Few works about MSO have well coordinated the above factors. Several works [12], [13] study MSO with the assumption that replication points in multicast routing can only occur after the data flow has been processed by all VNFs in the SFC. Although some works [14], [15] optimize MSO by considering the flexible replication that replication points in multicast routing can exist before and after the flow processing of arbitrary VNFs, either resource capacity (e.g., link bandwidth capacity) or the delay requirement is ignored.

As depicted in Fig. 1, with flexible functional split, the RAN is reconstructed into the modular architecture [5], [6], which consists of three modules: centralized units (CUs), distributed units (DUs), and radio units (RUs). The deployment of each module is supported by the infrastructure of cloud and edge computing. The CU, DU, and RU are connected through the midhaul and fronthaul networks respectively. The RAN baseband is decomposed into multiple function units (FUs), such as RF, PHY, MAC, RLC, PDCP, and RRC, which can be virtualized as virtual network function units (VNFUs) and flexibly deployed in the three modules to form a service chain, i.e., a functional split strategy [7]. Functional split enables the RAN to be a scaling network, which can flexibly select splits for specific use cases to provide customized service. Further, flexible functional split selection can relax the bandwidth requirements for fronthaul networks by carefully choosing splits considering resource saving [8].