DBDW: A Dual-Branch Database Workload Prediction Model

Scalable cloud databases allow users to dynamically adjust computational resources based on business needs. However, most current elastic scaling techniques face challenges in making sound decisions due to complex workload variations. Achieving efficient resource allocation requires accurate workload modeling and timely prediction of future workload changes. Therefore, we propose DBDW, an accurate and lightweight model for database workload prediction. DBDW consists of Local Branch and Global branch. The Global branch uses a Mixer architecture to model long-term cloud database workloads at a holistic level. The Local Branch employs multiple components to selectively emphasize local features via a gating strategy. It also uses a channel sparse clustering method to collaboratively represent multichannel information with low computational overhead. Experimental results show DBDW effectively models workload sequences using historical data. Its lowest prediction MAE/MSE for future workloads are 0.5953/0.5971. Furthermore, DBDW demonstrates significantly reduced computational costs: GPU memory usage during training is 304.14 MB , training speed reaches 0.3238 s/ epoch , parameter count is 0.34 million, and FLOPs are 4.19 million. This confirms DBDW ensures accurate predictions while greatly reducing overhead, allowing deployment without affecting database operations.