Optimising for energy or robustness? Trade-offs for VM consolidation in virtualized datacenters under uncertainty

Abstract

Reducing the energy consumption of virtualized datacenters and the Cloud is very important in order to lower CO 2 footprint and operational cost of a Cloud operator. However, there is a trade-off between energy consumption and perceived application performance. In order to save energy, Cloud operators want to consolidate as many Virtual Machines (VM) on the fewest possible physical servers, possibly involving overbooking of resources. However, that may involve SLA violations when many VMs run on peak load. Such consolidation is typically done using VM migration techniques, which stress the network. As a consequence, it is important to find the right balance between the energy consumption and the number of migrations to perform. Unfortunately, the resources that a VM requires are not precisely known in advance, which makes it very difficult to optimise the VM migration schedule. In this paper, we therefore propose a novel approach based on the theory of Robust Optimisation (RO). We model the VM consolidation problem as a robust Mixed Integer Linear Program and allow to specify bounds for e.g. resource requirements of the VMs. We show that, by using our model, Cloud operators can effectively trade-off uncertainty of resource requirements with total energy consumption. Also, our model allows us to quantify the price of the robustness in terms of energy saving against resource requirement violations.

On the energy cost of robustness for green virtual network function placement in 5G virtualized infrastructures

“…Authors in [15] show how the emerging area of robust optimization can advance the network planning by a more accurate mathematical description of the demand uncertainty. This concept is applied in [16], where the VM consolidation problem is modeled as a robust MILP and the resource requirements of the VMs are allowed to variate between 120 specific bounds. The price of the robustness is quantified in terms of energy saving against resource requirement violations.…”

A fast robust optimization-based heuristic for the deployment of green virtual network functions

“…in terms of CPU utilization of the virtual components or network latency), and the well-discussed price of robustness [3,6,8,9] due to higher cost (e.g. energy consumption) required to protect from parameter deviations [10].…”