Granularity load scheduling auto-tuning for multi-core processors applied to reverse-time migration

• Ano de defesa: 2020
• Autor(a) principal: Fernandes, João Batista
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Rio Grande do Norte; Brasil; UFRN; PROGRAMA DE PÓS-GRADUAÇÃO EM ENGENHARIA ELÉTRICA E DE COMPUTAÇÃO
• Palavras-chave em Português: Auto ajuste; RTM; CSA; HPC; Chunk Size; Memória compartilhada
• Link de acesso: https://repositorio.ufrn.br/handle/123456789/30251

Reverse-time migration (RTM) is an algorithm widely used in the oil and gas industry to process seismic data. It is a computationally intensive task that can be designed to run in parallel computers. Because of it being massive and regular, this type of task is often equally and statically distributed among the available parallel processors. However, this strategy might often not be optimal. When the processors are heterogeneous, and even when most have similar processing power, many of them might still have to wait idly for the slower processors. In this paper, we show that even among homogeneous cores here might be load imbalance that can considerably affect the overall performance of a 3D RTM application. We show that dynamic load distribution has a significant advantage over the conventional static distribution, and other default OpenMP schedules, such as auto and guided. However, the granularity of the dynamically distributed chunks of work plays a key role in harvesting this advantage. In order to find the optimal granularity, we propose a coupled simulated annealing (CSA) based auto-tuning strategy that adjusts the chunk size of the work that OpenMP parallel loops assign dynamically to worker threads during the initialization of a 3D RTM application. Experiments performed on computational systems with different processor and memory specifications for different sizes of input show that the proposed method is consistently faster than the default OpenMP loop schedulers.