Evaluation of non-linear power estimation models in a computing cluster

Abstract

The data center industry is responsible for 1.5–2% of the world energy consumption. Energy management technologies have been proposed for energy-efficient scheduling of computing workloads and for allocating resources in such computing infrastructures. One of the important factors for this energy management is the estimation of power consumption as a result of the workload schedule to be carried out. The commonly used power models are a linear function of resource features. Based on measurement data sets from our cluster, we extended power model study to multiple non-linear dependencies. We provide several novel contributions: we apply neural network models and unsupervised classification models with basic OS-reported resource features for power estimation; we build and test the power estimation models in a cluster environment from a large size of measurement data; we evaluate the power estimation models in terms of not only accuracy but also portability and usability.We prove that a multiple-variable linear regression approach is more precise than a CPU-only linear approach. The neural network approaches have a slight advantage – their mean root mean square error is at most 15% less than that of the multiple-variable linear model. The neural network models have worse portability when the models generated on a node are applied on other homogeneous nodes. Gaussian Mixture Model has the highest accuracy but requires the longest training time. In the end, we prove that models trained using the system-level full features have the highest accuracy comparing to only use part of features.