From Path-Aware to Application-Aware Source Routing using Traffic Classes

Abstract

As one of the most complex challenges in networking, traffic engineering (TE) involves difficult decisions on how to do capacity planning, allocate flows, or even split traffic among multiple paths, in order to minimize congestion. Where creating an optimized solution is cumbersome and difficult, researchers have turned to AI/ML techniques to enable the network to learn based on traffic characteristics and network patterns, making optimal decisions. SDN-based TE tools, such as Teal, Hecate, and B4, among others, have utilized classification or deep reinforcement learning to identify optimal TE solutions. The additional advent of source routing tools, such as PolKA, has introduced the capability for finer-grained control of flows. In this paper, we explore the concept of traffic classification at the edge, enabling source routing to perform flow allocation based on incoming flow requests. We develop new traffic class patterns and queuing methods to ensure fair queuing allocation, thereby satisfying all traffic classes and achieving an optimized network utilization and allocation scheme. This work contributes to in-band network telemetry (INT) and P4 programmability, demonstrating how source routing can enable finer-grained decision-making for self-engineering networks.