NFScaler: AI-Powered 5G-and-Beyond Network Function Scaler for QoS Assurance and Energy Efficiency

Efficient resource management and orchestration are essential to maximise the benefits of network slicing in 5G and beyond networks. This paper presents NFScaler, an Artificial Intelligence (AI)-powered 5G network function scaler that can perform a zero-shot sim2real transfer. The proposed NFScaler automatically adjusts the number of 5G network function instances running in a network slice based on changes in user traffic to meet the QoS requirement of the slice while reducing the resource consumption (i.e. CPU and power). The NFScaler consists of two main parts: a domain randomiser and a deepreinforcement learning (DRL) agent. The domain randomiser randomises the dynamic parameters of a simulation to provide the DRL agent with a range of simulation environments. The DRL agent interacts with both the original simulation and the randomised simulation to learn a generalisable autoscaling policy. The performance of the proposed NFScaler is evaluated in a live 5G network and compared with the threshold-based KEDA scaler, which is the industry standard event-based autoscaler used in cloud native environments, as well as a DRL method that does not involve sim2real transfer. The experimental findings demonstrate that NFScaler effectively guarantees the QoS requirements of the network slice, outperforming the benchmark methods. In particular, NFScaler achieves an improvement of almost 40% in QoS (throughput) performance compared to KEDA Scaler. Furthermore, the proposed NFScaler allocates a significantly lower number of CPU cores to the user plane of the network slice, with an average of 75% fewer cores than when the user plane is hosted on a dedicated bare metal server. In addition, the NFScaler reduces the power consumption of the user plane of the network slice by an average of 45% compared to hosting the user plane on a dedicated bare metal server.