Towards In-Band Telemetry for Self Driving Wireless Networks

Self-driving network is an emerging network automation design principle for building next generation autonomous networked systems based on machine learning algorithms trained on real-time experiences, i.e., network state measurements. However, existing network measurement techniques are designed on centralized architecture leading to considerable control overheads in wireless networks. In this work, we designed and implemented a distributed In-band network telemetry system (S-INT) and Wireless Network Operating System (WINOS) for self-driving wireless networks. On one hand, our proposed S-INT system significantly reduces network measurement overhead by embedding telemetry into flowing data traffic with a specialized packet header. WINOS system, on the other hand, seamlessly integrates programmable measurement, i.e., the proposed S-INT framework, with the programmable network control, while providing rich APIs to facilitate fast implementation of machine learning algorithms for intelligent and distributed network control. To show the effectiveness of our proposed system design, we implemented a multi-agent reinforcement routing as a traffic engineering application to optimize end-to-end delay performance. To the best of our knowledge, our implementation is the first one in the literature that enables multi-agent reinforcement learning algorithm to run on an actual physical wireless multihop network.