Workload prediction for resource management in data centers

Resource management is to arrange and allocate resources for computing operations and applications. In large scale data centers that contain thousands of servers, resource management is critical for efficient operation. To know workload characteristics in advance helps us proactively control resources in data centers, leading to benefits such as power savings and improved service performance. Workload prediction can be used, e.g., to decide how many resources to allocate for each application in a data center in the future. The accuracy of workload prediction varies depending on the used prediction methods and the characteristics of the workload. In this thesis work, we investigate three different methods: Linear Regression (LR), Adaptive Neuro-Fuzzy Inference Systems (ANFIS), and Nonlinear Autoregressive Network with Exogenous Inputs (NARX). These methods are used to build models of resource consumption such as memory, CPU, and disk. Based on these models, future workload resource usage is predicted, and the accuracy of prediction is assessed. We analyze a trace from a production cluster at Google, predict resource consumption for different time intervals, and compute the error between predicted and actual values. The results show that NARX gives higher accuracy than ANFIS and LR when forecasting one-step ahead prediction, and that the ANFIS method provides the best result with multi-step ahead prediction compared to the others. Finally, time to train and re-train LR, ANFIS and NARX are computed. The running times are short, suggesting that the methods can be used in real-time operation