An Interference-Aware QoS Violation Alleviation Framework for Multi-Tenancy Public Clouds

Public clouds employ multi-tenancy to increase resource utilization, but this can lead to co-location interference on shared resources, potentially causing performance degradation of cloud applications. In addition, virtual machines (VMs) are black boxes to the cloud providers, which makes performance monitoring challenging. To address the black-box scenario, we propose a novel machine learning framework, Alioth, to detect the quality of service (QoS) violations of cloud applications running inside VMs using low-level metrics. The framework then migrates the interfered VM to other hosts to mitigate QoS degradation and improve resource utilization. We first buildAlioth-datasetvia comprehensive co-location experiments to reflect the complexity and dynamicity in real-world scenarios. Then we construct the Alioth predictor by augmenting input features using denoising auto-encoders and employing XGBoost for QoS violation prediction. We improve the generalization and interpretability of the predictor by incorporating the domain adaptation neural network and the SHAP explainer. Moreover, we propose a VM migration algorithm based on the predictor and the interference-balancing strategy to optimize both VM QoS and resource utilization. Experimental results on a real-world dataset demonstrate that Alioth outperforms other state-of-the-art methods in terms of prediction accuracy, generalization, robustness, and VM migration performance.