Oversubscribing micro-clouds with energy-aware containers scheduling

Cloud computation is being pushed to the edge of the network, towards Micro-clouds, to promote more energy efficiency and less latency when compared to heavy resourced centralized datacenters. This trend will enable new markets and providers to fill the current gap. There are however challenges in this design: (i) devices have less resources, leading to a frequent use of oversubscription (ii) lack of economic incentives to both provider and application owner to cope with less than full requests fulfilled. To support this trend, the virtualization layer of Micro-clouds is currently dominated by containers, which have a small memory footprint and strong isolation properties. We propose an extension to Docker Swarm, a widely used containers orchestrator, with an oversubscribing scheduling algorithm, based on improving resources utilization to levels where the energy efficiency is maximized. This solution improves CPU and memory utilization over Spread and Binpack (Docker Swarm strategies). Although we introduce a small overhead in scheduling times, our solution manages to allocate more requests, with a successful allocation rate of 83% against 57% of current solutions, measured on the scheduling of real CPU- and memory-intensive workloads (e.g. Video encoding, Key-value storages and a Deep-learning algorithm).