Building networks that can run themselves, without requiring any human intervention has been the ultimate goal for networking researchers since inception. However, recent advances in the area of the distributed system, machine learning, and networking have provided the technological impetus required to achieve this goal. Our group has been leading and contributing to the development of self-driving networks. Our efforts can be divided into four categories.
- Network streaming analytics. We developed the first network streaming analytics system, Sonata, that lets network operators express network monitoring tasks as flexible dataflow queries and uses reconfigurable data-plane targets (e.g., PISA-based switches) to scale query execution. We are currently improving Sonata by making it more robust to traffic dynamics, compiling more dataflow operators in the data plane (e.g., Join), developing new data structures for compiling stateful operators (e.g., reduce), etc.
- Operationalizing “AI/ML for Networking (NetAI)” applications. We are currently working on operationalizing AI/ML-based solutions to networking problems. Mostly, we are working on problems related to automating network management for enterprise networks and Internet service providers, such as QoE-aware network configurations, in-network pre-filtering for intrusion detection, in-network dynamic firewalls, etc. In most cases, we strive to develop learning models that can make the best use of limited network resources and whose decision making is explainable to network operators.
- Self-driving ISP Networks. We are currently working with our industry collaborators (Verizon, Beegol, etc.) to develop an RL-based system that can detect, diagnose, localize, and fix network issues (e.g., outages) at the last mile (e.g., CMTS, home routers, etc.) as well as core (e.g., packet gateway).
- Democratizing Networking Research. Recognizing the obstacles for networking researchers in pursuing networking research in the era of AI/ML. Currently, our group is developing: (1) a new data collection and analysis pipeline that collects fine-grained networking data at scale while preserving the privacy of end-users for training learning models; (2) a versatile testbed for road testing learning models in realistic settings; and (3) a model-sharing system that lets different networks share learning models with each other without leaking privacy-sensitive information from the training data.