Energy consumption and workload prediction for edge nodes in the Computing Continuum

Abstract

The Computing Continuum paradigm provides developers with a distributed infrastructure for deploying applications through the network, improving performance and energy consumption. However, to maintain applications’ efficiency, their deployment in the Computing Continuum has to be continuously adapted to the varying load of different nodes of the network. In practice, existing support frameworks allow developers to automatically identify how to deploy applications based on the infrastructure status. However, as the application takes time to be deployed, the chosen deployment is outdated once it is applied through the network, as workloads change over time. To address this practical engineering challenge and plan deployments that foresee changes in energy consumption and workload, predictive solutions are needed. To fulfill this need, this work presents the Microservice Energy consumption and Workload Forecaster (MEWF), a prediction system that leverages artificial intelligence techniques to precisely predict CPU usage and energy consumption in varying circumstances. Our practical evaluation over multiple real microservices shows that MEWF improves prediction precision by up to 55% w.r.t. state-of-the-art benchmarks, enabling efficient resource management and demonstrating significant value for real-world deployments.