Abstract:
Power consumption prediction for virtual machines (VMs) is key in improving energy efficiency, capacity planning and sustainability in cloud computing. In this work, we present a comprehensive empirical study that compares several Graph Neural Network (GNN) architectures such as Graph Convolutional Network (GCN), GraphSAGE, Graph Attention Network (GAT), a Temporal GAT / dynamic graph design, GIN, R-GCN, Graph Transformer and Graph NODE for predicting VM power consumption using a real VM monitoring dataset. We assess the impact of different graph construction strategies including k-nearest neighbour, feature similarity threshold and temporal adjacency on predictive performance. Using standard regression metrics (MAE, MSE, RMSE, R²), we reveal that while all GNN models struggle to consistently surpass a naïve mean baseline, feature similarity threshold provides better performance and T-GAT architecture perform better than other. We analyze failure modes and offer recommendations for future work, including richer feature engineering, learned graph structures, and hybrid temporal-spatial modelling.
