Pattern Mining-Based Integrated Prediction Method for Cloud Workload
Jinhui Jiang, Xiushuang Yi, Yuting Zhao, Tengsheng Tu · 2024
Cloud service providers require accurate prediction of cloud workloads to promptly determine resource allocation strategies and enhance resource utilization while meeting Service-Level Agreements. However, existing workload prediction studies face significant challenges in addressing the high variability and diversity of cloud workloads across different time scales. To tackle this, we propose a pattern mining-based integrated forecasting approach for cloud workloads (PM-SGRUs-DTW). This method initially identifies potential workload patterns through pattern mining, then trains multiple stacked GRU models in parallel, and finally utilizes Dynamic Time Warping to compute weights for obtaining the ultimate prediction results. Experimental findings demonstrate that when using real workload traces from Alibaba Cloud data centers, PM-SGRUs-DTW achieves higher prediction accuracy across various forecasting horizons compared to GRU-based and other workload prediction methods.