Eficiencia energética en autoescalado predictivo de Kubernetes: evaluación experimental con Kepler y RAPL

Abstract

Data centers account for a relevant share of global electricity consumption, and demand keeps increasing as cloud services and artificial intelligence workloads expand. In this context, Kubernetes autoscaling decisions affect not only latency and availability, but also the energy required to serve each request. This paper evaluates five scaling strategies —fixed provisioning, HPA at 50% CPU, HPA at 70% CPU, Long Short-Term Memory (LSTM)-based predictive scaling, and a predictive-reactive hybrid strategy— measuring joules per request (J/req) with Kepler and Running Average Power Limit (RAPL) counters. These measurements are analyzed together with P95 latency, service-level violations, and pod-seconds. Based on 81 valid experiments and a non-parametric statistical protocol, the predictive strategy reduces energy consumption by 14.7% compared with fixed provisioning and cuts pod usage by 49%. The hybrid strategy is the most suitable option for production: with only 2.2% additional energy compared with the reactive scaler, it reduces service-level violations from 40.47% to 9.95%.