Don’t hurry be green: scheduling servers shutdown in grid computing with deep reinforcement learning

Abstract

Grid computing platforms dissipate massive amounts of energy. Energy efficiency, therefore, is an essential requirement that directly affects its sustainability. Resource management systems deploy rule-based approaches to mitigate this cost. However, these strategies do not consider the patterns of the workloads being executed. In this context, we demonstrate how a solution based on Deep Reinforcement Learning is used to formulate an adaptive power-efficient policy. Specifically, we implement an off-reservation approach to overcome the disadvantages of an aggressive shutdown policy and minimise the frequency of shutdown events. Through simulation, we train the algorithm and evaluate it against commonly used shutdown policies using real traces from GRID’5000. Based on the experiments, we observed a reduction of 46% on the averaged energy waste with an equivalent frequency of shutdown events compared to a soft shutdown policy.