Enhancing End-to-End Delay Satisfiability for IEEE 802.1Qav

Abstract

IEEE standard 802.1Qav evolves from Ethernet Audio Video Bridging (AVB) to support real-time transmission by performing Credit-Based Shaping (CBS) of flows at each hop. Prior work primarily focuses on achieving a tighter upper bound of the End-to-End (E2E) delay using network calculus, given the CBS configuration of flows. However, it is important to adjust the CBS configuration at each hop to meet the deadlines for various data flows. To this end, we address the CBS scheduling problem to enhance the satisfiability of E2E delay requirements. First, we formulate a linear constraint model and propose a Double Deep Q-Network (DDQN) reinforcement learning method, called $ \mathtt{CBS-DDQN} $, to solve the model. Second, to improve model training and scheduling, we present a fast method using equivalence reduction to calculate the E2E upper bounds of data flows for the evaluation of E2E delay satisfiability. Finally, we conduct an extensive evaluation of $ \mathtt{CBS-DDQN}$, demonstrating its ability to efficiently schedule thousands of flows and increase the number of flows meeting the delay requirements by 18–26% over prior methods.