Um estudo sobre a otimização de aplicações paralelas através de aprendizado de máquina

Abstract

Finding the number of threads that delivers a satisfactory outcome to an application is not a trivial task because of there are many variables involved, such as the application input set, hierarchical memory organization, and the processor microarchitecture. However, if an application has similar behavior when it is executed with different input sets, one can use a small input size to predict its behavior when a larger input set is considered. In this sense, the objective of this work is to use the data collected through the execution of a parallel application with a small input set to estimate its behavior when executing with a larger input set. From this estimation, we intend to find the degree of parallelism exploitation at the thread level that provides the best performance, energy consumption and/or Energy- Delay Product (EDP) to execute the application with the large input set. Through the execution of seventeen benchmarks well-known and widely used by the academic community on an AMD processor with 16-core, this work shows that it is possible to use the behavior of the execution with a small input set to optimize execution with a large input set. Considering the geometric mean of all applications, the results show a reduction of 9% in energy consumption and 10% in the EDP when compared to execution with the number of standard threads of the system. The work also shows that the proposed optimization method has an average difference of only 2% in energy consumption and 1% in EDP with respect to an oracle solution, which executes the application with the best number of threads.