Towards Proactive Resource Management in Virtualized Datacenters

Proactive resource management involves taking actions to make resources available for upcoming load spikes. Proactive resource management is highly desirable for managing resources in a multi-tenant cloud where it is common to have hot-spots at some loca-tions, while still having spare resources scattered throughout the datacenter. Inability to quickly use fragmented spare resources in a datacenter during load spikes causes administrators to do host-level over-provisioning. Running hosts at low utilization leads to lower consolidation ratios directly impacting the bottom-line. In this paper, we present key challenges in doing proactive re-source management in virtualized environments. We argue that many of the building blocks necessary for proactive management are available today – specifically, (a) analytics engines that collect performance metrics and apply sophisticated statistical learning al-gorithms to predict various metrics, including future resource de-mands for hosts, virtual machines (VMs), storage arrays and (b) cluster-level schedulers that react to resource demands and suggest VM migrations for load-balancing. We propose a preliminary solu-tion that combines these building blocks, while handling the hard issue of dealing with mis-predictions. Our key insight is to conduct proactive management only using spare resources without hurting the ongoing workloads. We have validated some of our ideas using a simulation-based evaluation. We hope to inspire further research in this area, making cloud environments more efficient, nimble and adaptive. 1