We presented our workshop paper “How Accurately Can the Energy Use of Spark Applications Be Estimated Based on Resource Utilisation?” at the PECS 2026 workshop (co-located with Euro-Par 2026) in Pisa, Italy on August 24, 2026.

The paper investigates model-based energy estimation for Apache Spark running on Kubernetes by comparing resource-utilisation-based estimates to direct Intel RAPL package and DRAM energy measurements. The experiments assess estimation accuracy for several commonly used Spark applications on both an AWS bare-metal cloud and an on-premises cluster. The results show that incorporating external monitoring signals reduces the energy estimation error for the package domain relative to Spark task traces, improving underestimation from -29.58% to -24.41% on AWS and from -24.00% to -16.22% on-premises.

The artifacts archive contains raw experimental results and the prototype code used for the experiments.