Enabling Energy-Efficient Kubernetes Cluster Autoscaling

Cloud computing offers unprecedented advantages in terms of scalability, reliability, and rapid resource provisioning. However, the underlying servers waste a significant amount of energy just to remain powered on. Therefore, shutting down underutilized servers can significantly reduce operational costs. This thesis proposes DREEM, a Kubernetes-based cluster scaling mechanism to predict future workloads and make intelligent decisions about powering servers on or off. DREEM optimizes the overall energy consumption while ensuring optimal performance and low latency. In this thesis, DREEM's architecture and implementation is discussed. In addition, its effectiveness is showcased by monitoring a small cluster and dynamically scaling it based on the measured and forecasted load. Then, it is compared againts Cluster Autoscaler, an already existing solution for scaling Kubernetes clusters. The demonstration highlights how this new approach effectively reduces the average number of active nodes, thereby lowering energy consumption without compromising performance or responsiveness.