Performance evaluation of data-driven techniques for the softwarized and agnostic management of an N×N photonic switch

Abstract

The emerging Software Defined Networking (SDN) paradigm paves the way for flexible and automatized management at each layer. The SDN-enabled optical network requires each network element’s software abstraction to enable complete control by the centralized network controller. Nowadays, silicon photonics due to its low energy consumption, low latency, and small footprint is a promising technology for implementing photonic switching topologies, enabling transparent lightpath routing in re-configurable add-drop multiplexers. To this aim, a model for the complete management of photonic switching systems’ control states is fundamental for network control. Typically, photonics-based switches are structured by exploiting the modern technology of Photonic Integrated Circuit (PIC) that enables complex elementary cell structures to be driven individually. Thus PIC switches’ control states are combinations of a large set of elementary controls, and their definition is a challenging task. In this scenario, we propose the use of several data-driven techniques based on Machine Learning (ML) to model the control states of a PIC N×N photonic switch in a completely blind manner. The proposed ML-based techniques are trained and tested in a completely topological and technological agnostic way, and we envision their application in a real-time control plane. The proposed techniques’ scalability and accuracy are validated by considering three different switching topologies: the Honey-Comb Rearrangeable Optical Switch (HCROS), Spanke-Beneš, and the Beneš network. Excellent results in terms of predicting the control states are achieved for all of the considered topologies.