A Supervised Learning Solver for Proactive Optimization of Cross-Cloud Applications

Autonomic Cloud application optimization is necessary to maintain the maximal application utility while ensuring efficient use of Cloud resources across providers and infrastructures. This requires continuous monitoring of the application to detect the needs for application reconfiguration. The vector of monitored parameters is a multivariate time series and, therefore, one can predict the future metric values and optimize Cloud applications proactively ensuring the new resources to be available when needed. However, the predictions are uncertain, and the optimization must be resilient to the inherent prediction errors. This paper presents a new supervised learning solver to find the Cloud application configuration with the highest utility value for predicted runtime conditions. The solver is trained on uncertain measurements, and it is evaluated for a real-world data intensive Cloud application in terms of the quality of returned solutions for various loss functions used during training.