Arbitrated Packet Switching With Machine Learning Driven Data Plane

Modern router architecture typically separates the control plane and the data plane. The control plane, which provides routing information, usually does not need to run at line rate. The data plane, on the other hand, must keep pace with the rate of the incoming packets from the input ports, processing every byte and transmitting it on the output ports. To prevent internal congestion that could result in packet loss, each internal node must operate at least at the same rate as that of the input port. With network speeds continually rising, designing the data plane, which directly affects system performance, has become increasingly more challenging. A widely-used architecture is the arbitrated packet switching where an arbiter manages the flow of traffic from input (ingress) to output (egress) using credits. The credits indicate the available space on the egress side corresponding to the amount of data that can be transferred from ingress to egress. To ensure a sufficient supply of credits for the system, an arbitrarily large number is often assumed. While this is a good architecture, it is quite costly. This paper proposes a more efficient approach that integrates machine learning techniques, specifically Q-Learning (QL) and Mamdani Fuzzy Inference System (MFIS). QL determines the optimal policy that guarantees the shortest route from ingress to egress. MFIS enables real-time control of credit assignments allowing credit sharing. By employing both QL and MFIS the overall number of credits could be significantly reduced resulting in substantial cost savings without sacrificing performance.