Umeå University's logo

Efficient workload prediction is essential for enabling timely resource provisioning in cloud computing environments. However, achieving accurate predictions, ensuring adaptability to changing conditions, and minimizing computation overhead pose significant challenges for workload prediction models. Furthermore, the continuous streaming nature of workload metrics requires careful consideration when applying machine learning and data mining algorithms, as manual hyperparameter optimization can be time-consuming and suboptimal. We propose an automated parameter tuning and adaptation approach for workload prediction models and concept drift detection algorithms utilized in predicting future workload. Our method leverages a pre-built knowledge-base based on historical data statistical features, enabling automatic adjustment of model weights and concept drift detection parameters. Additionally, model adaptation is facilitated through a transfer learning approach. We evaluate the effectiveness of our automated approach by comparing it with static approaches using synthetic and real-world datasets. By automating the parameter tuning process and integrating concept drift detection, in our experiments the proposed method enhances the accuracy and efficiency of workload prediction models by 50%.

Cloud management systems increasingly rely on machine learning (ML) models to predict incoming workload rates, load, and other system behaviors for efficient dynamic resource management. Current state-of-the-art prediction models demonstrate high accuracy, but assume that data patterns remain stable. However, in production use, systems may face hardware upgrades, changes in user behavior etc. that lead to concept drifts - significant changes in characteristics of data streams over time. To mitigate prediction deterioration, ML models need to be updated - but questions of when and how to best retrain these models are unsolved in the context of cloud management. We present a pilot study that address these questions for one of the most common models for adaptive prediction - Long Short Term Memory (LSTM) - using synthetic and real-world workload data. Our analysis of when to retrain explores approaches for detecting when retraining is required using both concept drift detection and prediction error thresholds, and at what point of retraining should actually take place. Our analysis of how to retrain focuses on the data required for retraining, and what proportion should be taken from before and after the need for retraining is detected. We present initial results that indicate that retraining of existing models can achieve prediction accuracy close to that of newly trained models but for much less cost, and present initial advice for how to provide cloud management systems with support for automatic retraining of ML-based methods.