Spatiotemporal Transformer-Based Latency Prediction for Large-Scale Cloud Computing Clusters

Abstract

This study addresses the key problem of latency prediction in cloud clusters and proposes a spatiotemporal Transformer-based modeling method to overcome the limitations of traditional approaches in capturing complex dependencies and achieving high accuracy in large-scale dynamic environments. The research first analyzes the spatiotemporal characteristics of latency, noting that latency shows both periodic fluctuations and sudden spikes in the time dimension, while in the spatial dimension, it is influenced by the coupling of topology structures and resource competition among nodes. To address this, a unified spatiotemporal Transformer framework is constructed, where temporal self-attention captures long-range dependencies, and spatial self-attention with adjacency matrices models node interaction features, enabling a comprehensive representation of dynamic latency patterns. The method design integrates an embedding module for high-dimensional representation of multi-source inputs, residual connections, and normalization to enhance stability and generalization, and a linear mapping output layer for precise latency predictions. Experiments are conducted on the public Google Cluster Trace dataset with comparisons against mainstream models and multidimensional sensitivity analyses. Results show that the proposed method achieves superior performance in MSE, RMSE, MAE, and R², significantly outperforming Transformer, LSTM, TimeMixer, and ITransformer. Furthermore, the method maintains stable performance under hyperparameter sensitivity tests, environmental disturbances, and data missing scenarios, demonstrating strong robustness and adaptability in complex distributed environments. In conclusion, the proposed framework achieves high accuracy and reliability in latency prediction tasks and provides effective technical support for performanceoptimization andresource management incloud computingsystems.