A Machine Learning-Driven Approach for Outbound Traffic Engineering in Edge Networks

Edge networks operated by Internet Service Providers (ISPs) and Content Delivery Networks (CDNs) face significant challenges, e.g., managing traffic flows during periods of high demand, link failures, and organic traffic congestion. Traditional traffic engineering methods that rely on manual adjustments to routing protocols, such as the Border Gateway Protocol (BGP), often fail to address the dynamic and unpredictable nature of network conditions. Although Software-Defined Networking (SDN) offers flexibility through centralized control, it is typically reactive and non-persistent, often failing to simultaneously account for capacity utilization, availability and performance. This paper proposes MLOTE, a machine learning-driven outbound traffic engineering approach tailored for edge networks to mitigate the problems experienced by conventional BGP and SDN. It leverages historical and real-time data to proactively predict traffic demands and dynamically select and adjust the most optimal network paths. Experimental results indicate that MLOTE outperformed the BGP and SDN-based approaches by 23% and 9%, respectively, in terms of peak interface utilization and 91% and 77% in terms of maximum latency reduction.