ADAPTIVE SURVIVABLE ROUTING IN LARGE-SCALE WDM OPTICAL NETWORKS USING TRAFFIC-REGIME-CONDITIONED MACHINE LEARNING

Large-scale Wavelength Division Multiplexed (WDM) optical networks are becoming more and more essential to be deployed in heterogeneous and fast-changing traffic regimes induced by cloud interconnection services, elastic data center synchronization and backbone applications with strict latency requirements. The routing and protection in these networks are primarily reliance on static heuristics, shortest-path formulations, fixed-alternate policies and etc. which are mostly designed to highly predictable traffic behaviours and stable operation conditions. But under dynamic traffic regimes, such approaches have low adaptability because of, among other factors, the strong coupling of wavelength contentions, congestion propagation and survivability. This restriction often leads to higher blocking probability, less efficient use of the spectrum and unpredictable restore situation for large scale optical backbone networks. In this paper, to tackle this problem, Traffic Regime Conditioned Survivable Routing Policy Learning (TRC-SRPL) framework for adaptive routing and protection in WDM optical networks is proposed. The framework model problem of routing and survivability decisions as a regime-aware supervised learning problem, where a protection-available wavelength is selected based on the learnt policy that conditions on network state variables like the occupancy status of wavelengths, congestion indicators, path feasibility, protection availability, and characteristics of the traffic regime. Routing-protection decisions generated by Oracle under the two scenarios (static and dynamic) are used to train a deterministic low-latency policy suitable to be deployed within realistic optical control planes. For the proposed framework, large-scale simulation analyses on NSFNET and USNET topologies are carried out under mixed traffic regime with around 48,000 connection requests for both static and dynamic operating scenarios are performed. The results show that the proposed TRC-SRPL policy always performs better than the traditional SRPL policy, SPP and FPR. The proposed framework achieves significantly better blocking probability reduction, mean (spectrum) wavelength utilization, and congestion exposure reduction - up to 26.5%, ~12.3% and ~16%, respectively, under dynamic traffic conditions, while demonstrating stable spectrum fragmentation behavior while keeping the inference latency non-prohibitive in operation. In addition, the learnt policy demonstrates consistent routing behaviour and restoration performance, no matter how large the traffic volume is, without the need of regime-specific re-training. The results obtained show that explicit traffic-regime awareness is indeed a wholesome, non-subsidiary, factor in survivable optical routing. The above adaptive survivable control requirements thus suggests an immediate direction for the proposed TRC-SRPL framework for survivable control in next generation large-scale WDM optical backbone networks.