Introduction
Preliminary background
Vehicular edge computing architecture
Computation offloading
Notation | Description |
---|---|
VEC | Vehicular edge computing |
MEC | Mobile edge computing |
IoV | Internet of vehicles |
IoT | Internet of things |
V2X | Vehicle to everything |
UE | User equipment |
RA | Resource allocation |
RSU | Road side unit |
VECC | Vehicular edge cloud computing |
CAVs | Connected autonomous vehicles |
SDN | Software defined network |
DRL | Deep reinforcement learning |
AI | Artificial intelligence |
VANET | Vehicular adhoc network |
TaV | Task vehicle |
SeV | Service vehicle |
PTE | Primary task entity |
CO | Primary computation offloading |
QoS | Quality of service |
NFV | Network function virtualization |
Related work
Computational offloading in IoT-enabled VEC-survey overview
Classification of RA frameworks in VEC based on optimization goals for IoT applications
Power/energy consumption optimization
QoS optimization
Delay optimization
QoE/user satisfaction optimization
Utility optimization
Maximizing reliability
Optimizing memory allocation
Summary of this section
Mathematical/computational models and algorithms | References | Advantages | Disadvantages |
---|---|---|---|
Gauss Seidel method | [29] | Used for finding frequency for calculating the power | Makes the programming of the network difficult as each iteration requires more computation time per iteration |
Online Lyapunov optimization | Enhances the user experience by providing the trade off between energy consumption and delay | There is no general method available to construct lyapunov functions | |
Game theory | Reduces the inter-cell interference, increases the throughput and guarantees the reliability of URLLC V2X communication by choosing the best offloading decisions | Sometimes the assumptions can be unrealistic while making certain offloading decisions which may not produce the best outcome | |
T-slot look ahead algorithm | [36] | Reduces the computing quality loss and provides a computationally efficient solution | Energy consuming |
Bipartite graph | Solves the NP-hard energy optimization problem when all the vehicles offload their tasks to the same VEC server | Solve the energy optimization problem but does not solve the latency issues | |
Heuristic algorithm | Used for reliability maximization and solves the problem of allocating resources to fog computing in vehicular applications. | Often used only when approximate solutions are sufficient and exact solutions are not necessary | |
Markov model | Reduces the computational limitations and delay of the VEC | Markov Models are generally inappropriate over sufficiently short intervals | |
Fault-tolerant particle swarm optimization (FPSO-MR) | [20] | Converts non-convex functions into convex functions for to solve the formulated problem with less convexity | It has a very low convergence rate in the iterative process. |
Convex optimization | [36] | Replaces the non-convex functions and restrictions by acceptable convex functions which helps in reducing the energy consumption | Difficult to design a suitable model and develop efficient, fast, scalable and distributed algorithm to solve large practical issues |
Reinforcement learning | [28] | Optimizes the computation offloading scheduling problem by which energy and delay are minimized in long term | Needs a lot of data and requires extensive computation which may sometimes diminishes the result |
Newton–Raphson method | [25] | Reduces the complexity in case of huge overheads using convex optimization algorithm | Sometimes the convergence of Newton–Raphson method is not guaranteed especially in case of multiple roots |
Distributed deep-Q learning | [28] | Minimizes the overall computational and energy consumption | Requires a large amount of data for better performance and also it is overly sensitive to the changes in the data set |
Non-dominated sorting genetic algorithm | [39] | Improves the user experience and reduces the latency | Genetic algorithms are non-deterministic method. Thus, it may give different solution each time the algorithm is run for same instance which makes it non-optimal in some cases |
Multiple criteria decision making | [39] | Provides the most optimal solution for the computational offloading strategies | Time consuming |
Select sort algorithm | [34] | Allocates the minimum required resources which helps in reducing the energy consumption | This algorithm is beneficial when used for small data but its efficiency decreases when dealing with large amount of data |
Randomly select algorithm | [34] | Uses random resources for allocation based on game theory which helps in reducing the latency | There is very little control on what the model does and sometimes its efficiency is not optimal as it does random selection |
Classification of RA frameworks in VEC based on mathematical and computational models/algorithms/techniques used
Gauss Seidel method
Online Lyapunov optimization
Game theory
Cooperative and non-cooperative games
Static and dynamic games
Bipartite graph
Markov model
Heuristic algorithm
Fault-tolerant particle swarm optimization (FPSO-MR)
T-slot look ahead algorithm
DC (difference of convex) programming technique
Reinforcement learning algorithm
Newton–Raphson method
Distributed deep-Q learning
Non-dominated sorting genetic algorithm-III (NSG-III)
Multiple criteria decision making (MCDM)
Simple additive weighing (SAW)
Bracketting
Binary searching
Distance based heuristic algorithm (Dbha)/Euclidean distance-based algorithm
Maximum value density based heuristic allocation (MVDHA)
Randomly select algorithm (RSA)
Select sort algorithm (SSA)
Summary of this section
Classification of RA frameworks in VEC based on underlying technologies
SDN
Blockchain
Artificial intelligence
[References] | Problem addressed | Formulation/modelling | Solution | Parameter/s optimized | Underlying technology |
---|---|---|---|---|---|
[29] | Joint optimization of power consumption and network stability in MEC assisted Cellular V2X | Non-deterministic polynomial-time hard problem, two subproblems: 1. URLLC RA and 2. Computation resource decision | Online Lyapunov optimization method to solve CO. Gauss–Siedel method to compute CPU frequency | Delay optimization and network/system utility | – |
[26] | Service hole problem in dynamic networks with static edge servers in MEC scenario | Semi-Markov process with Stochastic traffic, time-varying comm. and CO requests | A novel VEC network architecture: vehicles function as mobile edge servers. Q-learning & DRL to obtain optimal CO and RA | Power/energy optimization | AI |
[35] | Computationally intensive IoVs require minimal delay in MEC. Sudden increase in computations difficult to handle due to limited resources | RA modelled as Markov decision process | A regional intelligent vehicular system with dual MEC planes. MEC servers lying in same region co-operate to achieve resource sharing | Delay optimization and QoS | AI |
[36] | Need for low latency and high reliability among capacity-constrained and distributed MEC nodes. Long-term satisfactory UE for dynamica nd uncertain vehicular environments | A mix-integer non-linear stochastic optimization problem for computing quality loss minimization considering queuing, dynamics, throughput constraints, and worst-case delay bound | Adaptive RA for enhancing UE in VEC networks | Delay optimization | SDN |
[37] | Need for a collaborative MEC-cloud approach for offloading services to vehicles | Jointly optimizing CO decisions and RA—A non-convex NP hard problem | Collaborative CO & RA (CCORAO) scheme and distributed CO and RA algorithm (DCORA) | Delay optimization, QoE, user satisfaction, network/system utility | Blockchain and cloud computing |
[38] | Present resource-constrained vehicles do not meet low/ultra-low latency demands in delay-sensitive applications such as autonomous driving | Offloading modelled as a constrained delay optimization problem | Fiber-wireless (FiWi) technology that provides centralized network management and supports multiple comm. techniques | Delay optimization | SDN, cloud computing |
[25] | Need for ubiquitous connections and high QoS for vehicles in a vehicular network | Vehicles divided into groups and RA formulated as convex optimization problem | A mobility-aware task offloading along with a location-based offloading scheme | Delay optimization | Cloud computing |
[28] | Limited computational and battery capacity makes long term network participation difficult for vehicles in vehicular edge-cloud computing networks | An integration model of CO and RA optimizing weighted sum of energy consumption and latency as a binary optimization problem | A distributed DL approach to find near-optimal CO decisions using a set of DNNs parallelly | Power/energy optimization | AI and cloud computing |
[39] | Vehicles have computationally intensive tasks but limited storage and computation resources. | Resource utilization and Latency reduction in IoV modelled as a multi-objective optimization goal | A CO method that employs V2X with edge computing. A non-dominated sorting algorithm-III (NSG-III) for balanced offloading using additive weighting and multiple criteria decision making | Delay optimization and network/system utility | Cloud computing |
[33] | The difficulty to enhance precision in resource provisioning among distributed edge clouds in high mobility scenarios to optimize QoS and provisioning costs | An optimization problem to minimize cost of provisioning resources at edge cloud with a specified service blocking probability threshold | A stochastic traffic analysis based framework to optimize resource provisioning cost and keep blocking probability within a preset limit | QoS and network/system utility | Cloud computing |
[34] | Intelligent RA strategies required for intensive and low-latency vehicular applications in dense networks | Modelled as a special knapsack problem—NP-hard category in time taken to obtain solution | An RSU backed network with three layers-Cloud, RSU-Cloulet and vehicular cloud for real-time response to resource request | QoS, QoE, user satisfaction | SDN, cloud computing |
[30] | Excessive computing and energy requirements in heavily time and resource constrained vehicular scenarios | A block level minimum assignable resource for vehicular CO | A three-layer VECC framework to support real-time computation augmentation, energy conservation, and interconnection CAVs at network edges | Power/energy optimization | Cloud computing |
[40] | Difficulty for vehicles to provide required computation and performance levels in computationally intensive and delay sensitive applications | Model the joint load balancing and offloading under the permissible latency constraint to maximize utility of the system | Jointly optimizing VEC server offloading and selection in a distributed way with lower overhead | Network/system utility | – |
[20] | Cloud computing may result in excessive delay for latency-sensitive applications like assisted driving etc | NP-hard, non-convex optimization for maximizing reliability performance of CO jointly considering offloading in parts, and task allocation and reprocessing for maximizing reliability | A low complexity heuristic algorithm to maximize reliability in latency constrained scenarios | Maximizing reliability | SDN |
[37] | Vehicle dynamicity, delay-sensitive data processing and RA at EC server are challenges in intelligent driving | Objective function to maximize user satisfaction and three models to optimize user satisfaction: 1. Markov based state prediction, 2. Resource requirement, 3. Auction | Resource transaction architecture based on dynamic edge resource allocation and blockchain using double auction technique to optimize user and service provider satisfaction | Delay optimization, QoE and user satisfaction and network/system utility | blockchain and cloud computing |
[47] | Limited number of vehicles are able to use fog computing due to its resource restrictions | Allocating limited fog resources to vehicular applications such that the service latency is minimized and user satisfaction is improved, by utilizing parked vehicles | A VFC RA algorithm that considers both short-term and long-term RAs obtained from a heuristic and RL (reinforced learning) algorithms respectively | Qos | AI |