Due to the high mobility of vehicles, service migration is inevitable in vehicular edge computing (VEC) networks. Frequent service migrations incur prohibitive migration cost including the computing cost (e.g., increased computing delay) and communication cost (e.g., occupied backhaul bandwidth). Yet existing service migration schemes are usually designed without considering the impact of the computing cost. This paper considers the impact of computing and communication cost jointly, and proposes a computing and communication cost-aware service migration scheme for VEC networks (i.e., CA-migration). Taking the service delay as a QoS metric for VEC networks, this paper formulates a migration optimization problem aiming to maximize the services’ satisfaction degree of delay (i.e., the probability that the service delay is smaller than the service delay requirement), where both the communication cost and computing cost affect the services’ satisfaction degree. Since the optimization problem is a constrained non-linear integer programming problem, it is difficult to solve. Moreover, the VEC networks are highly dynamic. Thus, a fast transfer reinforcement learning (fast-TRL) method combining transfer learning and reinforcement learning is proposed to provide an adaptive service migration scheme in dynamic VEC networks. Simulation results show that compared with existing schemes, the proposed CA-migration scheme can increase the satisfaction degree by up to 30%, and needs 25% less training time to obtain the optimal service migration policy.

With the explosive growth of computation-intensive and latency-sensitive vehicular applications, limited on-board computing resources can hardly satisfy these heterogeneous requirements and task offloading becomes a potential solution. However, task offloading in vehicular networks may face the dilemma of unaffordable uploading time caused by the huge amount of uploading traffic. Therefore, considering the applications which use the environmental data as their input, the sensing abilities of serving nodes (SNs) are exploited and a traffic-aware task offloading (TATO) mechanism based on convergence of communication and sensing is proposed. In the TATO mechanism, a task vehicle can adaptively upload the input data to some SNs and transmit the computation instructions to others which use the environmental data sensed by themselves as input. The objective is to minimize the overall response time (ORT) by jointly optimizing the task and wireless bandwidth ratios. Next, a binary search and feasibility check (BSFC) algorithm is designed to solve the optimization problem. Simulation results demonstrate the effectiveness of the proposed BSFC algorithm and show that the TATO mechanism always outperforms the benchmark mechanisms (i.e., communication-based offloading and sensing-based offloading) in terms of the ORT. Specifically, when the task offloading traffic is huge and the wireless transmission capability becomes a bottleneck, TATO can reduce the ORT by 42.8% compared with that of the communication-based offloading.

Federated learning (FL) over mobile networks has attracted intensive attention recently. User mobility is a fundamental feature of mobile networks, which leads to dynamic network topology and wireless connectivity losses. As such, user mobility is usually considered a “trouble maker" and a great challenge to FL over mobile networks. Interestingly, we found that small user mobility can positively contribute to improving FL performance. This is because the total dataset size and the data diversity that the FL can utilize are increased by user mobility. Based on this observation, we aim to tame and exploit mobility instead of treating it as a hostile “trouble maker”. To this end, we first investigate how the FL performance changes with user mobility theoretically by jointly taking into account the positive and negative aspects of mobility. Specifically, a closed-form expression to quantify the impact of mobility on the FL loss is derived, which explains when negative or positive aspects of mobility dominate the FL performance. Next, a joint FL and communication optimization problem is formulated based on theoretical analyses to minimize the FL loss function by optimizing wireless resource allocation. Finally, we propose a two-step optimization algorithm to solve the formulated problem. The simulation results verify the theoretical analyses. It is also shown that the proposed method can significantly enhance learning performance considering users with high mobility. When the average velocity is larger than 150 km/h, the proposed method achieves more than 80% accuracy in the MNIST dataset, while the existing methods may fail during training.

Collaborative mobile edge computing enables resource-constrained edge facilities to work cooperatively for computation-intensive tasks. However, as the number of tasks demanded by various applications increases, resource competition is inevitable in edge facilities. Existing works tackle the resource competition problem with a first come first served (FCFS) scheme, which is blind to different delay requirements among tasks. This may result in tasks with higher delay requirements waiting a long time for service, thereby reducing overall service quality. This paper proposes a prioritized queuing scheme with task dependency (PQTD), which allows high-prioritized sub-tasks with higher delay requirements to jump into the queue ahead of low-prioritized sub-tasks with lower delay requirements. To describe the complicated delay change caused by queue-jumping, a joint DAG-queue delay (JDQD) model is proposed, which analyzes the chain reaction of delay changes caused by the processing queue on the server and the task dependency. With JDQD, a multi-task assignment optimization problem is formulated to maximize the average satisfaction degree (AvgSatD), which is defined according to the priorities of the tasks and their delay requirements. Then, a tree-based algorithm is proposed to solve the NP-hard optimization problem, i.e., Monte Carlo Tree Search (MCTS). Simulation results demonstrate the effectiveness of the PQTD queuing scheme and tree search mechanism of MCTS. Overall, PQTD + MCTS can increase AvgSatD by at least 45.8% with an acceptable complexity.

Federated Learning (FL) empowers wireless intelligent applications, by leveraging distributed data of edge clients for training without compromising privacy. Client selection is inevitable in FL, since clients have diversified learning capabilities arising from heterogeneous computing and communication resources. Existing methods like fair-selection and dropping-straggler are either inefficient or unfair (resulting in a less effective trained model). Therefore, we propose FedAoI, an Age-of-Information (AoI) based client selection policy. FedAoI ensures fairness by allowing all clients, including stragglers, to submit their model updates while maintaining high training efficiency by keeping round completion times short. This trade-off is achieved by minimizing Peak-AoI (PAoI), the interval between a client's consecutive participations. An optimization problem is formulated by minimizing the Expected-Weighted-Sum-of-PAoI. This NP-hard problem is addressed with a two-step sub-optimal algorithm, PriorS. It first calculates client priority in a round using Lyapunov optimization and then selects the highest-priority clients through G-FPFC (Greedy minimization of the round weighted-sum-of-PAoI with First-Priority-First-Considered). Simulation results demonstrate that, compared to fair-selection, FedAoI improves average efficiency by 83.8% and achieves an average model accuracy of 97.3% (or at the cost of averaging 2.7% degradation in model accuracy). Compared to dropping-straggler, FedAoI reduces the average model accuracy degradation from 9.5% to 2.7%.