To tackle the difficulties of running sophisticated deep neural network (DNN) models on mobile devices, split computing presents a viable solution by offloading computations to the edge server. Current split computing schemes typically aim to lower either inference latency or energy use separately; however, optimizing both simultaneously is quite challenging due to numerous shifting factors, such as intensive continuous DNN model inferences, DNN model traits, and device/network conditions. Moreover, in practical applications, edge server overload might lead to substantial queuing delays, adding complexity to the optimization process. This paper outlines a joint optimization problem that simultaneously seeks to minimize both inference latency and energy consumption, with a distinct inclusion of queue clearance latency for an accurate analysis of the continuously generated DNN model inferences. To address this intricate optimization challenge, we introduce a low-complexity heuristic algorithm that sets split point decisions based on the residual energy of mobile devices for each DNN inference cycle. Upon evaluation, our proposed algorithm demonstrates notable improvements by reducing inference latency by between$73.37\%$and$99.39\%$, and cutting down energy usage by between$39.97\%$and$94.67\%$compared to fully local processing on mobile devices.

Current 5G networks suffer from high signaling overhead due to highly mobile vehicles. In this paper, we first introduce a personalized handover function (denoted μ HF) that consolidates all handover-related functionalities for individual mobile devices (MDs). Recognizing that the location of μ HF affects overall handover performance , we propose a mobility-aware μ HF provisioning system (MA- μ HFPS), which utilizes a central controller to collect mobility information for each MD and provisions μ HFs in the edge cloud for MDs that are expected to have high mobility for a long time. To minimize signaling overhead and migration cost for handover-related information from the central cloud to the edge cloud while ensuring that the average required resource of the edge cloud remains below a specific threshold, we formulate a constrained Markov decision process (CMDP) problem. By converting the CMDP problem into a linear programming (LP) model, we can achieve an optimal stochastic policy using a traditional algorithm with low complexity. Evaluation results demonstrate that MA- μ HFPS significantly reduces signaling overhead with a small state migration cost compared to the traditional handover management system .

In the conventional split computing approach based on the external computing node (e.g., cloud), Internet of Things (IoT) devices suffer from high network latency. In this article, we introduce an unmanned aerial vehicle (UAV)-assisted split computing system (USCS) where UAV patrols around the IoT device and IoT device offloads performing the tail model inference to UAV. To minimize the energy consumption while maintaining a sufficiently low inference completion time, IoT device makes two types of decisions: 1) the timing of starting the split computing (i.e., whether to conduct the split computing or delay) and 2) the splitting point. By formulating a constrained Markov decision process (CMDP) problem and converting the CMDP model into a linear programming (LP) model, the decisions of the IoT device can be optimized. The evaluation results show that the USCS can significantly reduce energy consumption while satisfying the inference completion time requirement.

In the distributed task offloading system, the desired task completion time cannot be achieved when lots of mobile devices offload simultaneously the tasks with high complexity to a specific edge cloud. In this research, we present a neighbor-aware distributed task offloading algorithm (NA-DTOA) where the IoT device takes its energy status and the decisions of neighbor devices into account for the decisions on whether to process the task by itself and on which edge cloud is selected to offload the task. To decrease the task completion time while maintaining the average energy outage probability below the specified threshold, we design a constrained stochastic game model. To achieve the solution of the model, a best response dynamics-based algorithm is devised. The evaluation results reveal that, compared to a probability-based scheme, NA-DTOA reduces the average task completion time by almost 47% while ensuring a substantially low-average energy outage probability of 0.03.

In federated learning (FL), which clients and quantization levels are selected for the deep model parameters has a significant impact on learning time as well as learning accuracy. This is not a trivial issue because it is also significantly affected by factors, such as computational power, communication capacity, and data distribution. Considering these factors, we formulate a joint optimization problem for clustering and selecting clusters with quantization levels. Due to the high complexity of the formulated problem, we propose a situation-aware cluster and quantization level selection (SITUA-CQ) algorithm. In this algorithm, the FL server first assembles clients into clusters to mitigate the impact of biased data distributions and determines the most suitable clusters and quantization levels based on their computing power and channel quality. Extensive simulation results show that SITUA-CQ can reduce the round time by up to 80.3% compared to conventional algorithms.

In federated learning (FL), if the participating mobile devices have low computing power and poor wireless channel conditions and/or they do not have sufficient data for various classes, a long convergence time is required to achieve the desired model accuracy. To address this problem, we first formulate a constrained Markov decision process (CMDP) problem that aims to minimize the average time of rounds while maintaining the numbers of trained data and trained data classes above certain numbers. To obtain the optimal scheduling policy, the formulated CMDP problem is converted into an equivalent linear programming (LP). Additionally, to overcome the problem of the curse of dimensionality in CMDP, we develop a joint client selection and bandwidth allocation algorithm (J-CSBA) that jointly selects appropriate mobile devices and allocates suitable amount of bandwidth to them at each round by considering their data information, computing power, and channel gain. Evaluation results validate that J-CSBA can reduce the convergence time by up to$49\%$compared to a conventional random scheme.

Owing to the advantages of satisfying service requirements and providing strong security, non-public networks (NPNs) are considered as a promising technology in vertical industries. However, to efficiently manage cloud-native network functions (CNFs) in NPNs, a sophisticated control plane management scheme should be designed. In this paper, we propose a deep Q-network-based CNF placement algorithm (DQN-CNFPA) that jointly minimizes the costs incurred by launching and operating CNFs in edge clouds and the backhaul control traffic overhead. In addition, DQN-CNFPA learns the spatiotemporal patterns in service requests and adaptively places CNFs in edge clouds according to the expected incurred costs. The evaluation results demonstrate that DQN-CNFPA can reduce the total cost by up to 26.2% compared with a conventional scheme that does not learn spatiotemporal service request patterns.

In-network aggregation facilitates accelerated distributed deep learning by utilizing a programmable switch to aggregate gradient packets. However, a straggler problem should be addressed to avoid performance degradation in terms of training time. In this paper, we propose a straggler-aware in-network aggregation (SAINA) scheme to mitigate the straggler problem while preventing accuracy degradation. In SAINA, the programmable switch aggregates local gradients of the fastest$k$workers to exclude stragglers and changes$k$adaptively to balance the tradeoff between training speed and accuracy. To this end, we design a switch-friendly convergence detection (SFCD) algorithm which detects a convergence point and determines$k$at the convergence point. SAINA is implemented over a software programmable switch and experimental results show that the accuracy of SAINA can reach a target accuracy up to 2.84x faster than the existing in-network aggregation scheme.

In federated learning (FL), which clients are selected and which quantization levels are chosen for the deep model parameters have significant impacts on the learning time as well as the learning accuracy. In this paper, we formulate a joint optimization problem on the client and quantization level selections. As a low complexity solution to the formulated problem, we develop a performance-aware client and quantization level selection (PA-CQLS) algorithm where the FL server estimates the individual round times of clients based on their computing power and channel quality, and determines the most appropriate clients and quantization levels accordingly. Simulation results show that PA-CQLS can reduce the round time by up to 70% compared to conventional algorithms.

To support network slicing and service function chaining (SFC) at a time, we propose a novel hierarchical identifier (HID)-based 5G architecture. For this, we first introduce HID which consists of network slice selection assistance information (NSSAI) and service path ID (SPI). Based on HID, a user can attach a specific network slice and flows generated by the user can be processed by a set of service functions (SFs) in a sequence. Meanwhile, when the incoming flow to a specific slice unexpectedly increases, the slice cannot handle incoming flow due to its limited capacity, which degrades users' quality of service (QoS). To alleviate this issue and efficiently utilize network resources, we introduce a concept of the backup slice shared by different services. Evaluation results demonstrate that the proposed architecture can achieve better performance in terms of the average system blocking probability and utilization in dynamic environments.

To overcome the limitation of standalone edge cloud in terms of computing power and resource, a concept of distributed edge cloud has been introduced, where application tasks are distributed to multiple edge clouds for collaborative processing. To maximize the effectiveness of the distributed edge cloud, we formulate an optimization problem of task allocation to minimize the application completion time. To mitigate high complexity overhead in the formulated problem, we devise a low-complexity heuristic algorithm called dependency-aware task allocation (DATA) algorithm. Evaluation results demonstrate that DATA can reduce the application completion time up to by 15%-32% compared to conventional dependency-unaware task allocation schemes.

In vehicular edge clouds, tasks from vehicles are processed nearby edge nodes (ENs) and thus low latency services can be provided. However, under high vehicular mobility, frequent service migration between two ENs and increased handover latency can be observed. In this paper, we introduce a trajectory-aware edge node clustering (TENC) scheme in which multiple ENs form a cluster depending on the trajectory of a target vehicle. To attain the optimal performance, we formulate an optimization problem by means of a constrained Markov decision process (CMDP). Evaluation results demonstrate that the obtained optimal policy can minimize service delay significantly.

To overcome the limitation of standalone edge computing in terms of computing power and resource, a concept of distributed edge computing has been introduced, where application tasks are distributed to multiple edge clouds for collaborative processing. To maximize the effectiveness of the distributed edge computing, we formulate an optimization problem of task allocation minimizing the application completion time. To mitigate high complexity overhead in the formulated problem, we devise a low-complexity heuristic algorithm called dependency-aware task allocation algorithm (DATA). Evaluation results demonstrate that DATA can reduce the completion time up to by 18% compared to conventional dependency-unaware task allocation schemes.

Service function chaining (SFC) is an emerging technique that provides steering of traffic flows through an ordered set of service functions (SFs). In SFC, high availability is one of the most important issues to be addressed. SF instances within the chain can become unavailable when SF instances are overloaded or failed (e.g., power outage). Therefore, in order to realize highly available SFC, load balancing and fault management for SF instances must be provided. In this paper, we propose an overload and failure management (OFM) module in SFC that consists of the overload management (OM) module and the failure management (FM) module. In the OM module, when the current load at an SF instance exceeds a low-level threshold, a backup SF instance is prepared in advance. Meanwhile, if the current load further exceeds a high-level threshold, flow migration from the current SF instance to the backup SF instance is triggered. The FM module detects the failure of the SF instance by using a failure alarm. Upon detecting the failure, flow migration to the backup SF instance is triggered. We implement the OFM module in OpenDaylight (ODL) and present the experimental validation results.

Recently, femtocells are widely deployed to offload the traffic from the macrocell. Since conventional location management schemes of femto/macrocell networks do not consider mobility pattern of the mobile node (MN), unnecessary location updates can occur. Specifically, when an MN moves along the contour of the femtocell coverage, the MN frequently executes location update procedures, which causes significant location update cost. To address this problem, we propose a mobility-aware location management (MALM) scheme, where the MN conducts location update only at specific femtocells in which it is expected that the MN stays for a longtime. To optimize MALM, a Markov decision process (MDP) problem is formulated and the optimal policy is determined. Evaluation results demonstrate that MALM with the optimal policy can reduce the number of location updates while providing sufficient offloading gain.

Article synchronization is one of the most fundamental issues in online social networking services (SNSs). In particular, when deploying multiple access networks with different transmission costs, an efficient synchronization scheme should be devised for users to enjoy SNSs in heterogeneous wireless networks. In this paper, we propose an opportunistic push scheme (OPS) that aggregates published articles, opportunistically pushing them through low-cost access networks (e.g., open Wi-Fi networks). To balance the reduced transmission costs and the increased page loading time, we formulate a Markov decision process problem that considers a mobility model based on the users' social contact pattern. Evaluation results demonstrate that OPS with the optimal policy can reduce the number of transmissions in high-cost access networks, while satisfying users' quality of experience in terms of the page loading time.