Vortragender: Thomas Treutner
Schedule: 1, 16:30
In this talk, we present MetaVisor, a prototypical framework based on KVM and libvirt that automatically and adaptively manages a distributed system of hypervisors and their VMs under realistic assumptions for a business scenario. The goal is to minimize energy consumption while avoiding under-provisioning. We show that such a visionary framework is indeed possible and has great potential to increase energy efficiency of servers.
The vast majority of servers in classical data centers are underutilized for a significant amount of time. These servers operate at a very low rate of efficiency and waste huge amounts of energy. To find a remedy, modern data centers frequently use system virtualization to consolidate a set of idle to a smaller set of highly utilized servers. Nevertheless, even modern data centers which currently use system virtualization mostly do not fully profit from its key benefits - dynamic resource allocation and live migration - due to a manifold lack of automation:
1. The VMs’ resource allocations are not automatically adapted to its given or expected resource requirements. 2. The VMs’ distribution across the RMs is not automatically adapted to the given or expected workload. 3. The RMs are not automatically powered down and woken up with regard to the resource demand.
MetaVisor is the effort to design and implement such an intelligent system with the power of open source tools and libraries, for example libvirt, a powerful library to manage the most popular hypervisors (Xen, KVM, VMware, HyperV, etc.) and their VMs. Different algorithms for solving the involved optimization problem (Greedy, Genetic) and results from simulations (maybe experimental results too) will be presented and discussed.
Any kind of feedback will be sincerely welcome.
Related work: VMware DPM/DRI, Entropy
File: metavisor.pdf
File: folien.pdf