MRPNet: Workload Utilization Prediction for Configuration Tuning in Distributed Stream Systems

In the context of accelerating digital transformation and intelligent decision-making, distributed stream processing systems have become key infrastructures for supporting real-time analyses of large amounts of unbounded streaming data. To address the complex and difficult-to-predict workload variations in stream processing that may easily lead to insufficient configurations, we propose MRPNet, a lightweight application workload utilization prediction method. This method integrates multilevel discrete wavelet decomposition and includes a lightweight forecasting branch based on a patching mechanism to enhance the ability to capture local bursty patterns. Moreover, considering the asymmetric cost of prediction errors in configuration decision scenarios, we design a frequency-weighted asymmetric loss function (FWA-Loss), which effectively reduces the underprediction rate and magnitude by imposing heavier penalties on underpredicted samples. The experimental results demonstrate that MRPNet achieves excellent forecasting performance while maintaining a lightweight architecture. Compared with the baseline methods, MRPNet achieves maximum reductions of 13.2% and 14.3% in average symmetric mean absolute percentage error (SMAPE) under 15-min and 5-min sampling intervals, respectively. Moreover, FWA-Loss reduces the underprediction rate and underprediction magnitude by 37.4% and 41.3%, respectively.