Multi-Resource Workload Forecasting in Cloud Data Centers via Spatio-Temporal Graph Transformers
- Authors
-
-
Author
-
Author
-
Author
-
- Abstract
-
Addressing the challenges of strong spatiotemporal coupling, significant cross-node propagation, and dynamic topology changes with scheduling and scaling in multi-resource load prediction for cloud data centers, this paper proposes a unified modeling framework based on the Spatio Temporal Graph Transformer. This framework organizes multi-resource observations in the form of a temporal graph and explicitly injects relational structure information. The method abstracts the data center into a dynamic graph composed of nodes and edges at each time slice. Node features aggregate multi-dimensional resource indicators such as CPU, memory, disk, and network, while relational features characterize structural cues such as link types and interaction strengths, thus unifying multi-resource and structural constraints at the input level. Spatial encoding employs a graph attention mechanism to adaptively weight and aggregate neighborhood information, strengthening the expression of critical dependency paths and suppressing noise propagation. Temporal encoding utilizes self-attention to establish global associations across time steps within a historical window, improving the ability to characterize long-range dependencies and sudden disturbances. Subsequently, a gated fusion residual mechanism adjusts the information injection intensity, balancing representation updates and robustness. The output layer adopts a structure combining a shared backbone with resource-specific prediction heads, preserving common cross-resource patterns while characterizing resource dimensional differences, achieving multi-step, multi-resource joint prediction. To verify the effectiveness of the method, a unified evaluation index system was set up, and a horizontal comparison with representative baseline methods was conducted. The results show that the proposed framework can provide more consistent multi-resource prediction signals and provide more reliable forward-looking support for capacity planning and flexible scheduling.
- Downloads
- Published
- 2025-08-01
- Issue
- Vol. 5 No. 8 (2025)
- Section
- Articles