Niche: Isolation-oriented competition-aware AQM on programmable switches
Abstract
To meet heterogeneous application demands, the Internet must employ multiple congestion control algorithms(CCAs) concurrently. Coexisting congestion control algorithms exhibit divergent performance when competing for resources such as bandwidth and buffers. However, mainstream queue management lacks algorithm-aware differentiation, leading fairness-focused bandwidth allocation to overlook latency and buffer utilization, thus failing to achieve holistic, system-level optimization. In this paper, we introduce Niche, an active queue management (AQM) for programmable switches with online contention awareness and dynamic isolation. Niche extracts buffer-behavior features to classify traffic online and isolates flows in physical queues to suppress aggressive flows. By integrating dynamic bandwidth allocation, Niche enforces bounded fairness while optimizing buffer utilization and minimizing latency and jitter. Validated on the P4-programmable switches, Niche employs physical-queue isolation to mitigate resource contention, which directly reduces buffer occupancy while significantly improving latency, jitter, tail latency, and throughput fairness, thereby achieving systemic performance gains.