End-User Performance Isolation at Scale in Access Aggregation Networks

As broadband access speeds rise, the Access-Aggregation Network (AAN) faces growing challenges in ensuring scalable, fair, and predictable resource allocation for millions of subscribers. Traditional fairness mechanisms like Fair Queueing struggle to meet the scalability demands of modern AANs, constrained by hardware limitations and the complexity of managing per-user states at high speeds. Emerging programmable data plane technologies like P4 enables non-traditional QoS and resource sharing solutions. Core-Stateless Active Queue Management (CSAQM) is a recent proposal leveraging edge-based packet marking to enable fine-grained resource sharing while maintaining simplicity in core network schedulers. In this paper, we present a scalable packet marking algorithm for CSAQM, implemented directly on a Tofino ASIC using P4 programming language. Our approach enables efficient, real-time marking of packets at line rate, supporting over one million users on a single switch. We detail the algorithm design and evaluate its performance in a high-speed testbed, simulating fair queueing for thousands of users on a 100 Gbps bottleneck. The results demonstrate that the proposed marker achieves significant scalability and performance, enabling resource allocation at scale in broadband networks.