O-RAN-Enabled Collaborative Multi-Access Edge Computing Over Optical Metro and Access Networks for Enhanced Load Balancing

The open radio access network (O-RAN) and multi-access edge computing (MEC) are promising techniques for 6G cellular networks. O-RAN promotes the evolution of cellular ecosystem through function disaggregation and artificial intelligence (AI) controllers, while MEC enables latency-critical tasks offloading at the network edge. However, the finite number of edge servers (ESs) and limited computational resources may lead to resource shortages, particularly in hotspot areas with dense MEC activities. Our work considers an O-RAN-enabled collaborative MEC (C-MEC) network that employs passive optical network (PON)-based metro-access network to establish virtualized connections between ESs, while utilizing AI to enable intelligent RAN/C-MEC resource management. With the objective of optimizing hotspot throughput, we propose a function splitting (FS)-based load balancing (<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">FS-LL</i>) problem to jointly schedule RAN slicing and manage C-MEC traffic migration. We formulate <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">FS-LL</i> problem as an integer linear programming (ILP) model considering capacity constraints of computing and bandwidth resources, while satisfying latency requirements. Furthermore, we develop a greedy-search-assisted multi-agent deep reinforcement learning (<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">GA-MADRL</i>) algorithm, where each ES cluster is equipped with a dedicated agent to orchestrate traffic for the implementation of C-MEC. The results demonstrate that <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">GA-MADRL</i> algorithm can obtain close-to-optimal solutions compared to the ILP formulation, while also achieving higher hotspot throughput and better workload balancing performance than benchmark algorithms and traditional FS methods.