Modeling and predicting quality of service for application migrations in the cloud-edge continuum

Abstract

In cloud-edge infrastructures, computing resources are distributed across centralized cloud servers and decentralized edge devices. This hybrid model allows for efficient resource utilization but also introduces challenges, as resource demands can fluctuate rapidly due to varying workloads. Managing application performance in this continuum requires efficient migration strategies to maintain Quality of Service (QoS). Some proposed reactive strategies based on immediate system conditions (i.e., previous value of QoS) which do not adapt to changing conditions neglect resource changes until they affect performance. This thesis explores how migration can be optimized to improve QoS in such cloud-edge infrastructure by proposing an AI-driven methodology that leverages lightweight models suitable for resource-constrained environments. Using a proactive migration approach to manage application migrations through the cloud-edge infrastructure, adapting in a smart, flexible way to handle resource changes before they affect performance. First, we create a controlled and reproducible testbed to simulate a realistic interaction between the baseline application and the cloud-edge infrastructure by applying internal and external stress, allowing for data generation. This comprehensive data generation serves as the foundation for developing robust predictive models. Then, we introduce linear models and lightweight Long short-term memory (LSTM) networks for predicting QoS based on real-time telemetry data. Our findings show that LSTMs outperform linear methods with a 34.28% reduction in Mean Squared Error (MSE). Additionally, we further investigate direct and iterative forecasting approaches for lightweight LSTM models, revealing that the direct approach provides better accuracy, achieving a 38.48% reduction in MSE compared to the iterative approach when predicting QoS. Finally, we enable the proactive migration policy based on the best model(i.e., direct approach LSTM). The results show that proactive migration policies outperform reactive ones, increasing migration detection by 66.65% and reducing time spent breaking Service Level Agreement by 56.92%. The proactive approach excels in detecting critical scenarios, accurately flagging more than 90% of critical cases with potentially large SLA violations.