The recent resurgence of research in server virtualization has created a lot of interest among providers of large-scale data centers in employing this technology to design improved solutions for managed hosting. Virtualization holds the promise of increased degrees of resource consolidation with accompanying reduction in operational costs of administration, repair, and electricity. Designing efficient virtualized data centers is therefore a desirable and worthy endeavor in multiple ways. The transition from traditional hosting models to a virtualized model is, however, not a trivial one. Realizing the consolidation-related benefits that virtualization has to offer requires a re-consideration of: (i) schedulers within the VMM, (ii) mechanisms for resource usage monitoring and accounting, and (iii) system-wide dynamic resource provisioning mechanisms. This proposal identifies the key challenges that arise on all these fronts in a virtualized data center, and develops a comprehensive solution, called River, to address them.
River will make the following specific contributions: (i) improved resource management algorithms within the virtual machine monitor (VMM) that use both horizontal (between the schedulers in the VMM) and vertical (between VMM and hosted OS schedulers) co-operation, (ii) efficient and stable dynamic resource provisioning, and (iii) systems primitives that would provide new opportunities to the provisioning algorithms for improved cost-cutting via reduction of virtualization overheads. A prototype River data center based on Xen will be developed and resulting code made available to the community. Finally, we will re-design the labs associated with the systems courses at Penn State to employ virtualized hosting of class projects.