Dynamic task offloading and collaborative task execution using three tier edge cloud computing (T 2 EC 2 ) system for autonomous vehicles

Abstract

Automobiles have undergone a transformation during the past two decades due to the merger of the electronics and automotive industries. The combination of autos and electronic sensors has resulted in a new generation of vehicles known as autonomous vehicles (AVs). These AVs have a few hundred thousand sensors, producing an enormous amount of raw data for computation. Data from the vehicular network can be offloaded to existing telecommunication infrastructure to address the problem of processing resources. In order to address vehicular network requirements, large-capacity servers deployed in major telecommunications networks are first used to offload resource-intensive tasks. Mobile Cloud Computing (MCC) is a critical enabling technology for 5 G networks, which has a key feature of offloading to divide application tasks into local and cloud server execution components. This paper proposes a novel Three TierEdge cloud computing (T² EC²) system which uses an Energy-aware Dynamic Task offloading and collaborative task execution algorithm (EA-DTOCTE) for multilayer vehicular cloud computing networks. The EA-DTOCTE algorithm is included in the decision-making engine in the proposed system, which selects whether to offload the task to the remote environment or implement it locally. EA-DTOCTE focuses on consumption of energy by tasks both locally and remotely since its goal is to efficiently and dynamically split the application into tasks and schedule them on local devices and cloud resources. The proposed T² EC² has been evaluated in terms of parameters such as energy consumption, completion time, and throughput. Experimental results indicate that the proposed T²EC² can save up to 28% of system energy consumption compared with other state-of-art techniques.

Keywords

Autonomous vehicles mobile cloud computing application partitioning offloading scheduling EA-DTOCTE decision making engine collaborative task execution

1 Introduction

The Internet of Things (IoT) [1] and wireless sensors have become increasingly popular in recent years. Furthermore, as 5 G technology evolves, communication capabilities will gradually improve, resulting in the creation of new applications with advanced features like virtual/augmented reality, autonomous vehicles (AV), e-Health and face recognition. As the number of vehicles on the road increases, the Internet of Vehicles (IoV) [2] is becoming increasingly popular in the transportation sector. But, due to the inadequate computing capability and vehicles’ battery power, satisfying these necessities and confirming the appropriate quality of service (QoS)level is a significant problem. Furthermore, automated driving, voice recognition, in-vehicle entertainment and video streaming are projected to be used in automobiles. Due to the large amount of data generated by these workloads, these applications require not only severe time limitations, but also massive processing resources. Due to the fact that present cloud servers are far from the vehicle, offloading [3] these tasks to the cloud server increases latency problems.

Data from vehicular networks could be offloaded through existing telecommunication infrastructure to reduce computing resources requirements. Initially, this idea is realized by offloading resource-intensive tasks to high-capacity servers deployed at core telecom networks. This method of offloading vehicular network data is known as Mobile Cloud Computing [4, 5]. Through efficient offloading, MCC connects Mobile and Cloud Computing. Certainly, communication channel accessibility and availability are critical in enabling this offloading as a viable option for offloading device-based application tasks or parts of application components to the remote cloud services. Certainly, by relieving computing loads from mobile elements, a significant amount of energy can be saved on devices, in addition to offloading. But this system does not consider the practical constraints about the variable moving speed of vehicles. However, energy conservation is only possible if the overhead in management of communication channel is monitored and regulated appropriately.

The offloading process often brings up a number of concerns that must be addressed in order to increase performance or reduce energy consumption. For example, moving application load may result in increased compute and communication overhead, increased consumption of energy of external resources, user and application usability problems, or application and service QoS requirements. All of these factors have an impact on, and even obstruct, increasing the energy efficiency of the task offloading scenario during execution. According to the major aspects of MCC environment, the offloading process is divided into three main parts: partitioning, environment profiling, and decision making and execution.

Offloading is done at the device or application level in Task Partitioning [6], depending on the number of apps and the size of the system. To make a decision about real-time offloading scheduling in the next step, environment profiling is done. The implementation environment information has been updated as references in this step, so as to reduce the likelihood that offloading overhead outweighs its advantages. Depending on the sensors used, LIDARs, cameras, or a combination of these two types of devices can be used to solve the environment perception task. Three parts of the profiled data are included: the present job load, offloading resources, the computing status of mobile devices and the networking between them. The final step in the offloading technique is to make a global judgement about distributed job scheduling. Although each task has previously been partitioned, tagged, and profiled, a global review of the complete task set is required to optimize executions.

Task partitioning techniques [6] has been used in existing systems that migrate the partitioned tasks to the serving Multiaccess Edge Computing Server (MECS) and cloud server to reduce the average total delay of tasks. However, MECSs have a certain amount of processing capacity, and the amount of processing that a vehicle may request from a MECS relies on the load level of the MECS, which causes significant traffic congestion. To overcome the problem in task partitioning a cooperative task scheduling technique has been used, but it did not evaluate many types of services in cloud environment. Then a collaborative route planning is considered for mobile edge servers to move in vehicular edge networks for computational offloading [23]. However, these collaborative route planning possess latency during decision process. Furthermore, the presence of unavoidable features [22] in real-time systems, such as external-internal parametric and non-parametric uncertainty (i.e., time-varying load, network bandwidth, and resource availability speed), leads to poor control performance.

Therefore, a novel Three tier Edge cloud computing (T² EC²) system has been proposed in this paper which overcomes the above-mentioned problems. The main contributions of the proposed system are given as follows:

The proposed method uses a decision-making engine that decides the task offloading based on a novel EA-DTOCTE algorithm.

Based on current conditions of vehicles such as available network bandwidth, time varying task load and cloud resource available speed, the decision-making engine will decide whether to offload a task to a remote location or to execute it locally for reducing the energy consumption.

The proposed T²EC² method will decide to offload the task in a small amount of time which reduces the congestion and also latency for processing the tasks from autonomous vehicles.

The performance of the suggested framework has been evaluated in terms of specific parameters such as, total system performance, time and energy by comparing it with other benchmark solutions.

The remainder of the paper is laid out as follows. Part 2 discusses similar work, while parts 3 and 4 discuss problem definition and formulation. Section 5 defines the proposed algorithm. Parts 6 and 7 define performance evaluation and conclusion, respectively.

2 Related works

Mobile computing systems are complementary to offloading of computation and data, but it has evolved into a powerful tool for enhancing autonomy and enabling sophisticated approaches. However, energy efficiency of computation offloading is a challenge for this system. There have been several works describing offloading solutions for an IoT-edge-cloud integrated computing system.

Sun et al. [7] introduced a cooperative task scheduling technique for compute offloading in Vehicular Cloud (VC) that is phrased as an NP-hard scheduling issue and takes into account the specific characteristics of VC such as instability, heterogeneity, and interdependency of computing jobs. Moreover, a genetic-based heuristic method was created to solve the stated NP-hard scheduling problem, which lowered the problem’s complexity while improving the usage of the vehicle’s onboard resources. It increased CPU resource usage while ensuring low latency and system stability. It did not evaluate many types of services in VC with specific needs.

Vemireddy et al. [8] proposed an energy-efficient vehicle scheduling issue for offloading tasks to mobile fog nodes while satisfying task deadlines and resource constraints. To address the issue of high dimensionality induced by the growing number of vehicles covered by Road side units (RSU), Fuzzy Reinforcement Learning (FRL) was introduced, which combines a fuzzy logic based greedy heuristic with an on-policy reinforcement bias-based scheduling technique. Using either energy-constrained stationary fog nodes (RSU) or mobile fog nodes, this compute architecture allowed operations to be completed near the network’s edge. The reduction of RSU energy usage is regarded as an evaluation metric when activities are completed within a reasonable response time.

Tsheng et al. [9] suggested two job offloading techniques for mobile cloud. In order to execute the algorithms in MCC, the algorithms took into account the shortest possible execution time. The minimum offloading time algorithm (MOTM) chooses offloading links depending on task priority, while the minimum execution time algorithm (METC) identifies the best virtual machines (VM)for the jobs.

The energy-efficient task offloading problem was described by Jang et al. [10] taking into account the vehicle mobility using time-variant channels, the job deadline, and computation and communication energy consumption with multiple access. The offloading process was also examined frame by frame, with the ideal allocation of bits being each frame’s number of uplink/computation/downlink bits. The combined optimization of allocation of bits and ratio of offloading resulted in significant reduction in energy use when compared to the benchmarks, according to numerical figures.

For MEC-enabled distributed AV architecture, Rasheed et al. [11] presented an application-aware hierarchical offloading strategy (HOS). This architecture separated the network into 3 layers based on the requirements of application, resulting in a network that is quick to respond and efficient. As a result, each application was treated at the proper layer to satisfy its latency and computing needs. As can be seen from the data, increasing the number of RSUs per Mobile Edge Computing (MEC)adds delay.

Zhu et al. [12] suggested a multiagent DRL-based computation offloading approach that takes into account the unpredictability of a multivehicle environment so that vehicles can make the best long-term offloading decisions. The actor and critic networks were centralized trained using the multiagent DRL algorithm. In a distributed manner, the cars made real-time task offloading decisions. The goal of this study was to determine the appropriate MEC server offloading decision for each observable system condition in order to reduce total task processing time over time.

Apostolopoulos, P.A. et al. [13] provided a paradigm for making data offloading decisions in Multi-access Edge Computing (MEC) environment, users could partially offload their data to MEC servers mounted on the ground and in UAVs. Detailed numerical data that show the framework’s effectiveness and superiority are presented along with the introduction of a low complexity distributed algorithm that converges to the PNE.

Tutsoy, O. and Barkana, D.E. [22] developed a model-free digital adaptive control technique for underactuated manipulators. Timing delay, actuator saturation, and non-parametric uncertainty are all taken into account. Dynamic tests were conducted using the suggested digital adaptive control.

The above-mentioned approaches have various problems in offloading tasks, such as energy consumption. As vehicle equipment creates massive amounts of data, the vehicular edge computing network cannot support the massive amount of data created. The network has limited capabilities in terms of communication, processing, and storage. To overcome these challenges a novel Three tier Edge cloud computing (T² EC²) system has been proposed and discussed in next section.

3 Proposed Three tier Edge cloud computing (T² EC²) system

Figure 1 depicts our three-tier edge-cloud computing (T² EC²) system concept for vehicular networks: 1) Autonomous Vehicles (AV) (2) edge computing, and 3) cloud computing. m number of vehicles move at a dynamic pace on a bi-directional road in the first layer. In terms of processing and storage capacities, the edge servers are diverse. A vehicle v_m can only communicate with an edge server e_i if it is inside RSU’s communication range. The third tier (cloud computing) is made up of H heterogeneous cloud servers, with a cloud server f_H having greater processing and storage capabilities than an edge server. As illustrated in Fig. 1, Each cell is served by a single RSU, the radius of whose coverage is assumed to be r. Furthermore, we believe that every RSU [13] outperforms the On-Board Unit (OBU) in terms of processing. We estimate RSU has ten times the processing capability of AV for our computing study. Furthermore, let S_m denote the AV m’s speed, and t_m,n denote the AV m’s resident duration inside the scope of RSU n. M also represents the number of AVs that want to assign their responsibilities to the RSU or EC. We further assume that each AV has C jobs to offload, with each task c being inseparable.

Fig. 1

System model.

In our model, each edge server receives a set of requests from communicating vehicles. For each request r_j, the decision-making engine determines whether to implement the request locally on the edge e_i or offload it to a cloud server f_H for execution. A decision is made to offload so that the execution time of all requests is minimized. When a request is offloaded, the edge server sends the request and data to the cloud management server f_H. The request is then scheduled to f_H by the cloud management. The notations employed in this paper are listed in Table 1 along with their definitions.

Table 1

Notations

Notation	Definition
v _m	M number of vehicles
e_i, f_H	edge server, cloud server
S _m	Speed of autonomous vehicle
t _m,n	AV m’s resident duration inside the scope of RSU n
r _j	j^th request
$T_{m, local}^{c}$	execution time for doing all of the computing activities locally
d _m,c	data size (KB) denoting the task c of AV m
cc _m	compute capability (bps) of the OBU installed inside the AV
$T_{m, \max}^{c}$	maximum delay for task c of AV m.
$E_{m}^{c}$	energy used with each CPU cycle
v _m	Coefficient
p _m	Entire number of CPU cycles necessary to finish the computational work
$T_{m}^{f}$	The execution time for remotely executing the computing work of vehicle m at the cloud server
${cc}_{m}^{f}$	Cloud server’s computational capability for vehicle m
c _m	Task running in a local environment
${cc}_{m}^{rsu}$	Computational capability of RSU for vehicle m
T_local (c_m)	Completion time of c_m
δ_m, δ_m+1	Workload of task c and c_m+1
φ	Decision variable
$Y_{local}^{- 1}$	Computation speed of local environment
E_local (c_m), E_cloud (c_m+1)	Function for consumption of energy
P _local	Power consumption of local environment
T _b	Network’s transmission bandwidth
Y _cloud	Computation speed of cloud
P _idle	The power consumed by the local environment when it is not in use
C₁, C₂	Constraints
Z _i	Overall consumption for processing the computational task of vehicle i at server j

3.1 Local execution

We assume that various vehicles are likely to have different processing capacities when performing local execution. Furthermore, the computing activity is performed locally by all vehicles. Thus, for each vehicle m, the execution time and energy usage for doing all of the computing activities locally can be calculated as follows: $T_{m, local}^{c} = \frac{d_{m, c}}{c c_{m}}$ (1)

The Equation 1 represents the execution time for doing all of the computing activities locally. where d_m,c denotes the data size (KB) denoting the task c of AV m and cc_m is the compute capability (bps) of the OBU installed inside the AV. Let us represent $T_{m, \max}^{c}$ as the maximum delay for task c of AV m. $E_{m}^{c} = v_{m} p_{m}$ (2)

Equation 2 represents the energy used with each CPU cycle for task offloading. where p_m signifies the entire number of CPU cycles necessary to finish the computational work and v_m is a coefficient which denotes the energy consumed per CPU cycle.

3.2 Remote execution

The EA-DTOCTE algorithm is included in the decision-making engine which determines whether or not to offload the task in the proposed architecture. The decision-making engine must compute the local environment’s energy usage before executing the task. As a result, the execution time for remotely executing the computing work of vehicle m at the cloud server can be represented as follows: $T_{m}^{f} = \frac{p_{m}}{{cc}_{m}^{f}}$ (3) $T_{m}^{p} = \frac{p_{m}}{{cc}_{m}^{p}}$ (4)

Equation 3 represents the execution time for remotely executing the computing work of vehicle m at the cloud server. Equation 4 represents the computational time for cloud server. Where ${cc}_{m}^{f}$ , ${cc}_{m}^{p}$ denote the cloud server’s computational capability for vehicle m.

3.3 Local environment energy consumption model

The task c_m is running in a local environment; the task c_m′s completion time [21] is shown as

$T_{local} (c_{m}) = δ_{m} (c_{m}) \times φ (c_{m}) \times Y_{local}^{- 1}$ (5)

Equation 5 represents the completion time of task running in a local environment. Where δ_m is the workload of task c, φ is the decision variable and Y_local is the computation speed of local environment.

The function for consumption of energy is stated as follows based on Equation (5). $E_{local} (c_{m}) = P_{local} \times T_{local} (c_{m})$ (6)

Equation 6 represents the function for consumption of energy. Where P_local is the power consumption of local environment [14].

If the task c_m+1 is offloaded to a remote environment [15, 16], the task’s time to complete and usage of energy in the local environment are both zero.

3.4 Cloud environment energy consumption model

The task c_m+1 is offloaded to a remote cloud environment; the task c_m+1 completion time is indicated as

$\begin{matrix} T_{cloud} (c_{m + 1}) = (T_{m, local}^{c} \times T_{b}^{- 1}) \\ + δ_{m + 1} (c_{m + 1}) \times (1 - φ_{m + 1} (c_{m + 1})) \times Y_{cloud}^{- 1} \end{matrix}$ (7)

Equation 7 represents the task completion time in the cloud server environment. Where T_b is the network’s transmission bandwidth, δ_m+1 is the workload of task c_m+1 and Y_cloud is the computation speed of cloud.

Equation (8) is used to define the following energy consumption function for cloud servers: $E_{cloud} (c_{m + 1}) = P_{idle} \times T_{cloud}$ (8) where P_idle is the power consumption at the idle state of the local environment

3.5 Problem formulation

The problem of establishing an energy efficient offloading of computations for vehicular MCC networks [17] is discussed in this section. The computational offloading problem is described as the following constrained optimization formulation problem using the preceding energy consumption and computational models: $min_{φ} \sum_{i = 1}^{m} Z_{i}$ (9) $C_{1} = \sum_{j = 0}^{M + 1} φ_{i, j} = 1$ (10) $C_{2} = φ_{i, j} ɛ {0, 1}$ (11)

The optimization problem’s objective function is to use task offloading to lower the weighted sum of system consumption relating to time and energy which is represented in Equation 9. Equations 10 and 11 represents the constraints for the proposed models. Constraint C1 ensures that each computing work performed by vehicle m is only performed once. The binary nature of the computational offloading choice variable is guaranteed by constraint C2. The optimal values of the task offloading decision φ can be found to identify the solution to this problem. The problem’s feasible set and objective function are not convex since φ is a binary variable, making it hard to resolve, particularly for a high number of vehicles, due to the well-known curse-of-dimensionality problem., in which the problem’s size grows exponentially with the number of vehicles [18, 19]. To efficiently identify the optimal settings, the Dynamic Task Offloading and collaborative task execution algorithm (EA-DTOCTE) is employed.

We consider the constraints c1 and c2 minimum data rate of both offloading Vehicle and offloaded RSU for smooth data transfer between vehicle and RSU. The proposed T²EC² takes into consideration an offloading vehicle’s connectivity time with the RSU, with other vehicles travelling in the same or the opposite direction, the offloading capacity, and the contact duration with nearby vehicles. Offloading capacity is the most amount of data that a vehicle can send to an RSU during a connectivity period. The connectivity time has a direct correlation with the unloading capacity. Offloading capacity will increase with increased connected time, while offloading capacity will decrease with decreased connectivity time. Vehicle I might be able to handle a higher data rate, but RSU or another vehicle to whom it is offloading its data might only be able to support a lower data rate, which could result in data loss. Therefore, RSU performance can satisfy the offloading tasks of multi vehicles based on connectivity time.

3.6 Proposed EA-DTOCTE algorithm

The proposed method as illustrated in Algorithm 1 EA-DTOCTE has two sub problems: dynamic task offloading and collaborative task execution. Task size, device, network bandwidth, and cloud resource considerations can all be used to create Dynamic task offloading. A decision-making algorithm is used to formulate the dynamic application problem. Algorithm 2 is used to divide applications into local and remote execution. Algorithm 2 always returns the best application partitioning for reducing energy consumption. The proposed Flow chart of SSC method is depicted in Fig. 2.

Fig. 2

Flow chart of SSC method.

Algorithm 1: EA-DTOCTE Algorithm

Input: v_m ∈ V;

Output:

min_{φ} Z

;

begin

Z←0;

Task Offloading Based On algorithm 2;

Workflow Task Sequence;

for each v_m ∈ V do

Collaborative Task Execution based on algorithm 3;

Task Sequence Adjustment;

returnZ;

3.7 Task offloading process

One of the essential modules for offloading the AV’s data for computation in the proposed method is the decision-making engine as shown in Fig. 3. The decision-making engine’s function is to predict whether or not to offload a task to a remote location or to execute it locally. Algorithm 2 describes task offloading, which prioritizes completion time and consumption of energy as main factors for offloading. The user criteria must be met during job execution in the local environment. As a result, it must meet the user’s deadline.

Fig. 3

Decision-making engine workflow.

Algorithm 2: Decision making engine process

Input:C=

\sum_{i = 1}^{m} c_{m}

Output: Execution of task c in local/cloud environment

begin

Step 1: if δ (c) ⩽ 0 and φ (c) ⩽ 0 and φ (c) ⩾ 1 then

go to step 4;

Step 2: for every δ (c),if (φ (c) = =1)

Compute T_rsu & T_m,n;

T_{rsu} < t_{m, \max}^{c} & T_{rsu} < T_{m, n}

T_{b} < T_{rsu} & d_{m, c} < {cc}_{m}^{rsu}

Execute c_m locally;

else

go to step 3;

End if

End for

Step 3: if ((φ (c) = =0)

Compute T_cloud& E_cloud ;

Offload c_m to remote server;

Break;

Step 4: return “no offload”;

Break;

End

The workflow of the algorithm as shown in Fig. 4. calculates the task’s completion time and compares it to the user-specified time T. It computes the resident time and total time of completion of the RSU and then checks the specified condition. Task C is computed in the immediate environment if the condition is satisfied. A neighbouring cloud server is assigned the work in any other case.

Fig. 4

Workflow of Task offloading process.

3.8 Collaborative task execution

It focuses on the execution of group tasks in both the local and remote environments. Depending on completion time and energy consumption, tasks will operate either in the nearby or distant cloud environment. The overall consumption of energy of the offloaded tasks is calculated as $\sum_{i = 2}^{m - 1} E (c_{i}) = E_{send} (c_{i}) + E_{cloud} (c_{i}) + E_{receive} (c_{i})$ (12) $\sum_{i = 2}^{m - 1} T_{b} (c_{i}) = {d_{m, c}}_{send} (c_{i}) + {d_{m, c}}_{receive} (c_{i})$ (13)

Equation 13 represents the transmission bandwidth of the offloaded tasks. The bandwidth can be calculated as the data that can be sent and received. The option to offload the tasks c₁ and c_n is different from the rest of the tasks, in that c₁ and c_n run in a local environment, and the resultant consumption of energy is given in equation 14. $E (C) = E_{local} (c_{i}) + \sum_{i = 2}^{m - 1} E (c_{i}) + E_{local} (c_{n})$ (14)

This method determines C’s total energy consumption. The consumption of energy of tasks that are offloaded to the cloud environment is determined by the cloud’s energy consumption, necessary input data, and transmitted output data.

Algorithm 3: Collaborative Task Execution

Input: from algorithm 2’s decision, dec

Output: overall energy usage of the task

begin

step 1: if (dec==offload)

Compute consumption of energy of c in cloud;

implement c;

step 2: if (dec !=offload)

Compute consumption of energy of c in local environment;

implement c;

step 3: Merge consumption of energy of environments together;

Compute E(C);

End

This algorithm computes the energy consumption in the local environment if the subtask runs in a local environment. Finally, by merging both environments, the resultant energy consumption is calculated

3.9 Numerical Analysis for computational complexity

The computational complexity of the proposed EA-DTOCTE algorithm is O(V), where V=|v_m| is the number of vehicles v at server j, and the computational work in collaborative task execution algorithm is to calculate energy consumption. Algorithm 3 selects tasks offloading with a minimum search problem having an O(V) complexity. Updating the offloading delay $T_{m, \max}^{c}$ , and the offloaded time $T_{m, local}^{c}$ is a complex operation. Due to this, running EA-DTOCTE to offload one task has an overall computational complexity of O(V). When offloading tasks separately, there is a total amount of computation of O (CV). In this case, O(KC² V) is the complexity, where K is the number of iterations needed to complete the suggested process in the decision-making and collaborative algorithm. Therefore, EA-DTOCTE is of lower complexity than the existing algorithms.

4 Results and discussion

We present simulation results in this section to assess the efficiency of our suggested technique. We have implemented the proposed system using NS2 simulator. We consider that all AV’s have the same deadline T and that the vehicles arrive at the coverage edge of the first RSU at random. Our studies suggest that the proposed architecture reduces latency by distributing work in a hierarchical manner. In this system, each user makes use of the common shared channel while consuming the least amount of energy possible. Energy consumption, completion time, and throughput are the metrics considered for user satisfaction. Simulation parameters for the proposed T²EC² technique has been given in Table 2.

Table 2
Simulation parameters

Parameters Value

Coverage area of RSU 200m

Data size 0.5MB

Computation capability of OBU 0.5GHz

Computation capability of RSU 10 GHz

Computation capability of cloud 1x10¹²

Wireless link capacity 1 Gbps

Arrival rate of RSU 50/s

Type of offloading dynamic

Speed of vehicle 15 m/s

Transmission and reception power of vehicles 24dBm

Parameters	Value
Coverage area of RSU	200m
Data size	0.5MB
Computation capability of OBU	0.5GHz
Computation capability of RSU	10 GHz
Computation capability of cloud	1x10¹²
Wireless link capacity	1 Gbps
Arrival rate of RSU	50/s
Type of offloading	dynamic
Speed of vehicle	15 m/s
Transmission and reception power of vehicles	24dBm

This framework was evaluated on a MacBook Pro laptop with an Intel Core i7 processor and 16GB of RAM. A set of S = 10 MEC servers is used to cover 200 users, each of which has a coverage area of 200 m. The users are distributed evenly and randomly within a two-dimensional grid of 1000 m by 1000 m.

When the user deadline is stricter, the energy usage of the suggested offloading strategy becomes more effective compared to FRL, MOTM, and HOS. Because of its dynamic nature, the proposed algorithm uses less energy. The proposed approach is more flexible since it uses a dynamic offloading choice algorithm instead of the whole task implementation in the remote environment to offload the task.

The overall consumption of energy of all vehicles is shown in Fig. 5 and the plotted function for the offloading proportion is shown in Fig. 6. We can see that local execution consumes a lot of energy, and that offloading around 60% of the operation is more efficient than offloading the entire task. Data offloading consumes more energy than local execution when a vehicle is far from the RSU.As a result, in order to reduce energy usage, a portion of the task must be completed locally.

Fig. 5

Energy Consumption of vehicles.

Fig. 6

Energy consumption of vehicles w.r.t offloading proportion.

The variance in latency versus the number of tasks is depicted in Fig. 7. The proposed algorithm EA-DTOCTE outdoes FRL, MOTM and HOS according to the results. In the performance analysis, we assumed a growing number of jobs for each vehicle. Most applications’ latency needs are met by the proposed algorithm. When number of tasks grows, the time required for local computation grows at an exponential pace, making future applications unfeasible. Because a huge number of jobs are offloaded by vehicles in an AV network, it is not viable to compute locally or offload to distant servers. The proposed algorithm appears to suit the requirements of the future AV environment.

Fig. 7

Latency for C tasks.

Another important metric for determining the compute load on RSU is the number of tasks that are processed locally. The decision on whether to offload to cloud or execute locally at RSU is influenced by task deadlines. When compared to tasks with relaxed deadlines, projects with tight (or short) deadlines appear to require immediate processing. A short task deadline appears to increase the percentage of tasks that are processed locally. The number of tasks sent to cloud increases when task deadlines are relaxed, and the percentage of local processing at RSU decreases. When compared to HOS and MOTM, the average number of local processing jobs is reduced by up to 8.33 percent and 19.86 percent, respectively.

The task’s completion time grows exponentially in proportion to the task’s data size, as seen in the Fig. 8. When compared to the proposed EA-DTOCTE algorithm, the fuzzy reinforcement learning completion time increases rapidly. For task execution, the proposed technique considers both the remote and local environments. When compared to existing methods, the proposed algorithm takes less time to complete and the Completion time of tasks of the proposed system is shown in Fig. 9.

Fig. 8

Local Processing tasks Vs deadlines.

Fig. 9

Completion time of tasks.

In Fig. 10, we show how vehicular density affects latency for EA-DTOCTE, FRL, MOTM, and HOS. The discrepancy between proposed algorithm and HOS observed is lower, but as the number of vehicles increases, HOS’s latency rapidly increases when compared to proposed algorithm. Because each RSU’s coverage area is narrower, only a small number of AVs may be found within it. A separate back haul link connects each RSU to the MCC, which speeds up data transmission.

Fig. 10

Latency vs N.o of vehicles.

When comparing the EA-DTOCTE and FRL, MOTM, HOS approaches, as illustrated in Fig. 11, the task completion time varies significantly. In FRL, the average task completion time is gradually increasing. This is owing to the fact that the load on the core network grows as a result of the back-and-forth data transmission, yet the proposed solution works effectively.

Fig. 11

Completion Time vs N.o of subtasks.

Figure 12 depicts how vehicle speed affects the rate at which packets are delivered. Due to the fact that packets must be sent to all available service vehicles, fuzzy reinforcement learning has a lower packet delivery rate. Therefore, certain packets that cannot be transferred to a service vehicle before the maximum waiting delay must be dropped. Hierarchical offloading results in less packet loss than the Minimum offload time algorithm since a task is assigned to a single service vehicle. Therefore, fewer packets were delivered. The proposed method records the highest packet delivery rate because tasks are broken down into smaller tasks before being sent to the appropriate service trucks.

Fig. 12

Packet delivery ratio.

The proposed strategy, which has the maximum successful packet delivery compared to the other schemes, produces the highest throughput, as shown in Fig. 13.

Fig. 13

Throughput.

In order to compare the proposed distributed task offloading strategy’s performance, it is evaluated with state-of-the-art methods like task partitioning, mobile cloud computing, and cooperative task scheduling techniques. Figure 14. displays the outcomes. The task offloading strategy suggested has the best performance, as shown by the aforementioned figure, when compared to other techniques. Local computing is used in the task prioritizing technique, and all mobile devices finish the computation locally. The energy overhead of local computation of computing tasks is substantially higher than other offload algorithms due to the constrained processing capability of mobile devices and the high energy consumption per unit of CPU. In mobile cloud computing, each mobile device is randomly assigned access to the MEC using a random calculation offload technique. Since the access is randomly assigned, it may result in more mobile devices accessing the same MEC at the same time, increasing device interference and the energy required for computing operations to transmit data.

Fig. 14

Comparison of energy cost for state of art methods.

The proposed approach uses less energy than the cooperative task scheduling technique, which selectively offloads some jobs to MEC. However, the offloading algorithm utilized is ineffective, therefore it consumes more energy. According to the findings, the suggested T²EC² can reduce system energy usage by up to 28% when compared to other state-of-art methods.

The computation time increased as task size increased, as shown in Fig. 15, demonstrating that the proposed approach has the lowest energy usage and computing time. In comparison, the longest computation time is associated with mobile cloud computing. Even if task division takes less time than local computation, transmission still takes more energy and time. For instance, the computation time was 0.5 s when the task size was 500 KB. For other state-of-the-art approaches, the computation time was 7, 7.5, and 8 s, respectively.

Fig. 15

Comparison of computation time for state of art methods.

The offloading of tasks makes it possible to process calculations of engine speeds in subsequent vehicles. Fuel consumption and fuel economy are adversely affected by speeding due to various road conditions. A vehicle’s fuel economy appears to drop abruptly above 50 mph, regardless of the speed at which it achieves its best performance. Fuel efficiency can be increased by 7 to 14% when speed is reduced by 5 to 10 mph. For light-duty vehicles, for instance, every 5 mph over 50 mph you travel is equivalent to paying 14.78 rupees per gallon of petrol (based on the current petrol price of 215.89 rupees per gallon).

5 Conclusion

In this paper, we have proposed a novel Three Tier Edge cloud computing (T² EC²) system which uses an Energy-aware Dynamic Task offloading and collaborative task execution algorithm (EA-DTOCTE) for multilayer vehicular cloud computing networks. Task offloading, decision-making engine, and collaborative task execution were proposed as three modules. Using an offload decision algorithm, the decision-making engine determines whether and where the task should be offloaded. The algorithm takes into account the task’s energy usage in both the local and remote environments. It is also suggested that collaborative task execution be used to reduce the overall energy usage of the submitted tasks. Furthermore, simulation findings show that the suggested architecture is effective for collaborative task execution, reducing local and remote energy consumption while increasing completion time. The proposed architecture outperforms classical MCC and local offloading, according to the results of the performance analysis. The results show that the proposed T²EC² can save up to 28% of system energy consumption compared with other state-of-art techniques. The experimental findings suggest that the proposed algorithm can find a wide number of optimal alternatives for adjusting the offloading choice in real-world situations. For achieving superior performance in terms of latency, energy consumption, completion time and processing tasks, the significance of RSUs is important. In real world the required distribution of RSUs is importance to improve the connectivity to the vehicle. Finding the appropriate number of RSU Installation is the future work of this work. And also, for avoiding conflict between vehicle and RSU priority-based task offloading rule has to be considered in future, to improve the traffic efficiency and avoid the unnecessary delays of emergency vehicle processing.

Footnotes

Acknowledgment

The author with a deep sense of gratitude would thank the supervisor for his guidance and constant support rendered during this research.

References

Seid

A.M.

, Boateng

G.O.

, Mareri

, Sun

and Jiang

, Multi-agent DRL for task offloading and resource allocation in multi-UAV enabled IoT edge network, IEEE Transactions on Network and Service Management 18(4) (2021), 4531–4547.

Wang

, Li

, Wu

, Chen

and Sun

, Joint optimization of task offloading and resource allocation based on differential privacy in vehicular edge computing, IEEE Transactions on Computational Social Systems 9(1) (2021), 109–119.

Islam

, Debnath

, Ghose

and Chakraborty

, A survey on task offloading in multi-access edge computing, Journal of Systems Architecture 118 (2021), 102225.

Rahmani

A.M.

, Mohammadi

, Mohammed

A.H.

, Karim

S.H.T.

, Majeed

M.K.

, Masdari

and Hosseinzadeh

, Towards data and computation offloading in mobile cloud computing: taxonomy, overview, and future directions, Wireless Personal Communications 119(1) (2021), 147–185.

Eshratifar

A.E.

and Pedram

, Energy and performance efficient computation offloading for deep neural networks in a mobile cloud computing environment. In Proceedings of the 2018 on Great Lakes Symposium on VLSI, Chicago, IL, USA 2018, pp. 111–116.

Feng

, Krunz

and Zhang

, Joint task partitioning and user association for latency minimization in mobile edge computing networks, IEEE Transactions on Vehicular Technology 70(8) (2021), 8108–8121.

Sun

, Hou

, Cheng

, Wang

, Zhou

, Gui

and Shen

, Cooperative task scheduling for computation offloading in vehicular cloud, IEEE Transactions on Vehicular Technology 67(11) (2018), 11049–11061.

Vemireddy

and Rout

R.R.

, Fuzzy reinforcement learning for energy efficient task offloading in vehicular fog computing, Computer Networks 199 (2021), 108463.

Tseng

F.H.

, Cho

H.H.

, Chang

K.D.

, Li

J.C.

and Shih

T.K.

, Application-oriented offloading in heterogeneous networks for mobile cloud computing, Enterprise Information Systems 12(4) (2018), 398–413.

10.

Jang

, Na

, Jeong

and Kang

, Energy-efficient task offloading for vehicular edge computing: Joint optimization of offloading and bit allocation. In 2000 IEEE 91st Vehicular Technology Conference (VTC-Spring), IEEE, Antwerp, Belgium 2020, pp. 1–5.

11.

Rasheed

, Anwar

, Sudheera

K.K.

, Chong

P.H.

, Liu

, Yaqub

M.A.

and Jafri

M.R.

, Application-Aware Hierarchical Offloading for MEC-Enabled Autonomous Vehicle Architecture. In 2020 IEEE Globecom Workshops (GC Wkshps) IEEE, Taipei, Taiwan 2020, pp. 1–6.

12.

Zhu

, Luo

, Liu

, Bhuiyan

M.Z.A.

and Zhang

, Multiagent deep reinforcement learning for vehicular computation offloading in IOT, IEEE Internet of Things Journal 8(12) (2020), 9763–9773.

13.

Apostolopoulos

P.A.

, Fragkos

, Tsiropoulou

E.E.

and Papavassiliou

, Data offloading in UAV-assisted multi-access edge computing systems under resource uncertainty, IEEE Transactions on Mobile Computing 22(1) (2021), 175–190.

14.

Gholivand

and Movahedi

, A Cloud-RAN based end-to-end computation offloading in mobile edge computing, Computer Communications 175 (2021), 193–204.

15.

Khayyat

, Elgendy

I.A.

, Muthanna

, Alshahrani

A.S.

, Alharbi

and Koucheryavy

, Advanced deep learning-based computational offloading for multilevel vehicular edge-cloud computing networks, IEEE Access 8 (1370), 52–137062.

16.

Appathurai

, Sundarasekar

, Raja

, Alex

E.J.

, Palagan

C.A.

and Nithya

, An efficient optimal neural network-based moving vehicle detection in traffic video surveillance system, Circuits, Systems, and Signal Processing 39(2) (2020), 734–756.

17.

Appathurai

, Manogaran

and Chilamkurti

, Trusted FPGA-based transport traffic injects, impersonate (I 2) attacks beaconing in the Internet of Vehicles, IET Networks 8(3) (2019), 169–178.

18.

Lakhan

and Xiaoping

, Energy aware dynamic workflow application partitioning and task scheduling in heterogeneous mobile cloud network. In 2018 International Conference on Cloud Computing, Big Data and Blockchain (ICCBB), IEEE, Fuzhou, China 2018, pp. 1–8.

19.

Vantsevich

V.V.

and Blundell

M.V.

, Advanced autonomous vehicle design for severe environments, IOS Press 44 (2015), 1–317.

20.

Balakirsky

S.B.

, A framework for planning with incrementally created graphs in attributed problem spaces, IOS Press 270 (2003), 1–247.

21.

Vankadara

and Dasari

, Energy-aware dynamic task offloading and collective task execution in mobile cloud computing, International Journal of Communication Systems 33(13) (2020), e3914.

22.

Tutsoy

and Barkana

D.E.

, Model free adaptive control of the under-actuated robot manipulator with the chaotic dynamics, ISA Transactions 118 (2021), 106–115.

23.

Lin

, Huang

, Zhang

, Yang

and Zhao

, A deep reinforcement learning based computation offloading with mobile vehicles in vehicular edge computing, IEEE Internet of Things Journal (2023).

Dynamic task offloading and collaborative task execution using three tier edge cloud computing (T 2 EC 2 ) system for autonomous vehicles

Abstract

Keywords

1 Introduction

2 Related works

3 Proposed Three tier Edge cloud computing (T2 EC2) system

4 Results and discussion

Footnotes

Acknowledgment

References

3 Proposed Three tier Edge cloud computing (T² EC²) system