Load prediction algorithm for multi-tenant virtual machine environments

Computer systems configured in Cloud architectures have become more prevalent in the data center. However, a growing concern is the increasing amount of energy required to operate the data center. This has led to an industry-wide effort to look for ways the energy footprint of the datacenter can be reduced. Efficient scaling of the cloud's nodes is one way to manage the energy consumption. This goal is achieved through the application of energy optimization techniques. One of the techniques is to dynamically provision IT resources by switching the state of the cloud nodes from active to sleep and from sleep to active in response to the actual network load. To compensate for the potential impact of the time delay inherent in changing the state of a system, there is need to accurately predict the future load. The multi-tenant nature of cloud computing systems require that the load prediction takes into account that there can be multiple deployments of the same configured system over many different physical virtual and physical systems as well as that different services can be deployed in clusters across both virtual and physical systems. It is therefore important to predict the future network load per service-cluster. This paper presents an algorithm for predicting future request workload for multiple services then tests the algorithm using public NASA web server traffic data on four different services.