LASER: Learning Enhanced Segment Routing Using Link Telemetry

Murali S. Kodialam, T. V. Lakshman

Abstract: This paper presents a novel approach to telemetry-based routing, aiming to minimize network congestion using multiple link load measurements collected at different points in time at a centralized management system. The objective is to determine a fixed routing policy that optimizes network performance by minimizing the maximum link utilization for any traffic matrix that could have generated any link load measurement. The key idea is to develop a routing mechanism that has the flexibility to handle a wide variety of traffic conditions without reconfiguration. We use a routing mechanism called deflection routing and develop a machine learning based gradient algorithm (LASER) to compute the deflection routing parameters. We use a combination of variable transformation and Lagrangian based techniques to transform the parameter optimization problem into an unconstrained loss minimization problem which is solved using a structured neural network in the PyTorch framework.