Machine Learning-based Deep Packet Inspection at Line Rate for RDMA on FPGAs

FPGAs are becoming increasingly important in the cloud and data centers, especially as network-attached accelerators or reconfigurable Network Interface Cards (NICs). In the cloud, Remote Direct Memory Access (RDMA) over Converged Ethernet (RoCEv2) has emerged as the de facto standard protocol for data transport due to its low latency and high throughput. However, RDMA has several access control weaknesses limiting its applicability in the cloud. In this paper, we explore using machine learning-based deep packet inspection (DPI) as an enhancement to an open-source FPGA RDMA stack. The ultra low-latency ML model is integrated on the RDMA datapath and allows for detection of specific content in RDMA payloads (e.g., executables) at a line rate of 100Gbps while using less than 1% of the available resources. Compared with existing work, our solution operates on the full message payload, at the transport level, and on a complete RDMA stack without sacrificing compatibility with RoCEv2 and its native performance characteristics, proving its potential as an end-to-end solution.