Dynamic Routing with Mixture of Experts for Distributed Sketch Measurement

Sketch enables line rate network measurement on the programmable data plane but suffers from resource constraints, making distributed deployment necessary. In distributed sketch measurement, dynamic routing planning (DRP) directly determines traffic distribution across sketches. Unbalanced loads cause sketch overflow and accuracy degradation. However, existing DRP methods face limitations in multi-objective planning, model interpretability, and timely model response. This paper proposes a MoE-based DRP framework with three core components. First, a MoE model comprising four expert networks and one gating network enables multi-objective planning (e.g., bandwidth, path length and switch overhead). Second, a teacher-student distillation framework injects domain knowledge into the MoE model, making each decision traceable to specific experts for interpretability. Third, a decision cache stores model decisions for recurring traffic states, reducing model response latency. Experiments on BMv2 show that our framework improves distributed sketch measurement accuracy by 6.7% and reduces maximum data plane overhead by 6.6%.