Predictive Modeling of Resource Utilization in Cloud Data Centers Using Multi-Output Regression
Mustafa Daraghmeh, Anjali Agarwal, Yaser Jararweh · 2024
Integrating accurate resource usage prediction with cloud management systems is critical to optimize resource utilization and improve operational efficiency. The dynamic and non-linear usage patterns of resources in cloud data centers pose significant challenges for predictive modeling. Traditional single-output models, designed to predict a single value, often struggle to capture the complexities of series resource usage patterns. Current predictive models do not consider the interdependencies and interactions of various resource usage patterns in a sequence. Therefore, it is necessary to develop more robust predictive methods that can predict a series of resource usage patterns. This study introduces an innovative predictive model that uses multi-output regression combined with time series windowing, usage pattern clustering, and various transformation methods to predict a series of resource usage with high precision for heterogeneous cloud computing systems. Transformation methods include a power transformer to normalize the data distribution, a standard scaler to standardize the feature space, a polynomial transformation to enhance model complexity, and principal component analysis to reduce the dimensionality of the training feature space. The proposed model is evaluated using multi-output regression benchmarks with real cloud workloads and various evaluation metrics. The results demonstrated that the proposed approach significantly improves prediction accuracy while reducing training costs, offering substantial potential for improved performance and efficiency in cloud computing operations.