TARA: Enhancing Real-Time Network Traffic Classification with Hardware Virtual Layers

In network monitoring, accurately classifying traffic is essential for resource management and Quality of Service (QoS) tasks. With the rise of IoT and an ever-increasing number of connected devices running modern AI-based applications, high volumes of data need to be processed at high-speeds. The state-of-the-art employs Deep Neural Networks (DNNs) for classifying the network traffic. However, DNNs are computationally heavy, challenging their adoption in the networking infrastructure. Because of that, FPGA accelerators have been used for Smart Network Interface Cards (SmartNICs). However, these FPGA accelerators need to be reconfigured when changing the network classification task - leading to significant time and energy overheads. To solve this issue, this work presents TARA, a novel solution that employs Hardware Virtual Layers (HWVL) to eliminate the need for FPGA reconfigurations, enabling fast switching for various classification tasks. Utilizing a comprehensive evaluation framework, TARA improves Quality of Experience (QoE) and energy efficiency across diverse network scenarios when compared to traditional FPGA reconfiguration methods.