Enhancing Microservice Workload Forecasting with Event Embedding

Accurately forecasting microservice workload is essential for resource allocation in cloud services. Many deep models have been developed for microservice workload prediction, with KAE-Informer being a state-of-the-art model in this field. It introduces historical similar events to provide richer information for prediction. However, it directly concatenates historical similar events with the corresponding input sequence as model inputs, which may introduce unnecessary noise and interfere with model predictions. To enable historical similar events to provide valuable information effectively, we propose an event-embedded microservice workload Transformer called EE-LoadFormer, inspired by TimeXer. Our proposed model employs a cross-attention mechanism to select valuable information from historical similar events, assisting the model in making predictions. Experimentally, EE-LoadFormer outperforms state-of-the-art models, reducing the MSE and MAE by about 64% and 56%, respectively.