Machine-Learning-Based Prediction for Resource (Re)allocation in Optical Data Center Networks

Traffic prediction and utilization of past information are essential requirements for intelligent and efficient management of resources, especially in optical data center networks (ODCNs), which serve diverse applications. In this paper, we consider the problem of traffic aggregation in ODCNs by leveraging the predictable or exact knowledge of application-specific information and requirements, such as holding time, bandwidth, traffic history, and latency. As ODCNs serve diverse flows (e.g., long/elephant and short/mice), we utilize machine learning (ML) for prediction of time-varying traffic and connection blocking in ODCNs. Furthermore, with the predicted mean service time, passed time is utilized to estimate the mean residual life (MRL) of an active flow (connection). The MRL information is used for dynamic traffic aggregation while allocating resources to a new connection request. Additionally, blocking rate is predicted for a future time interval based on the predicted traffic and past blocking information, which is used to trigger a spectrum reallocation process (also called defragmentation) to reduce spectrum fragmentation resulting from the dynamic connection setup and tearing-down scenarios. Simulation results show that ML-based prediction and initial setup times (history) of traffic flows can be used to further improve connection blocking and resource utilization in space-division multiplexed ODCNs.