Decentralized and Energy-Efficient Workload Management in Enterprise Clouds

We present a decentralized approach towards scalable and energy-efficient management of virtual machine (VM) instances that are provisioned by large, enterprise clouds. In our approach, the computation resources of the data center are effectively organized into a hypercube structure. The hypercube seamlessly scales up and down as resources are either added or removed in response to changes in the number of provisioned VM instances. Without supervision from any central components, each compute node operates autonomously and manages its own workload by applying a set of distributed load balancing rules and algorithms. On one hand, underutilized nodes attempt to shift their workload to their hypercube neighbors and switch off. On the other, overutilized nodes attempt to migrate a subset of their VM instances so as to reduce their power consumption and prevent degradation of their own resources, which in turn may lead to SLA violations. In both cases, the compute nodes in our approach do not overload their counterparts in order to improve their own energy footprint. An evaluation and comparative study of the proposed approach provides evidence of its merits in terms of elasticity, energy efficiency, and scalability, as well as of its feasibility in the presence of high workload rates.