Hecate: AI-driven WAN Traffic Engineering for Science
Mariam Kiran, Scott T. Campbell, Nick Buraglio · 2022
Science network traffic, captured from experiments such as the Large Hadron Collider, are significantly different than general internet traffic in data size, complexity and performance requirements. To deal with this complexity, Research & Education networks (R&E) like ESnet are specifically designed to carry science traffic across the world to other R&E networks, laboratories, and experiments. Designing optimum network topologies, where traffic flow is always efficient with minimal congestion points is imperative to guarantee successful science experimentations. In ESnet, we see a large percentage of long-running flows, mixed in with deadline-driven flows and remote analysis, which makes traffic engineering (TE) particularly challenging. In this paper, we divert from traditional TE approaches and use AI to improve real-time traffic path control such to improve flow quality and network performance proposing a deployable solution, Hecate. Hecate performs a two-stage optimization process, first learning traffic profiles and network health data to predict future statistics, and second, by leveraging deep reinforcement learning to optimize path routing over many optimization objectives. Hecate is designed to optimize network utilization and performance to reduce network hotspots over an operational network.