Network function virtualization aims to deploy virtual network functions (VNFs) on commercial off-the-shelf hardware, allowing for efficient implementation of network services by scheduling various VNFs. However, the requirements of deterministic networking poses higher demands on VNF scheduling to achieve guaranteed latency performance. In this paper, we study the online VNF scheduling problem with the objective of minimizing the system cost regarding to unfinished services while improving the determinism of latency. The formulated optimization task yields a mixed integer program problem and we propose an online VNF scheduling algorithm to tackle it, which transforms the end-to-end latency constraint into deadlines for each VNF and adopts an online primal-dual method to generate scheduling decisions. Numerical results show that the proposed scheduling algorithm outperforms two traditional rule-based scheduling methods in terms of latency and acceptance ratio.

Network function virtualization redefines a network service as a softwarized chain of virtual network functions (VNFs), which decouples the specific network service from dedicated devices and greatly reduces the hardware cost. The VNFs are generally deployed on common off-the-shelf servers and statistically share computational resources. Due to the inherent integer constraints, the corresponding VNF mapping and scheduling issue is a highly challenging task. In this paper, we consider the mapping and scheduling of VNFs for a given network service, for which a flexible job shop scheduling problem is formulated to optimize the max-min fairness while ensuring the delay requirements of different service chains. Specifically, we propose a deep reinforcement learning method based on offline proximal policy optimization, which dynamically determines the mapping and scheduling decision based on the state of unfinished service chains. The proposed algorithm is scalable to the number of service chains and can be enhanced by Monte Carlo tree search. Numerical results show that the proposed algorithm outperforms the traditional random forest and the greedy algorithms in terms of both service fairness and acceptance ratio.

Network function virtualization aims at deploying network functions on general-purpose hardwares, referred to as virtual network functions (VNFs), rather than the specialized devices. A network service can be implemented by scheduling different VNFs. In this paper, we study the VNF scheduling problem with the objective of satisfying the diversified end-to-end delay requirements while maintaining the fairness among different network services. We formulate the problem as a mixed integer nonlinear program and propose a VNF scheduling algorithm based on deep reinforcement learning, in which proximal policy optimization is adopted to optimize the policy network. Numerical results show that the proposed scheduling algorithm outperforms other canonical ones and the designed policy network can scale to fit different problem sizes.

Antenna selection (AS) is a promising technology that can reduce the implementation complexity and hardware cost of a massive MIMO system, in which a part of the attainable antennas are selected and connected to the radio frequency chains in each time slot. In this article, we propose a low-complexity AS algorithm to maximize the downlink achievable rate in the time-division duplexing massive MIMO system, which is based on online Thompson sampling technique and significantly reduces the pilot overhead required for channel estimation with only partial channel state information. We prove that the distribution-dependent upper bound of the proposed algorithm is sub-linear as a function of time slot by introducing the concept of regret. We also develop three discounting factors to accommodate the large-scale variations across antennas. Numerical results show that the downlink achievable rate can be greatly improved with our proposed scheme as compared to other typical ones.