Customer-aware resource overallocation to improve energy efficiency in realtime Cloud Computing data centers

摘要

Energy efficiency is becoming a very important concern for Cloud Computing environments. These are normally composed of large and power consuming data centers to provide the required elasticity and scalability to their customers. In this context, many efforts have been developed to balance the loads at host level. However, determining how to maximize the resources utilization at Virtual Machine (VM) level still remains as a big challenge. This is mainly driven by very dynamic workload behaviors and a wide variety of customers' resource utilization patterns. This paper introduces a dynamic resource provisioning mechanism to overallocate the capacity of real-time Cloud data centers based on customer utilization patterns. Furthermore, its impact on the trade-off between energy efficiency and SLA fulfillment is analyzed. The main idea is to exploit the resource utilization patterns of each customer to decrease the waste produced by resource request overestimations. This creates the opportunity to allocate additional VMs in the same host incrementing its energy efficiency. Nevertheless, this also increases the risk of QoS affectations. The proposed model considers SLA deadlines, predictions based on historical data, and dynamic occupation to determine the amount of resources to overallocate for each host. In addition, a compensation mechanism to adjust resource allocation in cases of underestimation is also described. In order to evaluate the model, simulation experimentation was conducted. Results demonstrate meaningful improvements in energy-efficiency while SLA-deadlines are slightly impacted. However, they also point the importance of strongest compensation policies to reduce availability violations especially during peak utilization periods.