HyPPO: Hybrid Performance-Aware Power-Capping Orchestrator

Energy proportionality is the key in order to reduce the Total Cost of Ownership (TCO) of data-centers and on-premise systems, yet is difficult to achieve in practice due to workloads heterogeneity. Critical services require all the servers to remain up regardless the current traffic intensity in order to scale up quickly if needed. Furthermore, a minimum number of resources remain reserved to such services in order to guarantee the Service Level Agreement (SLA) in case of a sudden increase of requests. This situation makes most of the existing power management techniques ineffective at reducing power consumption, especially in under-utilized computing systems. In this paper we propose Hybrid Performance-aware Power-capping Orchestrator (HyPPO), a distributed Observe Decide Act (ODA) control loop able to introduce energy proportionality in a distributed and containerized infrastructure. HyPPO leverages Kubernetes resource requests (e.g. milli-cpus consumption) in order to dynamically adjust node power consumption, while respecting the SLA defined by the containerized application owners. The first experiments conducted in our laboratory shows that we are able to reduce power consumption of 25% on average, depending on the kind of workload. Furthermore, the defined SLA is violated 2.5% of the time on average