Today’s concern is about the energy efficiency of servers and high power machines in a cloud datacenter infrastructure. According to Barroso et al. [1], an ideal machine consumes energy proportional to the work performed. In this case, an idle machine should consume no energy and a machine in operation should only consume energy proportionally to the number of tasks performed. Even though the energy efficiency of machines is constantly improving, they are still not perfectly energy-proportional. Therefore, Dürr proposed the concept of Elastic Tandem Machines Instances (ETMI) in [2] aims to improve the energy efficiency in particular for idle and weakly loaded instances. In this thesis, we attempt to improve the concept of Elastic Tandem Machines. The original concept only integrated one low-power system on a chip (SoC) machine, which operates during low load on the datacenter, and exactly one high-power vir- tual machine(VM) instance, powered on when the traffic increases and needs to be redirected. However, if the performance of the SoC and the VM instance differed too much, the efficiency of the approach suffered since at the performance limit of the SoC, when the transferred occurred, the high-power would be almost idle. There- fore, we integrate different performance classes of VMs (e.g., small, medium, and large instances) into Elastic n-Instance Machines to further improve the efficiency and scalability of the system. We then design a predictive algorithm and integrate it with the ETMI to decide, in advance, when the best time is, before overloading any server, to switch among the instances. The handover algorithm, based on a software-defined networking and the pre- dictive algorithm, based on an Autoregressive Integrated Moving Average (ARIMA) model are presented. The performance of the system with respect to the energy efficiency and machine elasticity is evaluated using experiments and performance benchmarks. The evaluations of the model demonstrate the applicability of low and medium power instances serving low and medium loads efficiently, in addition to the scalability of the solution among n-instances. The predictive method shows satisfactory results when forecasting seasonal data, different models may have to be implemented for non-seasonal series.