Scalable Acceleration of The Characteristic Mode Analysis Computational Toolbox Using Big Data Techniques

Abstract

Characteristic Mode Analysis (CMA) is used in the design and analysis of a wide range of electromagnetic devices such as antennas and nano-structures. CMA provides physical insight into a target’s electromagnetic response by analyzing its Method of Moments (MoM) complex impedance matrix at every frequency. Depending on the size or the complexity of the target, its MoM impedance matrix can be large, containing tens of thousands of rows and columns. Therefore if hundreds of frequencies are needed to accurately quantify the electromagnetic response of a target, hundreds of gigabytes of RAM and tens of terabytes of storage are needed creating a classical Big Data problem. This dissertation addresses this problem by proposing two approaches: one using data-parallel techniques on a cluster of nodes, and the other using TensorFlow2.0 and the CUDA platform. Apache Spark and Apache Hadoop are used in the first approach to scale CMA over a wide range of impedance MoM matrices. The second approach exploits GPUs, as most matrix operations in CMA run faster on this hardware. Furthermore, this dissertation explores the use of hybrid computational platforms that combine multi-core CPUs with GPUs. To optimize CMA efficiency, we broke down its algorithm into basic matrix operations and then performed exhaustive computational experiments to study the optimum numerical library to execute each matrix operation. Following these experiments, we measured the execution time for each matrix operation involved in CMA implementation. The first approach improves the processing time for 51 matrices of 14k x 14k by 10 times. Moreover, the second approach achieved up to 16x to 26x speedup of the CMA processing of a single 15k x 15k MoM matrix of a perfect electric conductor scatterer and a single 30k x 30k MOM matrix of a dielectric scatterer, respectively. Furthermore, optimized CMA implementation can analyze matrices that are too large to be studied by other CMA hardware. Optimized CMA implementation will pave the way for electromagnetic engineers to advance the state of the art by exploring the electromagnetic response of new targets that are ~20 times larger or ~20 times more complex than what was studied before.