Load Balancing Based on Flow Classification with Private Link in Programmable Switch

In data center networks, the traffic exhibits multiple patterns and possesses dynamic and bursty characteristics, which pose high requirements for load balancing strategies. Existing load-balancing schemes used in data center networks have some shortcomings, such as the lack of differentiation between flow types and high delays during rapid reactions to microbursts. Thus, we propose a load-balancing scheme based on flow classification with private link in Programmable Data Planes (PDP). In our scheme, we constructed a Binary Neural Network (BNN) flow classification model to identify elephant flows and mice flows. Subsequently, we establish private links for elephant flows and mice flows respectively. In each private link, we utilize the local switch queue state to obtain the congestion index for each egress port, consequently obtaining the forwarding probability for each port. Moreover, we propose a flowcell-based forwarding mechanism. The results demonstrate that our approach improves performance in terms of deadline meet rate for mice flows, and the flow completion time for elephant flows. At the same time, it also has a significant improvement in flow completion time under the mixed flow.