Adaptive data center activation with user request prediction

The problem of energy saving in data centers has recently attracted significant interest within the research community, and the adaptive data center activation model has emerged as a promising technique to save energy. However, this model has not integrated adaptive activation of switches and hosts in data centers because of its complexity. This paper proposes an adaptive data center activation model that consolidates adaptive activation of switches and hosts simultaneously integrated with a statistical request prediction algorithm. The learning algorithm predicts user requests in a predetermined interval by using a cyclic window learning algorithm. Then the data center activates an optimal number of switches and hosts in order to minimize power consumption that is based on prediction. We designed an adaptive data center activation model by using a cognitive cycle composed of three steps: data collection, prediction, and activation. In the request prediction step, the prediction algorithm forecasts a Poisson distribution parameter λ in every predetermined interval by using Maximum Likelihood Estimation (MLE) and Local Linear Regression (LLR) methods. Then, adaptive activation of the data center is implemented with the predicted parameter in every interval. The adaptive activation model is formulated as a Mixed Integer Linear Programming (MILP) model. Switches and hosts are modeled as M/M/1 and M/M/c queues. In order to minimize power consumption of data centers, the model minimizes the number of activated switches, hosts, and memory modules while guaranteeing Quality of Service (QoS). Since the problem is NP-hard, we use the Simulated Annealing algorithm to solve the model. We employ Google cluster trace data to simulate our prediction model. Then, the predicted data is employed to test the adaptive activation model and observe energy saving rate in every interval. In the experiment, we could observe that the adaptive activation model saves 30-50% of energy compared to the full operation state of data centers in practical operating conditions of data centers.