Prof. Haiying Shen gave a talk at Google this past week entitled:
“Optimizing Virtual Resource Management in Cloud Systems”
Cloud service providers maximize energy efficiency by operating the equipment in their data centers at high levels of utilization. But these providers must also satisfy Service Level Objectives (SLOs) for their tenants, which complicates resource provisioning. Existing work on improving resource utilization in data centers focuses mainly on Virtual Machine (VM) consolidation. As data centers are often oversubscribed, resources such as CPU and bandwidth are stretched thin as they are shared across many tenants. In particular, when VMs with intense resource requirements are located on the same physical machine, they compete for scarce resources, which may lead to poor performance and violations of SLOs. Much effort has been devoted to developing strategies for resource provisioning in the initial VM allocation and VM migration phases. Previous methods, however, neglect to take a number of subtle factors into consideration, which limits their effectiveness in practice. For example, in predicting resource demands, previous methods assume that applications exhibit certain demand patterns which may not be accurate, and they neglect misalignment between the utilization curves of VMs running the same application. These issues may lead to either resource underprovisioning or overprovisioning. Previous methods also neglect that Last Level Cache contention exists between VMs on the same physical machine, which degrades VM performance. In this presentation, I will introduce approaches that address the subtleties described above. These approaches should contribute to higher resource utilization, higher profit for cloud providers, and yet at the same time better quality-of-service for cloud users.