Performance assessment of time series forecasting models for simple network management protocol-based hypervisor data

Abstract

Time series forecasting is vital for predicting trends based on historical data, enabling businesses to optimize decisions and operations. This paper evaluates forecasting models for predicting trends in simple network management protocol (SNMP)-based hypervisor data, essential for resource allocation in cloud data centers. Addressing non-stationary data and dynamic workloads, we use PyCaret to compare classical models like autoregressive integrated moving average (ARIMA) with advanced methods such as auto ARIMA. We assess 30 models on metrics including CPU utilization, memory usage, and disk reads, using synthetic and real-time datasets. Results show the naive forecaster model excels in CPU and disk read predictions, achieving low root mean squared errors (RMSE) of 0.71 and 869,403.35 for monthly and daily datasets. For memory usage predictions, gradient boosting with conditional deseasonalisation and detrending outperforms others, recording the lowest RMSE of 679,917.6 and mean absolute scaled error (MASE) of 4.46 on weekly datasets. Gradient boosting consistently improves accuracy across metrics and datasets, especially for complex patterns with seasonality and trends. These findings suggest integrating gradient boosting and naive forecaster models into cloud system architectures can enhance service quality and operational efficiency through improved predictive accuracy and resource management.