| Abstract: |
Cloud computing and virtualization technologies are essential components of modern IT infrastructures, enabling flexible resource allocation and efficient use of computational resources. Accurate prediction of virtual machine workloads is critical for dynamic resource management, cost optimization, and maintaining quality of service.
This work investigates virtual machine workload forecasting using machine learning methods. We analyze the statistical properties of time series representing real-world virtual machine workloads from the Microsoft Azure Traces dataset. Our analysis shows that most virtual machines exhibit relatively stable workload levels over extended periods. By leveraging statistical features such as Mean Squared Deviation from a Moving Average and Spike Rate, we identify distinct workload patterns and group virtual machines with similar characteristics.
We find that certain workload classes can be forecasted accurately using simple models, with these classes accounting for over 50% of the dataset, highlighting potential computational savings. Comparative evaluation of forecasting models shows that preliminary clustering improves prediction accuracy and enables the use of lightweight models. Moreover, our approach significantly reduces inference time, which is crucial for real-time applications like load balancing in cloud environments. |