Towards a Pluggable and Implementation-Agnostic Learning-Based Congestion Control for QUIC

QUIC is rapidly becoming the de facto transport standard for HTTP Adaptive Streaming (HAS), offering multiplexed, lowlatency delivery over UDP. However, most existing QUIC congestion control (CC) algorithms are heuristic-based and largely inherited from TCP designs, limiting their adaptability across diverse network conditions and ultimately degrading the user’s Quality of Experience (QoE). Learning-based CC approaches have shown promise in TCP but are typically non-modular and tightly coupled to specific implementations, making their integration with QUIC impractical. To address these limitations, we propose AERO —the first learning-based, plug-and-play congestion control algorithm designed for portable integration across QUIC stacks. AERO dynamically adapts to varying network characteristics and can be seamlessly integrated into any QUIC stack without extensive engineering effort. We conduct extensive trace-driven evaluations to assess AERO ’s performance from both the transport and application layers, examining its influence on client-driven adaptive bitrate (ABR) algorithms. Experimental results show that AERO enhances video quality (VMAF) by approximately 13% and reduces rebuffering by 30% in video-on-demand scenarios, while achieving around 12% VMAF improvement and 65% fewer stalls in low-latency live streaming compared to state-of-the-art approaches. Moreover, AERO performs favorably in terms of delivery rate and delay across diverse network conditions, demonstrating its robustness and QUIC-friendliness.