Container Workload Prediction Using Deep Domain Adaptation in Transfer Learning

Containers are the primary deployment method for cloud applications, and accurate workload prediction is essential for resource allocation and energy optimization. Traditional statistical models struggle to capture complex workload variations, while classical neural network models require extensive historical data, which is challenging due to containers’ short lifespan. This paper proposes a container workload prediction model based on deep domain adaptation in transfer learning (CWPDDA). The model includes a feature extractor with self-attention and cross-attention mechanisms to extract private and shared features, a domain adversarial adapter to reduce distribution differences, and a workload predictor to directly apply source-target data for prediction, avoiding performance degradation from domain shift. To validate the accuracy of the proposed model, this study utilized the Alibaba cluster-trace-v2017 dataset as the target domain and the Google cluster-usage traces v3 dataset as the source domain. Experimental results showed that the proposed model achieved a substantial improvement in prediction accuracy compared to the Autoregressive Integrated Moving Average (ARIMA), Long Short-Term Memory (LSTM), Autoregressive Recurrent Neural Network (DeepAR), Deep Renewal Processes (DRP), and Multivariate Quantile Function Forecaster (MQF2) models.