high performance computing on graphics processing units: hgpu.org

The Finite-Difference Time-Domain (FDTD) solution of Maxwell’s equations, a grid-based differential time-domain numerical modeling method, is an approach for the direct modelling of the penetration of structures by continuous plane waves. Although FDTD techniques are considered to be relatively easy to understand and to implement in software, such modelling methods require a high level of computational resources and long execution times. Both high computational costs and simulation times are desired to decrease. The highly parallel architecture of GPUs makes these devices to be more effective than general-purpose CPUs for a range of complex algorithms that are rich in data parallelism. Today, parallel GPUs have begun making computational inroads against the CPU. For certain classes of applications that require massive vector operations, GPUs can yield several orders of magnitude higher performance than conventional CPUs. Being FDTD a grid-based differential time-domain numerical modeling method, the implementation of such technique can leverage from GPUs revolutionary capability to operate on large matrices in parallel, while still making use of the CPU when appropriate. This thesis aims to explore the viability of using GPUs to simulate electromagnetic wave propagation. For that purpose, a "one-click" application has been implemented, which runs a complete simulation of the behavior of an electromagnetic wave being propagated through an arbitrary media. The FDTD method has been determined as the best simulation method to be used for our purpose. The FDTD method used in this thesis considers the peculiarities of GPUs, which differ from those of CPUs. The application will take two inputs: a volume describing the media where the electromagnetic wave will be propagated, and a description of the electromagnetic source. The media volume can define any arbitrary space. The source can be a carrier signal or a modulated pulse signal. The output of the simulation will be a series of data files containing snapshots of the simulated media at different instants of time. These data files can be easily used in scientific data analysis and visualization applications to further analyze the simulated media. The correctness of the implementation has been verified by simulating the behavior of electromagnetic waves in a rectangular waveguide. The results obtained from the FDTD simulation have been compared against the results obtained from the analytical solution. The experimental evaluation points out that the results of the FDTD-based simulator are reliable and accurate enough. Experimental results of the simulation’s execution time suggest that GPUs offer high performance. Moreover, the experience of this thesis confirms that GPUs are well suited to deal with algorithms rich in data parallelism.