Controle do consumo de energia em data centers via aprendizado por reforço

Abstract

Nowadays the high demand for Internet services creates the need for systems with parallel processing and a large number of machines. Often, these systems operate with more servers than needed to ensure an appropriate service at moments of peak demand. However, this generates enormous energy consumption which could be avoided. This work proposes a method to control the energy consumption using reinforcement learning, aiming to find a better way to minimize the total cost while maintaining quality of service. The system considered works with one block of servers, called regular, which is always turned on and one or more blocks, called reserves, which can be turned on or off. The Q-learning algorithm is used to determine the best moment to turn on or off the reserve blocks. A simulator was developed to test the proposed method.