Recently open radio access network (Open RAN), which splits baseband functions into multiple process units at different locations has received considerable attentions from both industries and academia with the potential to enable a fully disaggregated RAN with more flexibility in delivering energy-saving and latency-sensitive applications. However, the significant increases of resource usage dynamics in both geographical and time and network complexity may lead to unnecessary high energy consumption in RANs without an efficient RAN function management policy. Many studies have proposed baseband function management solutions, however, the activation cost and data network resources of edge computing capacities have not been evaluated in much detail, as far as the authors know. In this paper, with the objective of minimizing energy consumption, meanwhile, satisfying the requests over the network under the constraints of latency and resource capacity, we propose a completed mixed integer linear programming (MILP) formulation, a multi-agent deep reinforcement learning-based algorithm and a heuristic (DCUH), to take user plane functions (UPFs) on the multi-access edge computing servers (MECs) and the activation consumption of MECs into consideration. Moreover, we prototype an OpenDaylight, OpenStack and Open Source Management and Orchestration-based Open RAN testbed to verify the feasibility of the proposed solutions. Results show the importance of hibernating the MEC after a certain time of network vacancy. DRL-based algorithm and DCUH can approach a similar performance as the benchmark of MILP and save more than 40% energy consumption compared to the first-fit algorithm. This study offers an important insight into the design of baseband deployment policies that greatly enhance user experience with better service and save Open RAN operational energy costs.

翻译：最近开放的无线电接入网络(Open RAN)将基带功能分成不同地点的多个进程单位,但各行业和学术界都十分关注这一网络,因为这一网络有可能实现完全分解的RAN,在提供节能和延时敏感应用方面具有更大的灵活性,然而,在地理和时间复杂和网络复杂的情况下,资源使用动态大幅度增加,可能导致在没有有效的 RAN 功能管理政策的情况下,RAN 网络的能源消耗不必要地高。许多研究提出了基带功能管理解决方案,然而,就作者所知,边缘计算能力的启动成本和数据网络资源尚未得到详细评估。在本文中,为了尽可能减少能源消耗量,同时满足网络对能源消耗量的要求,在弹性和资源能力的限制下,我们提出了完成的混合整数线性编程(MILP)配制、多试用深度基于学习的算法和高压(DCUH),在多接入端计算服务器(MEC)的启动成本和MEC启动的消费。此外,我们设计了On-SDRA公司在高额成本之后的O-RRL 测试系统后,可以提高某些成本。