PRESTO: a latency-aware power-capping orchestrator for cloud-native microservices

Power consumption is a major concern for cloud data-centers. In this context, cloud-native applications emerged in the last few years and fostered the adoption of the cloud computing model across many organizations. Cloud-native workloads are highly heterogeneous, co-located and latency-sensitive and are able to scale to a high number of machines. To properly manage their power consumption, within this paper we propose Power REgulator for Service Time Optimization (PRESTO), a latency-aware power-capping orchestrator. PRESTO defines an Observe Decide Act (ODA) loop to manage power consumption and average latency of microservice-based workloads by considering all the network interactions between microservices in the cluster. PRESTO reduces the power consumption by 37.13% on average with a control error that is below 12.5% and below 1.5ms on average w.r.t. an unconstrained execution.