DRLS: A Deep Reinforcement Learning Based Scheduler for Time-Triggered Ethernet

• Main Authors: Zhong, Chunmeng, Jia, Hongyu, Wan, Hai, Zhao, Xibin
• Format: Conference Proceeding
• Language: English
• Subjects: Deep learning; deep reinforcement learning; Job shop scheduling; Network topology; Processor scheduling; Reinforcement learning; Scalability; Schedules; time-triggered Ethernet; TTEthernet scheduling
• ISSN: 2637-9430
• DOI: 10.1109/ICCCN52240.2021.9522239

Time-triggered (TT) communication has long been studied in various industrial domains. The most challenging task of TT communication is to find a feasible schedule table. Network changes are inevitable due to the topology dynamics, varying data transmission requirements, etc. Once changes occur, the schedule table needs to be re-calculated in a timely manner. Solver-based methods and heuristic-based methods were proposed to solve this problem. However, solver-based methods employ integer linear programming (ILP) or satisfiability modulo theories (SMT) which have high computational complexity. On the other hand, heuristic-based methods are fast, but they need to be handcrafted based on the application characteristics. Thus, these methods are not general enough to work in complex scenarios especially in large networks.In this paper we propose DRLS - Deep Reinforcement Learning based TT Scheduling method. DRLS first trains an application or network specific scheduling agent offline. Then, the agent can be used for online scheduling of TT flows. However, off-the-shelf reinforcement learning techniques cannot handle the TT scheduling problem with typical complexity and scale. DRLS provides novel solutions to this challenge, including three key innovations: new representations for TT network adapted to various topologies, proper deep neural network (DNN) structures to capture network characteristics, and scalable reinforcement learning (RL) models to handle online TT scheduling. Comprehensive experiments have been conducted to compare the performance of DRLS and other methods (heuristics-based methods such as HLS, LS, HLD + LD, LS + LD, and ILP-based method). The results show that DRLS can not only adapt to specific network topologies, but also have better performance: runs much faster than ILP solver-based methods, and schedules about 23.9% more flows than traditional handcrafted heuristic-based methods.