Algoritmos de escalonamento para grades computacionais voltados à eficiência energética

Abstract

Recent advances in High Performance Computing have opened a wide range of new research opportunities. Modern parallel and distributed architectures present each time more and more processing units seeking for a higher computational power. At the same time, the gain of performance obtained with those platforms is followed by an increase in energy consumption. In this scenario, researches in energy efficient high performance environments have emerged as a way to find the causes of excessive energy consumption and propose alternative solutions. Nowadays, one of the most representative high performance platforms is the computational grid which is used in many scientific and academic projects all over the world. In this work, we propose the use of energy-aware scheduling algorithms to efficiently manage the energy consumption in computational grids trying to avoid excessive performance losses. Our solution is based on: (i) an efficient management of idle resources; (ii) a clever use of active resources; (iii) the development of a procedure to accurately estimate the energy consumed in a given platform; (iv) the proposal of several new energy-aware scheduling algorithms for computational grids. We evaluate our approach using the SimGrid simulation environment and we compared our algorithms against five traditional scheduling algorithms for computational grids that are not energy-aware and one new algorithm recently proposed in the literature that deals with energy consumption issues. Our results show that in some experimental scenarios using our algorithms it is possible to achieve up to 221,03% of reduction in the energy consumption combined with 34,60% of performance loss. This example confirms our assumption that it is possible to significantly decrease the energy consumption on a grid platform without compromising proportionally the performance.