Self-stabilising QoS through autonomic scheduling of latency-critical traffic

Abstract Conventional queuing systems such as First-In-First-Out (FIFO) scheduling often experience unfair latency spikes and unbounded jitter. This makes them inefficient when handling bursty and heterogeneous traffic. As a result, maintaining fair and reliable Quality of Service (QoS) in mixed-traffic networks remains challenging. This work targets heterogeneous traffic (specifically mixed application classes: Video, VoIP, Chat, Browsing, File Transfer, and P2P sharing a single wired edge-router queue) rather than heterogeneous wireless access networks. Recent research has explored biological regulation mechanisms as inspiration for adaptive and robust resource management in complex systems. This paper formulates and evaluates a lightweight closed-loop QoS scheduler that converts application-level traffic semantics obtained from ML classification into local waiting-time-based priority adaptation, and systematically characterises the resulting latency, fairness, and stability trade-off under normal load, erroneous classification, multi-hop, and overloaded conditions. The MA-QoS dynamically adjusts packet priorities over time according to $$P_{\text {dynamic}} = P_{\text {base}} + (T_{\text {queue}} \times W_{\text {metabolic}})$$ , where $$P_{\text {base}}$$ encodes the application class’s latency sensitivity, $$T_{\text {queue}}$$ is the packet’s waiting time, and $$W_{\text {metabolic}}$$ is a tunable ageing rate; as a packet waits longer, its effective priority rises, preventing starvation while preserving class ordering under normal load. The model is evaluated using the CIC-VPN2016 dataset, which provides labelled network traffic traces covering application types such as Video, VoIP, File Transfer, and VPN-encrypted flows. Traffic flows are first classified using a Random Forest with 93% accuracy. Each flow is assigned a base priority by class, and a “metabolic” weight is used to boost its effective priority as it waits in the queue. In simulations with 10,000 Poisson arrival tasks, the MA-QoS queue reduces the median delay for video flows from 5.46 time units under FIFO to 0.52 time units, while also achieving consistently lower delay variance compared with RED and WFQ. Notably, MA-QoS also matches or exceeds a Deep Q-Network scheduler on latency-sensitive traffic while requiring no training; MA-QoS produces deterministic decisions for a fixed traffic trace, whereas the DQN’s run-to-run variance across independently trained models can exceed its mean delay. These improvements are obtained under identical load conditions: aggregate throughput and Jain’s fairness index are provably identical across all evaluated schedulers, achieved with modest additional computational overhead relative to FIFO. Overall, the results demonstrate that the proposed class-aware priority scheduler produces substantial latency and jitter reductions for high-priority traffic, while the metabolic ageing mechanism adds starvation protection when active, making MA-QoS a practical approach for QoS management in heterogeneous networks.