Hórus: reduzindo o custo de aprendizado das estratégias de otimização offline para aplicações paralelas

Abstract

Several online and offline optimization strategies have been employed to improve the performance and energy consumption of parallel applications. While the first strategy can achieve some behaviors that can only be known at run time, the last one does not impose any overhead during execution and can use more complex and efficient algorithms. However, the offline strategy learning algorithm can take several hours, preventing its use or smooth portability between different systems. In this scenario, this dissertation proposes a methodology to decrease the learning time of offline strategies, inferring the execution behavior of parallel applications using smaller input sets than those used by the target applications. Two optimization strategies are implemented: SEA, where all the parallel regions of an application are executed with the same number of threads; and SPRA, which seeks to find an ideal number of threads for each parallel region of a given application. With an extensive set of experiments, we show that the SEA and SPRA strategies converge to results close to an offline approach applied to regular entry, but being 88% and 87% faster, on average, respectively. We have also shown that SPRA is better than SEA for unbalanced applications. Key-words: Parallel Computing. Runtime Optimization Systems. Thread-level Parallelism Exploitation. Energy-delay Product.