Enhancing Resilience and Flexibility of Reconfigurable Data Centre Networks

Objective

The rapid growth of Data Centres is driven by the surge in demand for online storage and Machine Learning (ML) applications. Based on fixed electronic switches and packet-level protocols, traditional data centre networks effectively handle general-purpose, uniform traffic but struggle with bursty, skewed ML workloads. Reconfigurable Data Centre Networks (RDCNs), employing Optical Circuit Switches (OCSs) that can dynamically alter the network topology, have emerged to address this challenge. These flexible, high-speed networks can efficiently adapt to changing traffic patterns, with improved Quality of Service and energy efficiency for advanced digital services. Despite their promise, critical knowledge gaps remain in the scientific literature. ReFlexRDCN (Enhancing Resilience and Flexibility of Reconfigurable Data Centre Networks) aims to address these by targeting two major goals. First, it will engineer advanced tools for rapid identification, localisation, and failover in RDCNs, tackling challenges specific to OCSs. Second, it will design forecasting-driven failure-resilient control algorithms for ML inference requests. By leveraging real-world ML workload traces and advanced forecasting, the project will optimise the allocation and adjustment of network resources, ensuring high efficiency and resilience under dynamic, real-time inference requests. These objectives will be addressed by integrating interdisciplinary methodologies, including network optimisation, optical communications, queuing theory, parallelisation of ML workloads, and validated with simulations and real data. The project outcome is a set of rigorously tested frameworks and algorithms that fill crucial gaps in current research, directly enabling scalable, high-performance, and reliable digital services. These advances will support next-generation ML applications and promote European priorities for trustworthy and efficient digital infrastructures.