Aquilas: Adaptive QoS-oriented Multipath Packet Scheduler with Hierarchical Intelligence for QUIC

Multipath packet scheduler is responsible for delivering each packet to an appropriate path. However, rules-based schedulers struggle to adapt to varying network conditions and diverse Quality of Service (QoS) requirements. Despite learning-based scheduler can adapt to various network conditions, reacting quickly to network changes is still challenging. Moreover, for diverse QoS requirements, learning-based schedulers often necessitates training from scratch. To solve the above challenges, we propose a multi-head mapping model that selects an optimal sub-scheduler based on the current state. It works with a shared state encoder, a multi-head Q-value decoder, and QoS-oriented reward decomposition. Furthermore, we propose Aquilas, an adaptive multipath packet scheduler with hierarchical intelligence for the Quick UDP Internet Connection (QUIC) protocol. Aquilas adopts a learning-based scheduler selector, thereby enabling selection of an optimal policy from a pool of sub-schedulers. This sub-scheduler pool encompasses a wide range of knowledge for handling various network conditions. In coarse time intervals, the learning-based scheduler selector operates, while during fine-grained time intervals, the selected sub-scheduler delivers each packet to the corresponding path. Aquilas has been evaluated in both controlled emulation and real-world networks. Compared with the state-of-the-art schedulers, Aquilas improves transmission performance in various network conditions and diverse traffic types.