Enhancing virtual environments with QoS aware resource management

Abstract

Nowadays, the consolidation of application servers is the most common use for current virtualization solutions. Each application server takes the form of a virtual machine (VM) that can be hosted into one physical machine. In a default Xen implementation, the scheduler is configured to handle equally all of the VMs that run on a single machine. As a consequence, the scheduler shares equally all of the available physical CPU resources among the running VMs. However, when the applications that run in the VM dynamically change their resource requirements, a different solution is needed. Furthermore, if the resource usage is associated with service-level agreements, a predefined equal share of the processor power is insufficient for the VMs. Within the Xen’s primitives, even though it is possible to tune the scheduler parameters, there is no tool to achieve the dynamic change of the share of the processor power assigned to each VM. A combination of a number of primitives, however, appears to be suited as a base for achieving this. In this paper, we present an approach to efficiently manage the quality of service (QoS) of virtualized resources in multicore machines. We evaluate different alternatives within Xen for building an enhanced management of virtual CPU resources. We compare these alternatives in terms of performance, flexibility, and ease of use. We devise an architecture to build a high-level service that combines interdomain communication mechanisms with monitoring and control primitives for local resource management. We achieve this by our solution, a local resource manager (LRM), which adjusts the resources needed by each VM according to an agreed QoS. The LRM has been implemented as a prototype and deployed on Xen-virtualized machines. By means of experiments, we show that the implemented management component can meet the service-level objectives even under dynamic conditions by adapting the resources assigned to the virtualized machines according to demand. With the LRM, we therefore achieve both fine-grain resource allocation and efficient assignment.