A Machine Learning Method and Device Based on Programmable Switch

Machine learning methods have many excellent properties, such as the quality and efficiency of algorithmic that increase with the number of training sessions as new data is fed in. In a large network topology, how to balance the traffic in the network and improve the link utilization has always been a concern of network engineers. We take advantage of the Programming Protocol-Independent Packet Processors(P4) language with protocol-independent features, use in-band telemetry to collect port traffic statistics, delays and other information on the link, use routing protocols to collect topology, weight etc and transmit these information to machine learning server through the Google Remote Procedure Calls(gRPC) interface. The machine learning server uses a machine learning algorithm to generate a policy for adjusting the traffic, converts the policy into a forwarding table and sends it to the forwarding plane to balance the link traffic.