Resource Demand Prediction in Multi-Tenant Service Clouds

Cloud computing is emerging as an increasingly popular computing paradigm. Cloud computing allows dynamic scaling of resources available to users as needed. The increasing demand for cloud computing resources has led to increased virtual machine provisioning on the fly. However, initializing a new virtual machine instantly in a cloud is not possible, and it could take several minutes thereby leading to increase in latency. This requires a highly accurate demand prediction framework that can provision resources in advance, thereby minimizing the downtime of virtual machine. In this paper, we propose a resource demand prediction framework in multi-tenant service clouds. The proposed framework employs data mining techniques, which extract high level characteristics from historical demand behavior and provision resources in advance. Our framework classifies the service tenants depending on whether the resource demand for them is expected to increase or not. Then our framework proceeds for prediction in service tenants in which resource demand would increase, so that the prediction time can be minimized. It also makes predictions of short-term and long-term resource demand. We demonstrate the accuracy of our framework via extensive experiments.