Estimating energy consumption during live migration of virtual machines

Abstract

The workload coming to the server clusters varies over the course of the day which results in disproportional resource utilisation. Server consolidation is a technique exploited by cloud service providers in order to optimally utilise available physical resources. During the time intervals of low resource usage the workload is consolidated from a big set of un-derutilised servers to a smaller set of optimally loaded servers. Afterwards idle servers are turned off which reduces energy consumption as well as running costs. Server consolidation can be realised via live migration of virtual machines (VMs). Though, the VMs' migration process requires additional CPU cycles and introduces energy overhead. The quantitative knowledge of the energy consumption during migration is important in order to realise sophisticated migration decisions. This paper presents how the energy consumption of the servers during migration can be estimated based on their resource utilisation parameters using linear regression techniques. We defined the most significant parameters that influence the energy consumption of the servers during migration. These are: CPU instructions retired, last level cache line misses, and "dirty" pages observed in the source server during migration.