Billing the CPU Time Used by System Components on Behalf of VMs

Abstract

Nowadays, virtualization is present in almost all cloud infrastructures. In virtualized cloud, virtual machines (VMs) are the basis for allocating resources. A VM is launched with a fixed allocated computing capacity that should be strictly provided by the hosting system scheduler. Unfortunately, this allocated capacity is not always respected, due to mechanisms provided by the virtual machine monitoring system (also known as hypervisor). For instance, we observe that a significant amount of CPU is consumed by the underlying system components. This consumed CPU time is not only difficult to monitor, but also is not charged to VM capacities. Consequently, we have VMs using more computing capacities than the allocated values. Such a situation can lead to performance unpredictability for cloud clients, and resources waste for the cloud provider. In this paper, we present the design and evaluation of a mechanism which solves this issue. The proposed mechanism consists of estimating the CPU time consumed by the system component on behalf of individual VM. Subsequently, this estimated CPU time is charged to VM. We have implemented a prototype of the mechanism in Xen system. The prototype has been validated with extensive evaluations using reference benchmarks.