PERF CLOUD: A Hybrid DTW–GRU Framework for Black-Box Virtual Machine Performance Degradation Prediction in Multi-Tenant Public Clouds

Abstract: Cloud computing environments host multiple virtual machines (VMs) on shared physical infrastructure, leading to unpredictable performance due to resource contention and workload variability. In public cloud settings, virtual machines operate as black-box systems, restricting access to internal application metrics and making accurate performance prediction highly challenging. To address this issue, this paper proposes PERFCLOUD, a hybrid machine learning framework for performance degradation prediction in multi-tenant public clouds. The proposed approach consists of three stages: (i) application type identification using Dynamic Time Warping (DTW) to align time-series workload patterns, (ii) highly correlated metric selection using Pearson Correlation to reduce noise and improve feature relevance, and (iii) time-series performance forecasting using a Gated Recurrent Unit (GRU) neural network. The model predicts both application type and VM performance status as Best or Degraded, enabling proactive resource management and interference mitigation. Experimental evaluation on a real-world cloud workload dataset demonstrates that the proposed DTW–GRU hybrid model significantly outperforms baseline LSTM-based approaches, achieving an accuracy of 99% with reduced computational overhead. The framework is further integrated into a web-based monitoring system for real-time VM performance analysis. The results validate the effectiveness, scalability, and adaptability of the proposed solution for intelligent cloud resource optimization.

Keywords: Cloud Performance Prediction, Black-Box Virtual Machines, Dynamic Time Warping (DTW), Gated Recurrent Unit (GRU), Performance Degradation Detection, Multi-Tenant Public Cloud, Machine Learning-Based Resource Optimization