Communication networks are being restructured by means of network function virtualization (NFV) and service-based architecture (SBA) to embrace greater flexibility, agility, programmability and efficiency. The deployment of service function chains (SFCs) to flexibly offer diverse network services is considered essential in NFV-based networks. Beyond the fifth-generation (5G) and sixth-generation (6G) eras, SFC deployment should be capable of satisfying various quality of service (QoS) requirements, coping with dynamic network states and traffic, handling urgent business in a timely manner, and avoiding resource congestion, all of which present significant scheduling challenges. In this paper, we propose a priority-aware deployment framework for autoscaling and multi-objective SFCs, which mainly includes 2 parts. First, to guarantee the diverse QoS requirements (e.g., latency and request acceptance rate) of various network services, a multi-objective SFC deployment scheme is established to optimize the service latency, deployment cost and service acceptance rate. Second, a deep reinforcement learning (DRL) algorithm, named the autoscaling and priority-aware SFC deployment algorithm (APSD), is further designed to solve the multi-objective optimization problem, which is NP hard. In APSD, we first prioritize requests with varying real-time characteristics to ensure that urgent services can be processed in a timely manner; based on the resiliency characteristics of virtual network functions (VNFs), we propose a hybrid scaling strategy to scale VNFs both horizontally and vertically to respond to changes in service requests and workload. We report comprehensive experiments carried out to assess the effectiveness of the proposed SFC deployment framework and demonstrate its advantages over its counterparts. Thus, we show that APSD is time efficient in solving the multi-objective optimization problem and that the obtained strategy always consumes the least resources (e.g., central processing unit (CPU) and memory resources) and surpasses two baseline algorithms with a 29.5% and 12.36% lower latency on average.
