Elastic Resource Prediction for Cloud Data Center

Abstract

Cloud data centers offer utility-oriented IT services to users worldwide. The nature of resource demand of cloud data centers is elastic as the overall workloads are always changing. For handling dynamic workload nature ahead of the needs, elastic resource demand prediction is the key issue in cloud data centers. If the cloud provider does not ensure they have enough resources to meet demand which will lead to under or over provisioning of resources. In this paper, integrated elastic resource prediction system is proposed by combining signaturebased prediction and state-based prediction approahces. The workload nature of the cloud data centers are both repeating pattern and nonrepeating pattern workload. Signature-based prediction is used to predict the repeating pattern workload and state-based prediction is used to predict the non-repeating pattern workload. Integrated Elastic Resource Prediction (IERP) system is used to predict the mixed workload pattern. Feature selection is conducted first to reduce processing overheads while achieving high prediction accuracy. The proposed predictors are implemented and evaluated with real world workload traces which show that they achieve high resource prediction accuracy with above 95%.