Prediction of Future Loads Using Neural Networks for Energy-Efficient Computing

Abstract

In modern data centers a large amount of energy can be saved by intelligently distributing load on the available servers and transferring idle nodes into low energy modes. Distributing load leads to a more energy-efficient usage of the servers within a server farm. Additionally, the use of energy saving modes like suspend to main memory can decrease the energy consumption dramatically. The selection of nodes to be transferred into a low energy mode is based on the information of an energy-efficient load distribution. The usage of low energy modes requires knowledge about future loads. Having a variable load profile, i.e. variations in loads over time, leads to time periods in which servers are idle (denoted as gaps). Within these gaps, servers can be transferred into one of various supported energy saving modes. It is crucial to have information about future gaps in advance to make the right decision in regard to the chosen energy saving mode. Usually, information about the future is not directly available but can be predicted using sophisticated algorithms. In this paper, we present an approach to predict future loads using trends, seasonal data, and neural networks.