Machine learning-centric prediction and decision based resource management in cloud computing environments

Shobhana Kashyap, Avtar Singh, Sukhpal Singh Gill

Abstract: In cloud data centers, precise resource prediction is a critical issue due to the dynamic environment, the presence of irrelevant data points, and the unpredictable nature of resource demand. An accurate prediction helps with resource management, cost planning, and improving cloud-related services, whereas an inaccurate prediction increases the budget because of unused and overused resources. The presence of dynamic and irrelevant data not only creates confusion in the model, resulting in inaccurate predictions, but also adds unnecessary complexity and cost to the entire process. To address these challenges, we propose an approach for multi-model methods that uses a sliding window method with an adjustable size to estimate important data points from the real trace. The current work conducts three experiments to assess the impact of unpredictable data and improve the models' performance for greater accuracy. Initially, we conducted the experiment on entire datasets, but this approach failed to produce accurate and efficient machine-learning models. The next fixed window technique provides activity recognition in real-time, making it suitable for applications that require immediate feedback or response. Finally, a novel Variable-Size Sliding Window (VSSW) is proposed that selects relevant data points that help to provide better performance. Additionally, a Model Selector Decision Support System (MSDSS) is designed for forecasting and optimizing resource demand. This system determines the best predictive model for a specific set of resources based on observations gathered over a defined time frame. The experimental outcomes demonstrate that the proposed algorithm has improved the Mean Absolute Error (MAE) by approximately 50.01% compared to the baseline method and approximately 31.75% compared to the fixed window size approach. Furthermore, the proposed model effectively addresses the challenge of predicting resource workloads in a dynamic environment.