Causality-aware Graph Mixture of Experts for Accurate Multi-cloud Workload Prediction

Multi-cloud deployments are becoming increasingly necessary to meet data locality and compliance requirements, reduce latency for AI services, and mitigate the risk of single-provider failure. However, multi-cloud telemetry forms high-dimensional, heterogeneous and nonstationary time series where CPU, memory, disk, and LAN/WAN I/O exhibit time-varying, directional lead–lag effects across VMs and regions. Previous approaches, from RNNs to recent attention/GNN models, either assume linear stationarity or learn correlation-driven, largely symmetric dependencies that blur directed causal influence. To better leverage temporal dependencies and leverage multi-cloud deployments, we propose CAGMoE, a causality-aware dual-router Graph Mixture-of-Experts for multi-cloud workload forecasting. First, CAGMoE constructs two complementary graphs per window: an inter-metric graph derived from a transfer-entropy proxy of Granger causality to encode directed cross-metric influence, and an intra-temporal graph to capture local temporal continuity. Furthermore, a shared graph encoder produces token states and path summaries that drive dual routers to form Top-K sparse mixtures over experts. Finally, each expert is a FiLM-conditioned feed-forward network that injects a window-level causal vector to generate the final prediction. Experiments on real-world multi-cloud dataset MUCEP, Google Cluster, and Ali traces demonstrate that our method robustly compares with previous baselines, which improve in both accuracy and reliability, and demonstrate potential for industrial use.