Escalonador adaptativo para laços paralelos em processadores multinúcleo assimétricos

Abstract

The growing demand for computing power and energy efficiency in mobile computing has triggered the development of heterogeneous processors with specialized cores for different types of computational tasks, such as ARM big.LITTLE processors, which have different cores that combine performance with low energy consumption. Such difference in the composition of the cores in this kind of processors ends up inducing an asymmetry in the computational performance of these systems, making complicated the task of predicting the behavior of parallel applications in relation to performance when using all their cores. This asymmetry can be detected in applications that use parallel loops, a parallel programming feature that allows to divide the workload of an iterative routine between the cores present in a processor. Parallel loop schedulers that are not designed to prevent loss of performance in asymmetric multi-core processors (AMPs) can compromise the implementation of software solutions designed for this type of architecture. This dissertation presents the implementation proposal of a scheduler for parallel loops that uses an adaptive algorithm to distribute the workload among the threads, aiming at a better extraction of performance in AMPs. The scheduler uses of parallel work-stealing and lock-free as possible sequential extraction of work to face other existing solutions. Its implementation was carried out in the C ++ language, with the possibility of portability to the C language. In order to evaluate the performance of the solution, an analysis was performed on the set of NAS benchmarks and four distinct well-established scientific applications in related literature, over two real asymmetric embedded environments, against two existing solutions (OpenMP and Intel TBB). The analysis shows that the scheduler is able to extract more performance in certain cases and is very close to the best solutions in most of the remaining cases, with greater scalability potential in theory for cases where the scheduling overhead becomes an obstacle in the other solutions.