Performance vs. Cost Tradeoff for Network Slicing in Open RAN: An Intelligent Hierarchical Algorithm for Flexible Utility-Control

The emergence of sophisticated applications and vertical services results in ever more complex mobile networks. Hence radio access network (RAN) slicing based on the traditional private network architecture is struggling to satisfy the resultant diverse quality of service (QoS) requirements. The fledgling next-generation Open RAN is capable of creating a large number of virtual network slices, with the promise of meeting these challenges. However, the existing slicing schemes fail to lend themselves to seamless integration with Open RAN due to offering no customized function splitting capability and owing to having a single time granularity. Therefore, we consider three typical customized RAN slices, namely, the high-throughput slices, the low-delay slices and the wide-coverage slices based upon the same underlying Open RAN. By formulating a utility function, we jointly optimize the operating cost from the perspective of operators, as well as the transmission rate, service delay and coverage from the perspective of users. We propose a joint function-splitting and resource-scheduling (JFSRS) strategy to make optimal deployment decisions for each slice and user under different timescales and resource granularities. Specifically, a hierarchical deep reinforcement learning (DRL) framework is conceived for making local/global decisions. This is achieved by optimization relying on real-time (RT) centralized units (CUs)/distributed units (DUs), on an intelligent near-RT RAN controller, and a non-RT controller. The simulation results verify the utility advantages of our proposed scheme, allowing us to strike compelling tradeoffs between the operating cost and network performance. Compared to the traditional “BBURRH” and “static FS” benchmarks, the network-wide utility of the proposed scheme increases by an average of 17.21%, with a maximum increase of 123%. In contrast to the “coarse granularity” and “single-layer” benchmarks, the utility of the proposed algorithm increases by an average of 23.2%; while the algorithm's training time for the “fine granularity” benchmark is twice that of our scheme.