The energy consumption of data centers is known to be a critical issue for both costs and environmental impact. Several approaches have been suggested to minimize energy consumption. However, most of them lack a precise understanding of the actual drivers of energy consumption and use either simplistic energy models or other simpler metrics as proxies for energy consumption (for example, minimizing the number of virtual machine migrations to minimize their energy impact). This may lead to suboptimal decisions, though: for example, without knowing the exact energy consumption that comes along with a virtual machine migration, it is unclear whether a planned migration will pay off.

In this paper, the authors elaborate very detailed energy models for two processes: data transfers and virtual machine migrations. The two models are important contributions on their own, but they are connected because the model for migrations uses the model for data transfers.

Both models are developed with a combination of the engineer’s intuition to determine the parameters that potentially impact energy consumption and a series of experiments to quantify their actual impact. The energy model for data transfers includes the effects of the number of bytes transferred, the payload per packet, the number of concurrent connections, and the size of burst and throttle intervals. The energy model for virtual machine migrations is more sophisticated as it considers the different phases of the migration separately: initiation phase, transfer phase, and activation phase. Moreover, the authors consider both live migration and non-live migration. The model includes the effect of the central processing unit (CPU) load of the source and target servers, the CPU load of the migrating virtual machine, the bandwidth between the two servers, and, for live migration, the memory dirtying ratio of the migrating virtual machine.

The resulting energy models are expressed by a set of equations, which are mathematically not very complicated (consisting of linear expressions and fractions thereof), but still significantly more sophisticated than models used by previous works. The authors tune and evaluate their models through a series of well-designed and well-documented experiments. The error of the energy model of data transfers is at most nine percent, whereas the error of the energy model of migrations is at most 17 percent and compares favorably to similar models from previous works.