An ensemble of automatic algorithms for forecasting resource utilization in cloud

摘要

Forecasting resource usage in cloud environment is a challenging problem due to heterogeneous characteristics of workloads running on cloud. Each virtual machine (VM) resource usage can be characterized as steady, trend, seasonal, cyclic or bursty pattern. Manually it is not feasible to fit the forecasting models for each of thousands of VMs running on cloud. Further different forecasting models are suitable for different type of workload. Manual selection of best model is time consuming and cumbersome. In this manuscript we propose, implement and evaluate an automated technique of combining the forecasts from multiple non-overlapping time series forecasting methods. The objective is to improve accuracy and robustness of the forecasting algorithm. The experiments were conducted using publicly available yahoo dataset and cloud datacenter dataset. The results show that proposed technique is successful in forecasting with better accuracy, irrespective of workload type and without any manual intervention.