Evolutionary neural network modeling for energy prediction of cloud data centers

Accurate forecasts of data center energy consumptions can help eliminate risks caused by underprovisioning or waste caused by over-provisioning. However, due to nonlinearity and complexity, energy prediction remains a challenge. An added layer of complexity further comes from dynamically changing workloads. There is a lack of physical principle based clear-box models, and existing black-box based methods such neural networks are restrictive. In this paper, we develop an evolutionary neural network as a structurally optimal black-box model to forecast the energy consumption of a dynamic cloud data center. In particular, the approach to evolving an optimal network is developed from several novel mechanisms of a genetic algorithm, such as a structurally-inclusive matrix encoding and species parallelism that help maintain an overall increasing fitness to overcome slow convergence whilst preventing premature dominance. The model is trained using part of the data obtained from a set of MapReduce jobs on a 120-core Hadoop cluster and is then validated against unseen data. The results, both in terms of prediction speed and accuracy, suggest that this evolutionary neural network approach to cloud data center forecast is highly promising.