FORECASTING COMPUTING RESOURCE UTILIZATION IN SERVER CLUSTERS USING A LIGHTWEIGHT TRANSFORMER

The computing resource utilization in server clusters changes dynamically and non-stationarily; thus it is critical to accurately forecast such utilization for effective resource scheduling, auto-scaling, and workload management. Existing forecasting approaches lack abilities to model complex temporal dependencies, while standard Transformers exhibit powerful long-range modeling but incur significant computational and memory overheads. In response, we propose a lightweight Transformer for computing resource utilization forecasting in server clusters. It leverages compact temporal representations and efficient feature interactions to lower computational complexity and enable multi-step forecasts of future resource utilization. We conduct experiments based on real-world server cluster traces, comparing our approach against several representative baselines: LSTM, GRU, DLinear, Transformer, PatchTST, and iTransformer. Both the forecasting accuracy and model efficiency are measured by prediction error, parameter count, computational cost, and inference efficiency respectively. Our result shows that our approach achieves good forecasting performance without significantly increasing computational overhead, which achieves a balanced trade-off between predictive accuracy and model efficiency. Thus, our framework offers an efficient solution for practical resource utilization forecasting in server-cluster environment.