Low-latency computational-task execution can be achieved by leveraging device-to-device offloading and parallel processing over nearby extreme edge devices (EEDs), a paradigm known as extreme edge computing (EEC). However, EEC performance is challenged by device spatial randomness with intermittent wireless connectivity, limited device computing power, time-varying availability, and device failures. This paper introduces a novel spatiotemporal analytical framework for EEC by integrating stochastic geometry with an absorbing continuous-time Markov chain (ACTMC) to capture the interplay between communication and computation. Modeling a large-scale millimeter-wave network, we derive tractable expressions for the average task response delay and the task completion probability under both random and location-aware EED selection. Numerical results quantify the impact of location-awareness and unveil the existence of an optimal task segmentation that minimizes delay, which depends on network parameters and EED capabilities. We also demonstrate that device failures and EED scarcity exacerbate delay, which can be mitigated through a collaborative load-balancing approach between EEC and Multi-Access Edge Computing (MEC) schemes. Simulations and sensitivity analyses validate the proposed framework and offer design insights for optimizing system performance.

Task offloading in Vehicular Edge Computing (VEC) can significantly enhance Cooperative Perception (CP) for Autonomous Vehicles (AVs), improving situational awareness and traffic safety. However, the widespread adoption of VEC is often constrained by the high deployment costs of Roadside Units (RSUs). In this paper, we propose the Truthful and Quality-Aware Task Offloading (TQTO) scheme. TQTO leverages the prolific yet underutilized computational resources of parked vehicles for CP tasks in VEC to alleviate RSU scarcity. Using vehicle-to-vehicle (V2V) communication, parked vehicles can be strategically involved in CP processing and are incentivized to contribute their resources. TQTO introduces a distributed, truthful, double-auction-based multi-agent deep reinforcement learning framework that enables user vehicles to offload CP tasks to parked vehicles in a utility-maximizing manner, while respecting their individual budget constraints. Concurrently, TQTO considers the provider-side (i.e., parked vehicles) costs and ensures a truthful, incentive-compatible, and budget-balanced marketplace for VEC. A critical value-based payment mechanism is used to ensure fair compensation for parked vehicles and to align task requesters’ payments with their utility. TQTO formulates the task offloading problem as a Double Auction Quadratic Multiple Knapsack Problem (DA-QMKP) and solves it using a QMIX-based heuristic for scalable decision-making under partial observability. Extensive evaluations show that TQTO outperforms a prominent non-auction-based scheme by up to 39% in terms of social welfare.

Task offloading in Vehicular Edge Computing (VEC) enhances cooperative perception (CP) for Autonomous Vehicles (AVs), improving traffic awareness. However, the high cost of Roadside Unit (RSU) and the underutilization of parked vehicles pose challenges. Leveraging Vehicle-to-Vehicle (V2V) communication, parked vehicles can form collaborative worker groups for efficient perception aggregation. We propose CoGroup, a two-tier framework integrating task offloading and dynamic worker grouping. Modeled as a double quadratic multiple knapsack problem, it employs Hierarchical Reinforcement Learning (HRL): QMIX for decentralized task allocation and DQN for optimized worker grouping. Experiments show that CoGroup improves traffic awareness by 21% over non-cooperative methods, reducing RSU dependence and offering a scalable, cost-effective solution for next-generation VEC systems.

Democratizing the edge by capitalizing the underutilized computational resources of end devices, referred to as Extreme Edge Devices (EEDs), can foster various IoT applications. In this paper, we propose the Community Edge Platform (CEP). CEP is the first platform that exploits business, institutional, and social relationships to build communities of requesters and EEDs to eliminate recruitment costs and preserve privacy in EED-enabled environments. CEP promotes service-for-service exchange and utilizes a hierarchical control paradigm to prioritize the enrollment of nearby devices as workers. CEP also considers the fact that community-imposed constraints can lead to unbalanced work distribution. To alleviate this issue, we propose the Community-Oriented Resource Allocation (CORA) scheme. CORA accounts for community restrictions and strives to minimize the execution time and makespan while retaining a reasonable scheduler runtime. Towards that end, we formulate the resource allocation problem as a Bipartite Graph Matching problem. Comprehensive qualitative evaluations demonstrate the superiority of CEP compared to 12 prominent edge computing platforms in terms of various system architecture and performance features. Additionally, extensive simulations show that CORA outperforms six prominent resource allocation schemes by up to 44% and 7% in terms of makespan and execution time, respectively, while achieving a much faster runtime, outperforming the best of the six baseline resource allocation schemes by a factor of six.

Extreme Edge Computing (EEC) exploits the untapped computational power of end devices, referred to as Extreme Edge Devices (EEDs), and thus holds the potential to revolutionize the Digital Twin (DT) technology. However, traditional reactive task allocation approaches fail to address the complexities of DT processing tasks, where the execution of all the underlying subtasks is crucial. Additionally, these approaches suffer due to the intermittent availability of EEDs, which compromises the Quality of Service (QoS). In this paper, we propose the Proactive Maximum Weighted Service Capacity (P-MWSC) scheme. P-MWSC is the first scheme to employ a proactive approach, utilizing predictions of the dynamic resource usage and resource characterization of EEDs to tackle the intricacies of DT tasks while taming the effects of EEDs’ dynamicity and intermittent availability. We formulate the task allocation problem as a Binary Integer Linear Program (BILP) that aims to maximize the service capacity, weighted by the achieved gain from each fully assigned task. We derive an analytical solution using the Karush–Kuhn–Tucker (KKT) conditions and Lagrangian relaxation, and use a top-down decomposition approach to provide a solution that achieves up to 80% runtime reduction. Additionally, we propose a heuristic scheme with a bottom-up decomposition approach that is suitable for certain practical scenarios, yielding up to 90% runtime reduction. Extensive performance evaluations using data from a realistic testbed demonstrate that P-MWSC outperforms representatives of prominent reactive and proactive schemes, achieving up to 70% increase in the task success rate and a 39% reduction in the average response delay.

Mobile edge computing (MEC) is a promising paradigm for Internet of Things applications requiring synchronized user experiences. However, sustaining scalable and reliable MEC services is challenging when computational resources are overloaded, especially as MEC service providers (SPs) must minimize operational costs to maximize profits while offering competitively priced services. This article proposes the cooperative multiprovider market (CMPM) scheme, the first to cooperatively enhance service scalability and reliability while addressing the profit-pricing dilemma in a multiprovider market. CMPM enables overloaded home SPs (HSPs) to leverage underutilized computational resources from reliable foreign SPs (FSPs) via reputation-based service replication, meeting the stringent Quality of Service (QoS) requirements for real-time applications involving user groups. CMPM resolves the pricing dilemma by applying a game-theoretic approach, allowing FSPs to dynamically optimize revenue and adjust prices when HSPs cannot meet user demand. We formulate the resource allocation and pricing problem as a Stackelberg game, establish the existence of the equilibrium, and develop a distributed algorithm to reach it. Extensive evaluations show that CMPM significantly reduces unit prices, attracts more HSPs, and better manages high-density user loads compared to state-of-the-art schemes that overlook SP reputation and social welfare. CMPM also achieves up to 84% higher FSP revenue, a 67% improvement in scalability, and a 70% higher task success rate compared to baseline schemes.

Task offloading in Vehicular Edge Computing (VEC) is crucial for enhancing cooperative perception (CP) in Autonomous Vehicles (AVs), thereby improving traffic situational awareness. However, existing approaches often neglect the balance between high-quality execution of interdependent tasks and conserving AVs limited budget, including communication and financial resources. To address this, we propose the Quality and Budget-Aware Task Offloading (QBATO) framework. QBATO is the first framework to balance the quality of cooperative perception with budget conservation. QBATO models the budget as a queue to ensure stability, balancing resource use while prioritizing situational awareness in CP. Additionally, QBATO enhances CP quality by predicting vehicles movements and estimating their regions of interest, thereby improving the Value of Information (VOI). The task offloading problem is modeled as a Quadratic Multiple Knapsack Problem (QMKP), an NPhard problem that optimizes vehicle allocation by evaluating the quality of assigning multiple vehicles to the same worker through a quadratic objective function.To manage resources effectively, we apply the queue stability Lyapunov drift-minus-bonus approach. We also introduce the QBATO-Heuristic (QBATO-H), which solves the problem in a decentralized, time-efficient manner using a multi-agent deep reinforcement learning technique that leverages the Q-Mixing Network (QMIX) method, which employs monotonic value decomposition to coordinate the actions of multiple agents. Extensive evaluations show that QBATO outperforms prominent quality and budget-oblivious schemes by up to 49%, 15%, and 35% in terms of budget conservation, situational awareness, and efficiency, respectively. QBATO-H also yields a small gap of up to 7% and 11% with QBATO in terms of budget conservation and efficiency, respectively.

Democratizing edge computing (EC) by harnessing the underused computational resources of extreme edge devices (EEDs) can revolutionize a broad spectrum of Internet of Things applications. However, studying the impact of EED/worker attrition on the Quality of Service (QoS) at the extreme edge has been overlooked. In this article, we propose the multitiered worker-oriented resource allocation (MWORA) scheme. MWORA is the first scheme that studies the impact of worker attrition on the QoS and mitigates its risk by optimizing resource allocation to ensure that workers receive a satisfactory profit to maintain their participation in the service. Furthermore, MWORA accounts for the fact that EEDs are user-owned devices, and are thus subject to a dynamic user access behavior, which can affect the level of computational resources they are willing to endow. MWORA accounts for such dynamicity by pioneering the notion of enabling multitiered computational capabilities to be solicited from each worker based on the profit gained from the assigned tasks. We formulate the problem as an integer linear program (ILP) to maximize the QoS while abiding by certain worker satisfaction, deadline, and budget constraints. We also propose the MWORA-weighted sum (MWORA-WS) scheme to derive an analytical solution using the Karush–Kuhn–Tucker (KKT) conditions and Lagrangian analysis. Extensive simulations show that MWORA outperforms prominent resource allocation schemes by up to 31%, 52%, and 85% in terms of response delay, service capacity, and worker satisfaction, respectively. Additionally, MWORA-WS yields a small gap of 7% and 8% with MWORA in terms of response delay and worker satisfaction, respectively.

Mobile edge computing (MEC) has emanated as a propitious computing paradigm that can foster delay-sensitive and/or data-intensive applications. However, it can be challenging to maintain a scalable MEC service when computational resources are overloaded. In this article, we propose the service replication between multiple service providers (SRMSPs) scheme. SRMSP is the first scheme that fosters service scalability in a cost-efficient manner, while considering the stringent QoS requirements of real-time applications involving groups of users. SRMSP enables SRMSPs to minimize the average response delay and the operational cost incurred by service providers, while satisfying the delay requirements of all user groups. We formulate the resource allocation problem as an integer linear program (ILP) and derive an analytical solution using the Karush–Kuhn–Tucker (KKT) conditions and Lagrangian analysis. In addition, we propose the SRMSP-distributed allocation (SRMSP-DA) scheme to provide a time-efficient solution in distributed scenarios. In SRMSP-DA, we use a game-theoretic strategy that formulates the resource allocation problem as a potential game. Extensive simulations show that SRMSP renders a 50% operational cost reduction compared to a baseline scheme that does not consider the operational cost. In addition, SRMSP-DA exhibits a relatively marginal difference of up to 20% and 4% in terms of the total operational cost and average response delay, respectively, compared to the optimal solution provided by SRMSP.

Edge computing can satisfy the requirements of latency-critical and data-intensive applications by exploiting com-putational resources of end devices. However, such devices inherently suffer from dynamic user behavior, cyclic task-switching, varying link qualities which often impact their reliability. In addition, in incentivized systems, they may often over-estimate their advertised capabilities and consequently fail on delivering. In this paper, we propose the Reputation-based Task Assignment and Replication (RTAR) scheme. RTAR is the first scheme that uses a black box approach to perform cost-efficient task replication that accounts for workers' reliability and preserves workers' privacy by not requiring or soliciting any information about their devices. RTAR incorporates a reputation model using beta distribution to estimate the worker's reputation based on past performance. We formulate the problem as an Integer Linear Program (ILP) that strives to maximize the overall reputation of recruited workers, while abiding by a certain budget limit for each task. We also propose the RTAR-Heuristic (RTAR-H) scheme. RTAR-H uses matching theory to solve the optimization problem in a time-efficient manner. Extensive evaluations show that RTAR yields 63% and 68% reduction in recruitment cost and number of replicas, respectively, compared to a baseline scheme that blindly maximizes the number of replicas. Moreover, RTAR-H closely approaches the optimal solution, rendering a small gap of up to 1% and 1.2% in terms of task drop rate and recruitment cost, respectively.

Extreme Edge Computing (EEC) can drastically curtail the delay, reduce network bandwidth consumption, and enhance system performance by providing computing resources closer to the data-generating Internet of Things (IoT) devices. However, the use of Extreme Edge Devices (EEDs) in EEC presents unique challenges imposed by the inherent dynamic user-access behavior, which introduces highly dynamic resource usage. To tackle such challenges, it is crucial to enable accurate resource usage predictions, which in turn requires having reliable datasets. In this paper, we cultivate the Dynamic Resource Usage Data Generation for EEDs (DRUDGE) methodology. DRUDGE generates datasets that capture the resource usage dynamics of EEDs running diverse user-end applications in fine-grained intervals over extended periods. We present an in-depth characterization of resource utilization in EEDs and make the datasets publicly available to the research community. We examine the temporal variation of critical system metrics, such as CPU usage, memory usage, temperature, and network traffic. Furthermore, we apply various statistical tests to gain valuable insights into the data characteristics, including skewness, kurtosis, stationarity, volatility, cointegration, multi-collinearity, Granger causality, and Pearson correlation analysis. These insights inform model selection, feature engineering, and preprocessing techniques, leading to more accurate and reliable forecasts and analyses for EEC systems.

Leveraging the copious yet underutilized computational resources of end devices, also known as Extreme Edge Devices (EEDs), can significantly enhance the performance of various Internet of Things (IoT) applications. However, EEDs are heterogeneous and user-owned devices, which causes their availability to be highly unreliable. In this paper, we propose the Dynamic Worker Availability Prediction (DWAP) scheme. DWAP is the first scheme that predicts the availability of EEDs (i.e., workers) and adapts to the highly dynamic computing environment at the extreme edge. DWAP employs the Continuous-Time Markov Model (CTMC) to forecast the availability of workers in the upcoming time step. It does so while continuously fine-tuning the model parameters to incorporate newly available data. We use a dataset that consists of real-world Google cluster workload data traces. Extensive evaluations show that DWAP significantly outperforms a representative of state-of-the-art prediction schemes by up to 74% and 59% in terms of the Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE), respectively. In addition, DWAP yields 97% and 48% reduction in task drop rate compared to prominent availability-unaware and availability-based resource allocation schemes, respectively.

Harvesting the profuse yet underutilized computational resources of IoT devices, also referred to as Extreme Edge Devices (EEDs), can significantly curtail the delay in parallelized Mobile Edge Learning (MEL). However, EEDs are user-owned devices, which causes them to experience a highly dynamic user access behavior. Such dynamicity can lead to uncertainty in the available computation and communication capabilities of learners. In this paper, we propose the Minimum Expected Delay (MED) scheme. MED is the first data allocation scheme in MEL that accounts for uncertainty in learners' capabilities and enables multi task allocation. Given the state probabilities of learners, MED strives to minimize the sum of the maximum expected delay of all tasks, while abiding by certain training time and budget constraints. Towards that end, MED formulates the data allocation problem as an Integer Linear Program (ILP) and makes uncertainty-aware decisions. We conduct rigorous experiments on a real testbed of Jetson Nano devices. Extensive performance evaluations show that MED outperforms a representative of state-of-the-art uncertainty-naive schemes by up to 11 %, 11 %, 42 %, and 5 % in terms of training time, satisfaction ratio, data drop rate, and occupancy time, respectively. In addition, MED approaches a baseline scheme that assumes a perfect knowledge of the learners' states, yielding a gap of up to 10%, 5%, and 14% in terms of satisfaction ratio, data drop rate, and occupancy time, respectively.

Autonomous vehicles heavily rely on sensor data to make pivotal driving and traffic management decisions. However, the reliability of such data can be profoundly impacted by many impairments, such as the adverse environmental and weather conditions, the presence of obstacles, and the vehicle's limited view of road and traffic conditions of larger areas. Collaboration between vehicles can help improve the perception of vehicles beyond their line-of-sight, and increase accurate detection of objects. Vehicular Edge Computing (VEC) has emerged as a propitious computing paradigm that can foster the realization of autonomous vehicles. However, maximizing the cooperative perception of vehicles has been mostly overlooked. In this paper, we propose the Cooperative Perception-based Task Offloading (CPTO) scheme. CPTO enables task offloading in VEC with the goal of maximizing the cooperative perception of vehicles and minimizing the latency of perception aggregation, while abiding by a certain deadline. Towards that end, we formulate the task offloading problem as a multi-objective 0–1 integer linear program (0–1 ILP). We also propose a greedy heuristic, called the CPTO-Heuristic (CPTO-H) scheme, to solve the optimization problem. Extensive simulations show that CPTO significantly outperforms the baseline task offloading scheme in terms of perception intensity, service capacity, and satisfaction ratio. Furthermore, CPTO-H closely approaches the optimal solution, with a small gap of up to 3.7% and 2.4% in terms of perception intensity and satisfaction ratio, respectively.

Leveraging task offloading in Vehicular Edge Computing (VEC) via V2X can present unique and robust solutions to the challenges associated with cooperative perception in Autonomous Vehicles (AVs). However, making task offloading decisions that account for the risk of communication failure due to path loss, while adhering to the stringent QoS requirements of cooperative perception has been mostly overlooked. In this paper, we propose PLTO, a Path Loss-Aware Task Offloading scheme that accounts for path loss for Line-of-Sight (LOS), Obstructed LoS (OLoS), and Non-LoS (NLoS) propagation in vehicular communications. We formulate the task offloading problem as a 0–1 Integer Linear Program (0–1 ILP) that aims to minimize the path loss and response delay, while sustaining a certain satisfactory level of improved perception and situational awareness demanded by users. We also propose PLTO-Heuristic (PLTO-H), a scheme to solve the task offloading problem using the MTHG heuristic. Extensive simulations show that PLTO yields significant improvements of up to 17%, 10%, and 23% in terms of packet delivery ratio, Received Signal Strength Indicator (RSSI), and average response delay, respectively, compared to a baseline task offloading scheme that does not consider communication efficiency. In addition, PLTO-H achieves a near optimal solution, with a small gap of up to 6%, 5% and 1.2% in terms of packet delivery ratio, RSSI, and satisfaction ratio, respectively.

Multi-access Edge Computing (MEC) is a revolutionary computing paradigm that facilitates delay-sensitive and/or data-intensive applications associated with the Internet of Things (IoT). Harvesting copious yet underutilized computational resources of the Extreme Edge Devices (EEDs) is foreseen as a promising endeavor. Such EEDs offer a unique opportunity to bring the computing service closer to IoT devices to curtail delay. However, the efficacy of extreme-edge parallel computing paradigm is profoundly impacted by i) wireless device-to-device communication performance, that is required for task offloading; and ii) computing capabilities of the EEDs, that governs the execution time of each task. In this context, we propose a novel spatiotemporal framework that employs stochastic geometry and continuous time Markov chains to jointly analyze the interwoven communication and computation performance of extreme edge computing systems. Based on the incorporated framework, we study the influence of various system parameters on the task response delay. Our findings reveal the existence of an optimal number of EEDs that need to be recruited in order to minimize the task response delay. Moreover, we show that in some cases, our model can outperform the normal MEC offloading systems.

Edge Computing (EC) has been evolving towards harvesting latent yet underutilized computational resources of the Extreme Edge Devices (EEDs), such as autonomous vehicles, smartphones, and tablets. However, EEDs tend to be user-owned devices. This triggers a high level of uncertainty, the impact of which is mostly overlooked. Such uncertainty can stem from the potential loss of network connectivity, battery depletion, as well as the dynamic user access behavior that can affect the computational capability of EEDs and compromise the convenience of users. This uncertainty can profoundly impact the devices' reliability of executing the offloaded tasks. In this context, we propose the Replica Maximization at the Extreme Edge (RMEE) scheme. RMEE employs task replication to achieve maximum reliability and improve successful task execution while abiding by certain QoS requirements. Towards that end, RMEE aims to maximize the number of offloaded replicas for each task, while ensuring that the task execution delay is kept within a certain threshold. We formulate the task replication optimization problem as a Mixed-Integer Linear Program (MILP) and devise an analytical solution using the Karush-Kuhn-Tucker (KKT) conditions and Lagrangian analysis. Extensive simulations have shown that RMEE outperforms other baseline schemes that involve single and fixed number of replicas, in terms of drop rate, satisfaction ratio, and the number of replicas by up to 100%, 100% and 60%, and 95.1 % and 85.4%, respectively.

Fostering Edge Computing (EC) by recycling prolific yet underutilized computational resources of the Internet of Things (IoT) devices, also referred to as Extreme Edge Devices (EEDs), has gained significant momentum lately. Fair resource allocation is a primary concern in such computing paradigms. However, fairness is typically considered from the requester's perspective, whereas fairness for workers (i.e., EEDs) is mostly overlooked. In this context, we propose the Multitiered Worker-Oriented Resource Allocation (MWORA) scheme. In MWORA, the resource allocation problem is formulated as an Integer Linear Program (ILP). MWORA aims to maximize service capacity and minimize the task response delay while enabling fair resource allocation that maintains a specific satisfactory profit for workers. Such a satisfactory level is maintained to prevent the workers from leaving the system and ensure their recurrent subscription to the service. This is done while abiding by the deadline demanded by each requester and without exceeding a certain budget. MWORA also accounts for the fact that EEDs are user-owned devices and are thus subject to a dynamic user access behavior, which can affect the level of computational resources that workers are willing to offer. In particular, MWORA enables multitiered computational resources to be granted by each worker depending on the price of the allocated task. Extensive simulations have shown that MWORA outperforms other baseline resource allocation schemes regarding average response delay, service capacity, worker satisfaction ratio, and fairness.

The surging demand for Edge Computing (EC) to cope with the proliferation of latency-critical and data-intensive applications has inspired the notion of recycling ample yet underutilized computational resources of end devices, also referred to as Extreme Edge Devices (EEDs). Maintaining data privacy and cost efficiency remain core challenges for the viability of EED-enabled computing paradigms. In this context, we propose the Community-Oriented Resource Allocation (CORA) scheme. CORA exploits business, institutional, and social relationships to build clusters and communities of requesters and EEDs that can eliminate recruitment costs and preserve privacy. However, community-imposed constraints on resource allocation can lead to unbalanced work distribution. To address this issue, CORA considers community restrictions, minimizes flowtime and makespan for the allocated services, and retains a reasonable scheduler runtime for real-time resource allocation. Towards that end, CORA formulates the resource allocation problem as a Bipartite Graph Matching problem. Furthermore, CORA exposes tuneable parameters that allow prioritizing flowtime or makespan, making it suitable for different scenarios. Extensive simulations show that CORA outperforms six prominent heuristic-based resource allocation schemes by up to 24% in terms of average makespan while sustaining the same level of flowtime and runtime.

Edge Computing (EC) has paved the way toward the realization of the Internet of Things (IoT). This can be attributed to the ability of EC to bring the computational resources within close proximity to end-users, which significantly improves the response time. However, performance gain in EC can be compromised by service interruptions triggered by various dynamic changes. Consequently, reliable service migration is crucial in EC. However, most service migration schemes either fail to consider the profound impact of the dynamic computation load on service continuity or provide impractical and time-inefficient solutions based on optimization techniques. This paper proposes the Heuristic-based Load-induced Proactive Migration (HLPM) scheme. HLPM incorporates a Finite State Machine (FSM) to model the dynamic computation load. It then makes proactive migration decisions based on the underlying transition probabilities. The proactive migration problem is solved using the MTHG heuristic algorithm. Performance evaluation shows that HLPM produces a significant decrease of up to 97% in migration decision latency compared to conventional optimization techniques. Furthermore, the performance gap of HLPM with respect to the optimal migration solution is just 1.44% latency and 3.89% number of migrations.

Edge Computing (EC) has emerged as a key enabling paradigm for latency-critical and/or data-intensive applications. Recently, recycling abundant yet underutilized computational resources of the Extreme Edge Devices (EEDs), such as smartphones, laptops, connected vehicles, etc, has been explored. This is since EEDs can bring the computation service much closer to the edge, which can drastically reduce the delay. However, resource allocation in such environments typically follows a reactive approach, which can lead to increased delay and wasted resources. In this paper, we introduce the Optimal Proactive Resource Allocation (OPRA) benchmark to quantify the potential gains of proactive resource allocation in EC environments. OPRA exploits the predictability of request patterns to proactively perform resource allocation and create compute clusters that take future task and resource dynamics into consideration. Specifically, OPRA formulates the resource allocation problem as a Binary Integer Linear Program (BILP) problem, where it aims to minimize the total delay under full task assignment and computation capacity constraints. The optimal solution acquired under perfect knowledge acts as the upper bound on the achievable potential of predictive proactive resource allocation schemes. The effect of erroneous predictions on the performance of OPRA is also investigated. Extensive simulation results show that OPRA outperforms a reactive baseline by yielding a 50% decrease in the subtask dropping rate and 97% decrease in the service capacity.

Multi-Access Edge Computing (MEC) has emerged as a computing paradigm that can facilitate the use of Mobile Edge Learning (MEL), where Machine Learning (ML) models are processed at the edge. In MEL, it is important to address system heterogeneity in a way that minimizes staleness to improve learning accuracy. To do so, a centralized data allocation approach is typically used. However, this approach tends to overlook the privacy of learners, since learners' capabilities are assumed to be known beforehand by the orchestrator. In this context, we propose the Data Allocation via Benchmarking (DAB) scheme. DAB is a decentralized data allocation scheme that eliminates staleness and achieves a certain QoS while preserving the privacy of learners. DAB does not allow any information about the learners to be known to the orchestrator. Instead, each learner estimates the upper bound on the amount of data that it can train such that a certain training deadline is not exceeded. In addition, DAB proposes a novel method to enable each learner to accurately estimate its own hardware characteristics via benchmarking. Extensive performance evaluations on a real testing environment have shown that DAB can outperform the centralized data allocation scheme by up to 12% and 26% in terms of loss and prediction accuracy, respectively. Performance evaluations also show that the proposed benchmarking scheme yields an 83% reduction in benchmarking error compared to a prominent baseline scheme.

Democratizing the edge by leveraging the prolific yet underutilized computational resources of end devices, referred to as Extreme Edge Devices (EEDs), can open a new edge computing tech market that is people-owned, democratically managed, and accessible/lucrative to all. Parallel computing at EEDs can also move the computing service much closer to end-users, which can help satisfy the stringent Quality-of-Service (QoS) requirements of delay-critical and/or data-intensive IoT applications. However, EEDs are heterogeneous user-owned devices, and are thus subject to a highly dynamic user access behavior (i.e., dynamic resource usage). This makes the process of determining the computational capability of EEDs increasingly challenging. Estimating the dynamic resource usage of EEDs (i.e., workers) has been mostly overlooked. The complexity of Machine Learning (ML)-based models renders them impractical for deployment at the edge for the purpose of such estimations. In this paper, we propose the Resource Usage Multi-step Prediction (RUMP) scheme to estimate the dynamic resource usage of workers over multiple steps ahead in a computationally efficient way while providing a relatively high prediction accuracy. Towards that end, RUMP exploits the use of the Hierarchical Dirichlet Process-Hidden Semi-Markov Model (HDP-HSMM) to estimate the dynamic resource usage of workers in EED-based computing paradigms. Extensive evaluations on a real testbed of heterogeneous workers for multi-step sizes show an 87.5% prediction accuracy for the starting point of 2-steps and coming to as little as a 16% average difference in prediction error compared to a representative of state-of-the-art ML-based schemes.

Social media traffic is considered the primary source of Internet traffic. Such traffic is largely facilitated by mobile devices. Consequently, cellular networks tend to experience significantly high traffic load, and mobile users tend to incur high cellular costs. In order to alleviate such effects, we strive to allow social media users to have more reliance on vehicular rather than cellular networks for data access. However, this can be impeded by the high delay and low packet delivery ratio often associated with content access from remote data providers in vehicular networks. Thus, we introduce the Vehicular Optimal Proactive Caching (VOPC) benchmark to quantify the potential gains of predictive proactive caching in improving the quality of Internet services in vehicular networks. In VOPC, we exploit the fact that some users tend to exhibit a somewhat predictable behavior in terms of the type and time of social media access during the daily route they follow, as well as the period of encounter with road segments along that route. Such a predictable behavior is utilized to pre-cache the data at parked vehicles to be proactively procured by requesters as they pass by. The objective is to maximize cache hits by assigning replicas to caching spots that yield maximum certainty in their spatiotemporal availability for requesters. This is while sustaining a cache capacity limit. VOPC formulates the caching problem as an integer linear programming optimization problem and can thus act as an upper bound on reachable potential. Performance evaluation substantiates the ability of VOPC to act as a benchmark that can quantify the potential gains of the heuristic-based predictive caching scheme in terms of delay, packet delivery ratio, and cache hit ratio.

Social media traffic constitutes the highest percentage of Internet traffic, which is mostly facilitated by mobile devices. This leads to high cellular costs incurred by mobile users. To reduce these costs, we strive to enable social media users to rely more on vehicular rather than cellular networks for content access. However, this can be hindered by the high delay and low packet delivery ratio often associated with accessing data from distant content providers in vehicular networks. Thus, to bring the data closer to the requester, we propose the Probabilistic Cooperative Caching at Moving and Parked Vehicles (PCCMPV) scheme. In PCCMPV, we exploit the static and mobile nature of parked and moving vehicles, respectively, to dynamically populate valuable road segments with diverse cached data. To do so, we dynamically assign a probability of caching to nodes along the data delivery path to assess their importance as caching nodes. For parked vehicles, such a probability relies primarily on the traffic density of the corresponding road segment, as well as its closeness centrality, and remoteness from the nearest data holder. PCCMPV provides an implicit form of off-path caching by assessing the trajectory of moving vehicles encountered along the data delivery path to calculate their probability of caching. Performance evaluation of PCCMPV demonstrates significant improvements in terms of delay, packet delivery ratio, and cache hit ratio compared to other caching schemes in vehicular networks.

The majority of Internet users are active users of social networks and thus social media traffic represents the highest percentage of Internet traffic. The average daily usage of social media has been rigorously growing. Mobile devices are considered to be the most common facilitators of such usage. The aforementioned facts contribute to the excessive traffic load on the Internet, poor users experience in terms of quality of service, and high cellular costs spent by mobile users. In this paper, we strive to reduce these effects by proposing a scheme called Proactive Caching at Parked Vehicles (PCPV). PCPV aims to provide a better quality of service to vehicular users who tend to have a somewhat consistent social networking behavior. In particular, users who have a predictive behavior in terms of the type and time of access of social media platforms as a part of their daily routine during transit from one place to another. This is done by having the required data pre-cached and ready for users to proactively acquire at roadside parked vehicles as they pass by, rather than sending a request to the far-away data center and waiting for the reply. Performance evaluation of PCPV shows significant improvements in terms of delay, packet delivery ratio, and cache hit ratio compared to the reactive approach typically used in infotainment applications.

Route guidance and navigation services have been widely attracting researchers and application developers due to the serious problems of traffic congestion and the ceaseless need to improve the driving experience. Motivated by such driving concerns, this paper proposes a real-time, dynamic route guidance system with the main focus on the driver safety and satisfaction. As a unique feature compared to other existing systems, the proposed driver-centric route guidance (DCRG) system considers the driver behavior in the route guidance process for the sake of boosting the safety levels on roads. The system also considers the driver preferences targeting a personalized satisfying driving experience. As most drivers prefer traversing the fastest and healthiest route to their destination, the DCRG system takes into account as well the real-time traffic and road conditions while guiding drivers towards their targeted destinations. Performance evaluation of DCRG shows significant improvements in the travel time, on-road safety, and preference satisfaction levels compared to the shortest and fastest route guidance schemes.