Understanding Energy Consumption Trends in High Performance Computing Nodes

This article presents a study of energy consumption behavior in high performance computing nodes in relation to the usage of computing resources. Linear models are constructed for identifying common patterns or differences in energy consumption across different architectures. The study is significant as it provides insights into the energy consumption of computing nodes, helping to build simple, yet useful and transparent models. Moreover, the methodology provides building blocks for building new models with high predictive quality and broad applicability across different architectures. The results reveal similarities in energy consumption across different architectures when compared in terms of CPU cycles and cache misses. Additionally, employing linear models based on CPU cycles and cache misses allows for both the explanation of energy consumption behavior and the achievement a reasonable predictive quality. Overall, a partition-based linear model outperforms a global linear and a mean partition-based model by up to 5.7%.