Accounting for Variability in Large-Scale Cluster Power Models

Presented in March, 2011 at the Exascale Evaluation and Research Techniques Workshop (EXERT), held at the International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS).Studying the energy efficiency of large-scale computer systems requires models of the relationship between resource utilization and power consumption. Prior work on power modeling assumes that models built for a single node will scale to larger groups of machines. However, we find that inter-node variability in homogeneous clusters leads to substantially different single-node models. Furthermore, these models have much higher error when scaled to the cluster level than models built using multiple nodes. We report on inter-node variation for model feature selection and model training for four homogeneous five-node clusters using embedded, laptop, desktop, and server processors. These results demonstrate the need to sample multiple machines in order to produce accurate cluster models. Furthermore, we determine the necessary sample size for the machines and applications in this study by applying a theoretical worst-case error bound based on the mean power interval across the cluster