Dynamic Management of CPU Resources Towards Energy Efficient and Profitable Datacentre Operation.

Energy reduction has become a necessity for modern data-centres, with CPU being a key contributor to the energy consumption of nodes. Increasing the utilization of CPU resources on active nodes is a key step towards energy efficiency. However, this is a challenging undertaking, as the workload can vary significantly among the nodes and over time, exposing operators to the risk of overcommitting the CPU. In this paper, we explore the trade-off between energy efficiency and node overloads, to drive virtual machine (VM) consolidation in a cost-aware manner. We introduce a model that uses runtime information to estimate the target utilization of the nodes to control their load, identifying and considering correlated behavior among collocated workloads. Moreover, we introduce a VM allocation and node management policy that exploits the model to increase the profit of datacentre operators considering the trade-off between energy reduction and potential SLA violation costs. We evaluate our work through simulations using node profiles derived from real machines and workloads from real data-centre traces. The results show that our policy adapts the nodes' target utilization in a highly effective way, converging to a target utilization that is statically optimal for the workload at hand. Moreover, we show that our policy closely matches, or even outperforms two state-of-the-art policies that combine VM consolidation with VFS - the second one, also operating the CPU at reduced voltage margins - even when these are configured to use a static, workload- and architecture-specific target utilization derived through offline characterization of the workload.