@

초록

In this paper, we propose a framework that performs traffic forecasting for cloud applications, including web services and mobile applications, in data centers. To support simultaneous traffic forecasting of multiple applications, we suggest a light-weight deep learning model that consists of only linear layers based on a popular time-series decomposition method. We also propose online learning and transfer learning methods, which help a traffic predictor learn the dynamic changes of network traffic patterns and perform traffic forecasting for a new application whose amount of the dataset for training is insufficient, respectively. We evaluate the proposed framework using all the packet traces captured at the entry switch in a real-world data center, operated by a popular Internet service provider (ISP) in South Korea. Our experiments on the traces of five representative applications used by more than 10 M users demonstrate that (i) the proposed traffic predictor cooperating with online learning can accurately estimate the traffic volumes of the applications, which outperforms state-of-the-art time-series forecasting models, despite its short processing times for inference and online learning (2 ms and 0.5 s, respectively) and (ii) a transfer learning method with online learning can be an effective solution to predict traffic volumes of a newly-deployed application. In particular, we reveal that choosing the predictor of the application whose traffic scale is similar to or lower than the target one is useful for transfer learning by reducing the convergence time of adjusting the transferred model to the traffic patterns of the new application. Finally, we introduce a deployment scenario of the proposed framework. We believe that our work can provide valuable insights for researchers and engineers who not only want to develop traffic forecasting or its associated tasks such as anomaly detection, but also aim to effectively manage resources for cloud applications in data centers.

키워드