Virtual machine priority adaption to enforce fairness among cloud users

Abstract

In recent years fairness problems in data centers have been pointed out and job/Virtual Machine (VM) scheduling has been chosen as a solution approach. Clouds are a special case of data centers, where resources are deployed by VMs in a highly dynamic manner during VM runtime. However, scheduling only allows influencing resource allocations, when VMs are instantiated, i.e., before runtime. Thus, runtime prioritization bears a great potential to manage cloud resources and promote fairness in clouds, especially, when VMs run over long periods. Nevertheless, runtime prioritization is not leveraged accordingly. This paper defines fairness as handicapping VMs of heavy users during runtime to allocate more resources to VMs of light users. Thereby, the need to make assumptions on user's utility functions is avoided, while different fairness notions can be captured by adapting the definition of heaviness. Guidelines for this definition are provided to ensure incentives to configure and utilize VMs adequately. Finally, OpenStack is extended in its implementation by a decentralized fairness service to enforce fairness according to this definition. The fairness service's functionality is certified by experiments in terms of overhead and fairness promotion.