(RP04) GPU-based Parallel PO-SWE Algorithm for the Design of Large-sized Dual-Reflector Antennas
Performance Analysis and Optimization
TimeTuesday, June 26th8:30am - 10am
DescriptionLarge-sized dual-reflector antennas are analyzed to evaluate their performance through long-time simulation using some accurate EM (electromagnetic) methods based on PO (physical optics). To reduce the time/term for the simulation/design, we focused on the parallelization of the fast and accurate hybrid PO-SWE (spherical waveguide expansion) method, which relaxes the computational complexity while maintaining accuracy on a single-node CPU/GPU-embedded system, where each GPU has several thousand cores on a discrete small-sized memory. To perform the hybrid PO-SWE on such a small-sized memory, we developed the CPUs-GPUs distributed algorithm, through which the CPUs iteratively compute the new small-sized regions of the main-reflector and the SWE coefficients of the sub-reflector behind the GPUs-based PO-SWE computation for the pre-computed them. We found that our implementation using 4 GPUs is up to 37 times faster than that using 24-threaded 2 CPUs.