GPU Based Parallel Smoothing Of Seismic Tomography Models

Three-dimensional models of the velocity structure of the Earth\u27s crust are an important and relevant factor for several types of analyses across disciplines. Crustal velocity models are also commonly used to analyze and search for different materials of interest or to determine and differentiate many aspects of life on Earth during different eras. Seismic tomography techniques, both in two and three dimensions, perform image reconstruction of the crust of the Earth. Seismic tomography algorithms can calculate crustal velocity structure through inversion of traveltimes of seismic waves produced by natural events, such as earthquakes, or controlled source experiments, such as explosions. The work presented in this thesis is based on the utilization of tomographic data sets produced through controlled-source experiments and the application of an iterative first-arrival traveltime seismic tomography algorithm ( STA ) to obtain 3D velocity models of specific regions of the crust of the Earth. The research focus of this thesis is to identify and exploit potentially parallelizable functions of smoothing algorithms, which are the STA performance bottleneck. Parallel tasks are then mapped to the architecture of a graphics processing unit ( GPU ) to accelerate the smoothing execution speed while maintaining the consistency and reliability of the outputs with respect to models considered as correct reference outputs. The implemented parallel STA smoothing algorithms deliver peak performance improvements from 31.9% to 73.1% and average improvements from 20.9% to 66.4% with respect to the fastest sequential implementations of the algorithms