Modelling the Live Migration Time of Virtual Machines

Dynamic server consolidation in data centres enables the efficient usage of resources, because it aims to minimise the underutilisation or overloading of physical servers, both of which produce a disproportional amount of energy consumption. Server consolidation takes place by migrating virtual machines from one server to another while the virtual machines are still executing. However, live migration comes with corresponding costs in terms of execution latency and additional resource and power consumption. Whether or not these costs are significant depends on how long a migration lasts. In this paper we propose models to estimate the time it takes to live migrate virtual machines at runtime. Our models are built using simple and multiple linear regressions. The paper reveals useful insights into the most important parameters which are strongly correlated with the migration time. These are: Instructions retired, last level cache line misses, and dirtying memory pages.