Lightweight Energy Consumption Models for High Performance Computing

This article presents the development of energy consumption models for high-performance computing nodes, aiming to reduce construction overhead while maintaining estimation accuracy. Based on a piecewise linear model derived from performance counters, several modeling strategies were proposed, including exact and heuristic approaches, which exploit specific knowledge of energy consumption behavior of high-performance computing to enable the reuse of submodels across different domain regions. The models were implemented and evaluated on four high-end computing nodes to identify common consumption patterns, supporting the generation of strategies applicable across architectures. Results indicate that reusing submodels have the potential of reducing the training domain fourfold, with an average error difference of 0.58 compared to the full-domain model. Furthermore, the strategy in which training regions are predefined in a node-agnostic manner achieves an average error difference of 0.7 while halving the training domain.