Paralelização do algoritmo de migração sísmica em plataformas heterogêneas

• Ano de defesa: 2010
• Autor(a) principal: Thiago Santos Faria Xavier Teixeira
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; UFMG
• Palavras-chave em Português: Ciência da Computação; Algoritmos de computador; Computação; Sistemas distribuídos; Algoritmos paralelos
• Link de acesso: http://hdl.handle.net/1843/SLSS-8B2HMZ

Oil is one of the most important and strategic energy resources in the world. The increasing oil demand and price now support its production where it was previously unfeasible. The most common strategy to nd potential reserves is through seismic methods, among which the Kirchho seismic migration is one of the most popular strategies. However, it is computationally intensive and improving its performance has been researched in several works, often through their parallelization.On the other hand, we are witnessing the emergence of novel parallel computing architectures where each processing node is organized as a hierarchy of several heterogeneous processors that may be multi-core or many-core, such as Graphical Processing Units (GPUs), used as a consequence of their computing power. Such architectures are becoming commonplace and represent a challenge with respect to design computationally ecient programs. In this thesis, we discuss the parallelization of the Kirchho seismic migration algorithm for execution on the aforementioned heterogeneous environment. We proposeand evaluate parallel implementations that use eciently the available processors. We exploited the trade os among three parallelism opportunities, namely asynchrony, data parallelism and task parallelism, by designing, implementing and evaluating various possible congurations. We also devised and implemented a ner grain dynamicscheduling for the devices, supporting executions that present higher eciency. The experiments showed an acceleration of 87 times compared to execution on a single CPU core. This was due to CPU and the GPU ecient use through the approaches presented.