Declarative Traffic Engineering for Low-Latency and Reliable Networking
Jacopo Massa, Stefano Forti, Federica Paganelli, Patrizio Dazzi, Antonio Brogi, Alexander Clemm, Toerless Eckert
TL;DR
The paper targets deterministic end-to-end latency for Cloud-Edge applications by using gLBF, which forwards packets with per-hop budgets and does not require per-flow state at each node. It then introduces dgLBF, a concise Prolog-based declarative control plane that selects end-to-end paths and computes per-hop delays to satisfy latency budgets, while extending with reliability, path protection, and anti-affinity (fate-sharing avoidance). The approach leverages a lightweight knowledge base and precomputed candidate paths to express and enforce constraints, aligning with DetNet/TSN objectives without clock synchronization. Empirical results on the Orion CEV topology and large-scale scalability experiments show quasi-linear growth in placement times up to thousands of flows, with modest CPU and memory usage, indicating practical viability for large deterministic networks.
Abstract
Cloud-Edge applications like industrial control systems and connected vehicles demand stringent end-to-end latency guarantees. Among existing data plane candidate solutions for bounded latency networking, the guaranteed Latency-Based Forwarding (gLBF) approach ensures punctual delivery of traffic flows by managing per-hop delays to meet specific latency targets, while not requiring that per-flow states are maintained at each hop. However, as a forwarding plane mechanism, gLBF does not define the control mechanisms for determining feasible forwarding paths and per-hop latency budgets for packets to fulfil end-to-end latency objectives. In this work, we propose such a control mechanism implemented in Prolog that complies with gLBF specifications, called declarative gLBF (dgLBF). The declarative nature of Prolog allows our prototype to be concise (~120 lines of code) and easy to extend. We show how the core dgLBF implementation is extended to add reliability mechanisms, path protection, and fate-sharing avoidance to enhance fault tolerance and robustness. Finally, we evaluate the system's performance through simulative experiments under different network topologies and with increasing traffic load to simulate saturated network conditions, scaling up to 6000 flows. Our results show a quasi-linear degradation in placement times and system resilience under heavy traffic.