On the feasibility of classifying flows in software-defined data center networks by an encoder-only architecture

Classifying elephant and mouse flows is pivotal for meeting Quality of Service requirements in Software-Defined Data Center Networks. Diverse approaches have addressed this classification problem by using machine learning and rule techniques in the data and control planes and the high-volume servers at the edge. However, those approaches can fall short when the classification task involves high-dimensional data generated by large networks. This paper presents CIEF-BERT, an approach devised to investigate the feasibility of using the encoder-only architecture of large language models to classify flows in data center networks that follow the software-defined networking paradigm. CIEF-BERT model leverages the self-attention capability to analyze large, complex, and high-multidimensional data and detect patterns provided by the Bidirectional Encoder Representations from Transformers to classify elephants and mice. CIEF-BERT was extensively evaluated using real traffic traces collected in a data center. Results show that CIEF-BERT is a promising solution for classifying elephant and mouse flows since the accuracy achieved by CIEF-BERT was equal to or slightly higher than that obtained by existing proactive controller and server-side flow classification approaches.