Modeling and reducing power consumption in large IT systems

This paper proposes a new energy optimization methodology using the Potluck Problem concept for large and complex IT systems. The proposed methodology involves analyzing the historical time-series data about resource utilization in the IT system, and finding interesting temporal patterns therefrom. Based on the analysis, a system of predictors is constructed wherein each predictor provides an estimate of demand for resources in the IT system for a future time period. These estimates are then used in the Potluck Problem solution strategy to obtain final demand predictions. This approach has been tested for efficacy using data from an actual large IT system. Before-and-after comparison is done for computing energy savings had the resources in the IT system been provisioned in accordance with our approach, in comparison with what actually transpired. Following these calculations, policies are designed with an aim to reduce the aggregate power consumption of the IT system.