Enabling Energy-Efficient Kubernetes Cluster Autoscaler

Energy waste in underutilized bare-metal clusters remains a critical challenge for cloud providers. Powering off such servers could significantly reduce operational costs. However, selecting the best server to be powered off is a complex optimization problem; incorrect selection risks resource saturation and compromised service availability. This work presents DREEM, a Kubernetes-based autoscaling mechanism specifically designed for bare-metal infrastructures. DREEM integrates a server selection algorithm that reduces overall energy consumption while preserving application performance, according to user-defined preferences. We showcase DREEM operating on a small cluster, dynamically adapting its size to the running workload. The results highlight how DREEM effectively reduces the average number of active nodes, lowering energy consumption without compromising performance or responsiveness, while consistently selecting near-optimal servers according to the defined constraints.