Dual-Timescale Control for SDN-Enabled Collaborative Inference

Abstract

In dynamic edge networks, multi-node collaborative inference often suffers from amplified tail latency due to background traffic competition and node load variations. While existing work focuses on static computation-side optimization, it lacks online control over dynamic network factors. To bridge this gap, we propose an SDN-enabled collaborative inference control framework based on a dual-timescale optimization approach. On the slow timescale, we optimize controller placement and domain partitioning by modeling the trade-off between management cost and synchronization cost, solved via a randomized greedy approximation algorithm. On the fast timescale, we enable real-time regulation of inference-critical flows, including path reselection and bandwidth guarantee, based on instantaneous link states. Experimental results show that under high-load scenarios, our enhanced strategy reduces P95 and P99 latency by 38.5% and 44.6% respectively compared to traditional networks, significantly improving service stability.