned host memory, and the asynchronization of data transfer between CPUrnand GPU. For the CPU-GPU collaborative mode, the problem size solved is not limited by the size of globalrnmemory on GPU because the computational capability of CPU is also used for the calculation of EM field.rnTo keep the load balance between CPU and GPU, 85% of the computation load is assigned to GPU forrnNVIDIA Tesla K40m, and the collaborative efficiency is found to be less than 75%.