Avaliação de estratégias de paralelismo em simulação de meios porosos

Abstract

High-performance computing provides for the acceleration of scientific applications th- rough the use of parallelism. Applications of this type usually demand a large amount of computing time for a version with a single code execution stream. There are different parallel programming models to indicate the concurrent execution of a code. The division in data decomposition processing is the best choice for scientific applications in general. This dissertation aims to evaluate different parallelism approaches for a porous media application. The application case study provides the simulation of grain drying. We im- plement four versions: OpenMP parallel do, target and teams, and OpenACC parallel do. The OpenMP Target and OpenACC versions required structural changes in the source code. Three mesh sizes were evaluated for each implemented version, varying the number of threads in 2, 4, 8, 16, and 32. The results obtained show that all versions showed performance gains. The best result for loop I was the Target 32 threads test case with speedup of 3.11. For mesh II the best result was the Teams 16 threads test case with speedup of 6.27. For mesh III, the highest speedup result was 7.54 for the OpenACC test case. We perform a statistical analysis of the results obtained with the T-Student distri- bution. The analysis validated the performance gain. Traces of parallel executions were also generated, which allowed, through the Vampir tool, to visualize the synchronizations among concurrent operations.