Assessing and forecasting energy efficiency on Cloud computing platforms

Abstract

IaaS providers have become interested in optimising their infrastructure energy efficiency. To do so, their VM placement algorithms need to know the current and future energy efficiency at different levels (Virtual Machine, node, infrastructure and service levels) and for potential actions such as service deployment or VM deployment, migration or cancellation. This publication provides a mathematical formulation for the previous aspects, as well as the design of a CPU utilisation estimator used to calculate the aforementioned forecasts. The correct adjustment of the estimators’ configuration parameters has been proved to lead to considerable precision improvements. When running Web workloads, estimators focused on noise filtering provide the best precision even if they react slowly to changes, whereas reactive predictors are desirable for batch workloads. Furthermore, the precision when running batch workloads partially depends on each execution. Finally, it has been observed that the forecasts precision degradation as such forecasts are performed for a longer time period in the future is smaller when running web workloads.