Multi-Layer Latency Aware Workload Assignment of E-Transport IoT Applications in Mobile Sensors Cloudlet Cloud Networks

: These days, with the emerging developments in wireless communication technologies, such as 6G and 5G and the Internet of Things (IoT) sensors, the usage of E-Transport applications has been increasing progressively. These applications are E-Bus, E-Taxi, self-autonomous car, E-Train and E-Ambulance, and latency-sensitive workloads executed in the distributed cloud network. Nonetheless, many delays present in cloudlet-based cloud networks, such as communication delay, round-trip delay and migration during the workload in the cloudlet-based cloud network. However, the distributed execution of workloads at different computing nodes during the assignment is a challenging task. This paper proposes a novel Multi-layer Latency (e.g., communication delay, round-trip delay and migration delay) Aware Workload Assignment Strategy (MLAWAS) to allocate the workload of E-Transport applications into optimal computing nodes. MLAWAS consists of different components, such as the Q-Learning aware assignment and the Iterative method, which distribute workload in a dynamic environment where runtime changes of overloading and overheating remain controlled. The migration of workload and VM migration are also part of MLAWAS. The goal is to minimize the average response time of applications. Simulation results demonstrate that MLAWAS earns the minimum average response time as compared with the two other existing strategies.


Introduction
These days, cloud computing-based smart applications are growing progressively to deal with different life activities. The applications are E-Healthcare, E-Banking, Augmented Reality and E-Commerce [1]. Cloud computing offers the abandonment of being ubiquitous and omnipresent in order to entertain the requests of applications. However, conventional cloud computing, located multiple hops away from E-Transport, and the offloading of the workload of applications, will suffer from long end-to-end latency. Cloudlet is an extension of cloud computing that brings computing resources to the edge of the communication network. Cloudlet is a latency-optimal paradigm and is easily connected with the network's base stations, controlled by the software-defined network. Offloading and workload assignment are two fundamental techniques in distributed cloudlet-based cloud computing networks where E-Transports send their workload for execution. Generally, it is called full offloading from E-Transport devices to the cloudlet-based cloud network [2].

•
Proposed System: Initially, this study proposes a mobile cloudlet-based cloud system where E-Transports can easily offload their coarse-grained applications. The mobile devices work as a thin client, where E-Transport IoT applications only offload their workload via a client cloudlet module. The cloudlet host is a distributed server-based network in which all servers are heterogeneous, and each workload is assigned a single virtual machine as a guest to run. The cloud remote is the conventional cloud that is the part of the system which supports the migrated workloads for further execution. The mobile cloudlet based cloud (MCBC) is a distributed system in which the workload can execute at different computing nodes during offloading and execution. • To solve the workload assignment in a distributed MCBC network, this study proposes the MLAWAS framework, which consists of different schemes which solve the problem into different steps. The schemes are defined in their next specific points in detail. • Initial Optimal Heuristic: This study proposes the improved Genetic Algorithm (IGA) based on a natural selection mechanism that mimics biological evolution to solve constrained and unconstrained optimization problems. A workload population of individual solutions is updated repeatedly by the algorithm. An optimal solution reaches the best point in the workload population. • Global Searching: For optimization, improved Genetic Algorithms are a global search technique. They are bad at hill-climbing. However, the potential for probabilistic hill-climbing is improved Simulated Annealing. Therefore, by adding a mutation operator such as Simulated Annealing and an adaptive cooling schedule, the two techniques are combined here to create an adaptive algorithm that has the merits of both Genetic Algorithms and Simulated Annealing. • Local Greedy Searching: The greedy-based search is proposed to deal with local search and to find the optimal solution of any requested workload within available cloudlets. However, it is working in the same environment. The main goal of this method is to verify that all workloads assigned to computing nodes have optimal average response time solutions compared to others. This searching repeats iteratively until all workloads are assigned to all available component nodes. • Q-Learning Aware Migration Technique: This study suggests a Q-Learning-based VM migration method where the current optimal solution being treated as a state is an improvement on the objective solution, and the workload to be migrated to benefit E-Transport benefits mobility. The Q-Learning method always adopts any dynamic changes of the environment and always chooses an optimal solution during the migration of the workload to maintain the workload and the overhead situation in the MCBC network.
The rest of the paper is planned as follows: Section 2 discusses the literature on the latency optimal workload assignment distributed network. Section 3 contains a mathematical model of how to define all kinds of delays with their constraints. Section 4 clarifies how we utilize the techniques chosen to lower the delays in conjunction with the mathematical model. Section 5 presents the simulation system and the results of the discussion of different schemes. Section 6 presents the summary of the study and the future direction of the research.

Related Work
The latency optimal workload assignment problem in the mobile cloudlet based cloud network (MCBC) has been widely studied in the literature. This problem is formulated as a linear integer problem, a greedy problem and a combinatorial problem by different researchers in the literature. Table 1 describes the existing problem formulation and the proposed schemes for the workload assignment problem in the MCBC network. Nouha et al. [1] formulated latency and flexible workload assignment distributed cloudlet networks. The goal is to minimize the response time of applications, which is determined by process delay. Binwei et al. [2] suggested a framework to deal with an energy and latency aware workload assignment in distributed edge cloud computing. The communication delay and error rates are taken as constraints, and the objective is to minimize the communication of applications during offloading and to improve the battery life of devices. Xiang et al. [3] formulated a single latency aware workload assignment in the cloudlet network, where the communication delay and the process delay are taken as constraints and the average response time of the application is the objective of the study. Qiang et al. [4] and Shanhe et al. [5] formulated latency aware workload problems for computing-intensive real-time workloads and healthcare applications in the distributed edge network. These studies tried to reduce the end-to-end latency of applications via offloading and resource allocation methods. Jungmin et al. [6] and Amit Samanta et al. [7] proposed latency optimal workload assignment greedy strategies to optimize the objective function of applications. The objective function was to communicate delay and process delay during workload allocation in the distributed edge cloud network. Huy Trinh et al. [8,9] devised solutions based on dynamic programming based iterative greedy solutions to solve an energy-efficient and low-latency optimal workload in distributed edge cloudlet networks. These studies optimize the objective function of the problem, that is, the energy and latency, and obtained near-optimal solutions in the iterative process.
A.Rasheed et al., Lixing et al. and Dileep et al. [10][11][12] proposed dynamic programmingbased iterative and full approximation methods based on Lyapunov optimization heuristics to solve the workload assignment problem in the distributed edge cloud network. The concave and convex optimization-based solution suggested solving the travelling salesmen's problem for workload in the network. All studies considered the energy and latency objectives during problem formulation and met the applications' Quality of Service requirements. Li et al. [13][14][15] and Ying Wah et al. [16] suggested energy, latency and cost-aware workload assignments in the distributed mobile edge/fog/cloudlet based cloud network. These studies solved the workload assignment based on NP-Hard scheduling heuristics and meta-heuristics. The goal was to minimize the makespan of applications. The considered problem is dynamic, and a Queue based solution is suggested to estimate the required resources for offloaded workload in the system. The communication delay, process delay and energy cost were taken as constraints of the study.
Aburukba et al. and Lakhan et al. [17] suggested mobility-aware workload assignment solutions for the distributed network. The processing delay, communication delay and prorogation delay are also considered during problem formulation and, based on these delays, the studies proposed solutions. The studies [19] suggested dynamic environment workload assignment aware strategies based on Genetic Algorithms and ant-colony meta-heuristics. The solutions are fully approximated, and exponentially gained optimal objectives in an exploration and exploitation environment. The studies [20][21][22][23] formulated dynamic content and failure-aware workload assignment-based scheduling and resource allocation techniques. Different heuristics, such as Hill Climbing, Earliest Deadline First, and Earliest Finish Time methods, suggested minimizing the tardiness latency of applications. Wang et al. and Lakhan et al. [24][25][26][27] introduced reinforcement learning-based resource allocation techniques and workload assignment strategies in the dynamic environment, where the network contents and load balancing situation of the computing nodes are considered to be constraints. The application-level latencies, such as average response time, were booted via a fully offloading technique in the Queue-based cloudlet cloud network. Li et al. [28] proposed mobility and fault aware methods in a dynamic environment where the latency and makespan of applications were considered as objective functions. The communication delay and the process delay were the constraints, and each workload had a deadline constraint.
To the best of our knowledge, the multi-layer latency-aware workload assignment in the cloudlet-based cloud network and the Q-learning migration workload have not yet been studied for mobile cloudlet based cloud applications. This study considered the QoS of applications and the multi-layer latency of the workload (e.g., round-trip delay, process delay and migration delay) with the roaming situation of E-Transports. This study suggested dynamic mixture solutions based on meta-heuristics and convex optimization to obtain the optimal application solution in dynamic and adoptive cloudlet-based cloud environments. In this way, the study optimizes the average response time of applications and runs them under QoS requirements.

Problem Description
This study devises a novel system, the Multi-Layer Latency Aware Workload Assignment of E-Transport IoT Applications in Mobile Sensors Cloudlet Cloud Networks, as shown in Figure 1. The primary objective is to boost the performances of E-Transport applications in distributed sensors cloudlet cloud networks. The system consists of different Internet of Things (IoT) E-Transport applications. At the same time, local mobile sensors decide where to send the workload for further execution. The computing nodes are local mobile sensors, cloudlets and cloud computing, which are geographically distributed and are connected to the different communication technologies. The software-defined network control plane is where communication nodes connect to each other to provide an efficient routing path to transport applications. This study considered the multi-latency aware workload assignment in the cloudlet-based cloud network. The considered workload is coarse-grained, where the thin mobile clients offload their workloads to the cloudlet based cloud for further processing. This study suggests using the Multi-Layer Latency Aware Workload Assignment (MLAWAS) framework, which consists of different global searching and local searching iterative methods and Q-Learning heuristics.
Mobility States Figure 1. IoT E-Transport System.

Application Workload Characterization
In this work, all workloads are full offloading coarse-grained, consisting of different types of tasks. The Poisson process follows the arrival of the workload to the MCBC. The M/M/S-based queue in the system immediately allocates these workloads to the available optimal cloudlets for further execution. All workloads have QoS requirements (e.g., maximum latency tolerance).

Sensors Cloudlet Cloud Assignment Mechanism
This study formulates multi-latency aware workload assignment problems in cloudletbased cloud networks where the objective is to minimize the average response of the workloads. This study proposed an offloading and execution system based on MLAWAS in a cloudlet-based cloud network as shown in Figure 2. There are three main layers in the system. The first layer is a mobile client where all applications are installed; the cloudlet client for offloading is asked if the total average response is optimal for offloading. The second layer is the cloudlet host, which consists of many homogenous servers. The client host consists of different virtual machine (VM) host components, which wrapped all coarse-grained applications into a single VM. The VM manager allocates the single computing node to a workload as a guest VM for execution. The cloudlet meta-data and service discovery are two components that aim to save application data and monitor the remaining resources of the cloudlet host during execution. The third layer is the remote cloud, which only executes the migrated workload of the application from the cloudlet host. The remote cloud also has meta-data and a discovery service, which saves the completed data and monitors cloud computing resources during the execution of the workload in the cloud computing. MCBC is a distributed network in which mobile devices are treated as thin clients, and cloudlet based clouds are thick computing nodes for processing the requested workload to their QoS requirements.

System Model
This study considered the fine-grained workload of applications. The computing environment is geographically distributed in the network as shown in Figure 3, and it is similar to existing work [4]. There are three main aspects of the considered work: mobile E-Transports, cloudlet, and cloud computing. All the cloudlets are placed flexibly and are connected to the different base stations in the network. At the same time, all base stations are managed by a Software Defined Network (SDN) to provide an efficient routing path to the E-Transports during offloading in the network. The study also considered cloud computing services that execute the migrated workload of applications from cloudlet servers to avoid overloading. Each cloudlet is made up of heterogeneous servers where they have different computing speeds and resources. Multi-latency-aware aspects are considered in this study-communication latency (round-trip delay and migration) and process delay.

Multi-Latency Based Problem Formulation
We examine the MCBC as the amalgam of a set of base stations (BS) and a set of mobile E-Transports under the coverage base station, where each base station is outfitted with one cloudlet as shown in Figure 3. As we already explained, the cloudlet is an extended small data center for computing and storage abilities. Every individual cloudlet has a set of similar types of virtual machines. Mobile E-Transports can execute and offload the task to the closest cloudlet, and uploading and downloading the data from the cloudlet are identical. The arrival rate of the mobile E-Transport i application workload requests follows a Poisson process, represented by λi, whereas i ∈ I. We describe c i as the number of computation tasks generated by mobile E-Transport i, whereas d i symbolizes the quantity of data to be downloaded by mobile E-Transport i. Let k be the range of the cloudlet k for data processing and storage. The average service rate of the mobile device and the cloudlet are defined as θ i and µ k . We mock-up each base station B and mobile E-Transport i inside the coverage area of the base station (BS) as an M/M/1 queuing system. Normally, MU sends data requests, and then either the cloudlet or the remote cloud provides the results. This procedure engages the transmission between MU and the BS and the BS to the remote cloudlet. In simple terms, we define the e UM transmission delay between MU and BS, whereas e MC is the transmission delay between the base station and a remote cloud.

Process Delay and Network Delay
Here, we are using binary variables to show the vector dimension of all delays such as X = {x ik : k ∈ K, i ∈ I} with x ik = {0, 1} shows whether the mobile E-Transport accesses the associated cloudlet or not, that is, In Equation (1), vector Y = {x ik : k ∈ K, i ∈ I} shows whether the computational task to offload is generated by the mobile E-Transport i to the cloudlet k or not, likewise, Equation (2) determines the offloading of the application workload either to the cloudlet or the cloud.

Problem Formulation
We use the queuing model M/M/1-PS [19] to calculate the processing delay of any computing node k as follows: Equation (3) determines the execution on the cloudlet, where y ik c i λ i is the workload computation task offloaded to cloudlet k, and In the same way, the computation delay on computing node k occurs in the following way: Equation (4) determines the computation delay at the mobile device, where 1 − ∑ k∈B c i λ i is the computation task amount that is locally executed at the mobile E-Transport device If the request for data access to the cloudlet server is k, then the transmission delay between the mobile E-Transport and the cloudlet is as follows: Equation (5) determines the communication delay between E-Transport and cloudlet. y ik c i shows the volume of data at cloudlet k requested by the mobile E-Transports.
Additionally, the delay between the cloudlet and the remote cloud for the additional process is expressed as follows: Equation (6) calculates the communication delay between the cloudlet and the cloud computing. (1 − x ik ) =0 shows that the requested amount of content for the processing is stored in the cloudlet k and the mobile E-Transport does not need to access the remote cloud.
Finally, the total Service Delay of all mobile E-Transports is the amalgam of the transmission delay and execution delay expression as follows: Consequently, we know it is an optimization problem and can be formulated as an optimization problem, which minimizes the τ with various constraints as shown in Equation (7).
Equation (8) determines the objective function of the study.
Equation (9) shows that the requested workload of all applications must be less than the capacity of any computing k.
Equation (10) shows that all computing nodes easily handle overloading and overhead situations before offloading and workload assignment in the network.
All computing nodes, such as mobile computing, cloudlet node and cloud node, must be stable in the network as defined in Equation (11).
Each computing node can execute one workload at a time as defined in Equation (12).
Each workload can be allocated to any single computing node as defined in Equation (13).
The respective Equations (14) and (15) determine the binary variable, which shows that either the workload is assigned to any computing node during offloading and workload assignment in the network.
All equations are numeric, which shows that the considered problem is a dynamic integer problem.

Proposed MLAWAS Algorithm Framework
This study considered the multi-latency aware workload assignment in a cloudletbased cloud network. The considered workload is coarse-grained, where thin mobile clients offload their workloads to the cloudlet based cloud for further processing. This study suggested the Multi-Layer Latency Aware Workload Assignment (MLAWAS) framework, consisting of different global searching and local searching iterative methods, and Q-Learning heuristics. We defined the entire process of workload assignment based on MLAWAS components, as shown in Algorithm 1.

Algorithm 1: MLAWAS Algorithm
Input : cloudlet service rate µ k ; Mobile device θ i ; Output : Workload assignment of the mobile to optimal cloudlet S = {x ik | i ∈ I, k ∈ K}; 1 begin 2 foreach (i to I) do 3 Call Modified Genetic Algorithm for Initial Workload Assignment; 4 i ← k based on equation (7); 5 Improve the global search call Simulated Annealing Algorithm method; Call LATAS method to improve the solution within local space; 8 Call VM migration method to move a workload from cloudlet to cloud; 9 Swapping Workload i ← from node k1 to k2 ; 10 The migration will be done based on VM migration scheme;

End MLAWAS Framework
We defined the respective techniques of MLAWAS in the following subsections.

Improved Genetic Algorithm for Initial Workload Assignment
A mixture evolutionary optimizer algorithm was developed for the Service Delay. The proposed improved GA (Genetic Algorithm) and SA (Simulated Annealing) are merged to provide the optimal geo-distributed cloudlet network. The proposed algorithm is proficient at defeating the untimely convergence of the Genetic Algorithm and helped the escape from the optimal local solution. We have proposed the modified classical Genetic Algorithm to resolve the optimization problem efficiently.
The above-given problem is not easy and is non-linear for the reason that x ik and y ik both have different directions. We implement our problem in a modified Genetic Algorithm and a Simulated Annealing algorithm, compared with classical Simulated Annealing and genetic heuristics. We compare individual components with classical heuristics and find some improvement as follows: 1.
In the Genetic Algorithm (GA) cross-process, we replicate the Simulated Annealing (SA) operation for incoming new individuals that can avoid heuristics converging to the optimum local value with some low fitness individuals passing onto the next generation.

2.
The fitness function calculated the incoming individual fitness value after the crossover operation and the mutation operation, correspondingly. This way, our techniques save the best individuals from being lost during the process of evolution.
It is worth remembering that the individuals are promising solutions to the given problem that is to be optimized.
Fundamentally, as articulated in Algorithm 1, given the solution n s = {x ik y ik } where k ∈ K and i ∈ I are the 2D-dimensional vectors that can be filled with either 1 or 0, where the first D represents the value of x ik and the second D represents the value of y ik , respectively. The intention of the feasible solution given in the proposed heuristics was to guarantee all constraints. Furthermore, the solution of s with the given fitness function is f (s) and the probability that every individual would be inherited by the incoming next generation is P(s) as follows: Equation (16) shows the probability of the function value. There is another type of cumulative probability as follows: Equation (17) shows the commutative probability of the objective function during the initial assignment of workloads. In Algorithm 2, the original population updated by steps 8-12, Pop, selects the better individual with respect to Q(k). Steps 15-21 show the crossover procedure from a solution; it randomly selects a figure of bit positions and then replaces the two bits in the same position as two individuals.
Step 22 provides the update value of Pop 1 to support Algorithm 3. In Algorithm 3, we measure all the individuals that belong to two dissimilar Pop 1 and Pop 2 populations. There exists the probability that bad individuals will be inherited by the subsequent generation with a certain probabilistic expression such as exp(−( f (n s ) − f (n s ))/T)).
Steps 16 and 30 provide the surety that the crossover procedure and the mutation solution strictly satisfy all constraints in their given equations. Steps 24-26 describe the comparison between Pop 1 and Pop 2 and record the lowest fitness values of both of them. Steps 28-35 express the procedure of the mutation, while step 32 initializes or generates the new optimal solution ni c and turns over (0 ← 1 or 1 ← 0). Furthermore, Steps 38-40 randomly recorded the lowest fitness values and compared them with the fitness values in steps 37 and 25, which is the optimal solution that we observed and found.
Algorithm 4 always returns the optimal solution through the global search space, whenever a Genetic Algorithm returns a feasible solution locally; then Simulated Annealing helps it to escape from the local optima.

E-Transport Applications Mobility and Migration Latency
The E-Transport i ∈ I roams in the geographically distributed cloudlets k ∈ K through the cellular network from base station to base station along with the VM service that is in progress, as shown in Figure 4. It needs to be migrated to the E-Transport, but this process is stochastic and network behaviors periodically change; this leads to many challenges associated with propagation latency, such as congested nodes, high traffic and lower bandwidth. Our goal is to migrate the running VM and transport application with lower propagation delay without interruptions and disruptions to the VM service. Our method always finds the optimal shortest path among all environment spaces. Here, we introduce the utility table of each cloudlet, which saves the data of the current E-Transport and the data to be migrated via the VM migration method, and we save another utility table of the cloudlet server as shown in Figure 4. The E-Transport mobility scenario will be discussed in the following way.

E-Transport Applications Mobility Scenario
All E-Transport applications can move from one place to another place, for example, during a business trip or roaming between different locations. The flexible cloudlet-based cloud is geographically distributed in the network, and all computing nodes are connected via base stations. If the application i = 1 migrates the transport from one node k = 1 to another node k = 2 and the utility table that transfers the workload runs the information to another cloudlet for further execution, the main advantage of the mobility and migration of the workload is avoiding overloading and overhead. Figure 4 shows the scenario of E-Transports roaming between different base stations, and all cloudlet and cloud services are ubiquitous and omnipresent.

Input :
size R of the population; Probability P o of the crossover; the P 1 is the probability of the mutation; Number of iteration is N Output : τ min 1 begin 2 initialization; 3 n s is the randomly generated of feasible solutions K ; We model the propagation latency as follows: 1.
The set of network and agent state cloudlets k ∈ K.

2.
The set of VM migration actions a ∈ A of the agent.

3.
P a (k, k ) = P r (k t+1 = k | k t = k, a t = a) describes the probability of state k to state k transition in action a.

4.
R a (k, k ) describes the instant reward following k to k transition in action a.

5.
All kinds of rules that explain what the agent cloudlet Controller observes.
All rules are stochastic because the behaviour of distributed cloudlet networks randomly changes node to node. It depends upon the cases. Sometimes agents know to observe the current environment cloudlet k ∈ K state; sometimes it is fully observed and sometimes partially observed. All sets of action a ∈ A are stopped if it reaches zero. Furthermore, it is not reducible. Because of the stochastic nature of networks, learning agents in cloudlet environments follow discrete time t, under action a t ; they contain a set of reward values from a set of actions a t ∈ A. The E-Transport moving from cloudlet state to cloudlet state can be expressed as k t + 1 and reward factor r t + 1 belongs to the transition process and is expressed as (k t , a t , k t + 1).

Algorithm 4: LATAS Algorithm
Input : The request of all transport application workloads applications generated in descending order ; 4 i * = 1; 5 while (i * ≤| I |) do 6 Find the underutilized optimal cloudlet with lower delay for transport application i * , based on equation ; 7 Set x * ik * = 1; In our problem, the data model of the cloudlet environment is known. Still, we are looking for the optimal solution, so this is the planning problem and it must be converted into a machine learning specified problem.
The learning cloudlet agent requires an optimal searching mechanism to randomly select an action without degrading the QoS, for example, the latency and bandwidth. We solve our problem of the principle of optimality with respect to two factors which are as follows: The action in the current cloudlet agent is replicated as a plan called policy π. The cloudlet policy: In Equation (18), the policy π always provides the possibility in cloudlet k state while taking action a.
The cloudlet value function: Value function V π (k) is expressed as the preliminary predictable return with cloudlet state k and policy π; the cloudlet value function approximates how high the quality is in the given cloudlet state.
Equation (19) measures the value function, whereas the randomly generated variable R shows the return and is expressed as the sum of the value of discount rewards; in R = ∑ ∞ t=0 γ t r t , value t t shows a reward at step t and γ ∈ [0, 1] expresses the discount rate. Algorithm 5 determines workload migration from one computing node to another computing node based on the Q-Learning model. States show that a new optimal objective function must be the best as compared to the current solution. Figure 5 shows that the trained model in Q-Learning performs well, and improves and chooses the optimal computing node when the workload migrates between computing nodes.

Experimental Design
The performance evaluation shows the effectiveness of the proposed algorithm frameworks when they run a healthcare workload in the cloudlet-based cloud network. This paper implemented the proposed framework in the simulation form. The related parameters of the simulation are described in Table 2. The coarse-grained healthcare workload consists of different types of tasks such as text, video and images. This study implemented different baseline approaches and proposed schemes as defined in the following.
The goal was to minimize execution costs and the communication of the workload during offloading and assignment in the cloud network. The method consisted of four main components: profiling of workload, workload assignment, migration, and ordering of tasks. The centralized cloud-based network implemented in these studies solves the delayed optimal workload of applications.
• Lean Algorithm: The Latency Effective Aware Algorithm is a framework which allocates workload to different available cloudlets. The framework consists of the assignment part and the migration part of the applications. This framework has been implemented in different studies to solve the healthcare workload of applications [9][10][11][12][13][14][15]. • MLAWAS Framework: This proposed method is composes of different components, such as assignment, migration, local and global search, and is considered the dynamic environment in this study.

Simulation Parameters
First scenario (static): We examined the performance of the proposed MLAWAS algorithm in static traffic conditions, where vehicles do not move in this simulation scenario. The static traffic scenario allows us to demonstrate the proposed algorithm's fundamental characteristics. The maximum density and standard deviation (SD) were set to 55 cars in a 100 m radius and 1000 m, respectively, on a straight double lane road (5 km). In the second scenario, vehicle mobility was taken into account, and the algorithm E-Transport Mobility and Migration latency was used to handle both application mobility and workload migration.
The simulation was performed on local machines, where baseline classes were taken from edgecloudsim and VANET++ [14] vehicular environment integrated into edgecloudsim [13]. The goal was to exploit the the available abstract of edgecloudsim and VANET++ for the applications' cloudlet implementation and vehicular environment. For all the cloudlets K, in the simulation environment, we implemented E-Transport classes of edgecloudsim and tested on the genymotion [15] environment with the first scenario. In the second scenario, we named the VANET++ classes and implemented waypoint mobility classes to analyze the mobility and migration of the workload in the system. The simulation parameter in Table 2 shows that the system implemented in JAVA and XML allows communication among different nodes such as cloudlet, cloud and applications interfaces. All the applications, as shown in Table 3, are E-Transport types in various domains such as E-Bus, E-Ambulance, and E-Taxi. These applications covered the healthcare aspects and e-passengers applications, which were integrated into the simulation environment at the config of edgecloudsim during simulation. This study implemented coarse-grained workloads with different types of tasks as defined in Table 3. The cloudlet based cloud network was deployed to run the workload tasks in this study. The cloudlet based cloud network configuration is defined in Table 4.  Table 4 defines the resource specification of the computing nodes with their characteristics and features.

System Implementation and Mobility Aware Dataset
This study designed transport applications on the specific interface, for example, the GenyMotion emulator of the android phone, which is then the directly connected base station. These base stations are geographically distributed and connected with distributed cloudlets and are controlled by the SDN controller. The study implemented a real-time dataset https://everywarelab.di.unimi.it/lbs-datasim (accessed on 18 May 2021), which is implemented in this work [29] on mobility for E-Transport movements in the network.
The study implemented real-time E-Transport mobility dataset-based simulation libraries, as defined in Table 5, in which the distance between the two base stations was about 1 KM. Thirty base-stations were considered with three cloudlet networks. Furthermore, 100,000 E-Transport users were considered during the simulation in this study. Figure 6 shows the map of the road of Milan where this dataset was implemented in practice, and we used this open-source data for the mobility features of the study.

Road Network
Simulation-Area

Data Performance Method
The study tested four benchmark mobile cloud applications; their specifications are presented in Table 3. We tested applications that generate data (i.e., config file obtained data via profiling technologies and task scheduling heuristics) from different applications via analysis of variance (ANOVA). At the same time, ANOVA is an efficient parametric technique for examining the algorithm-generated data of mobile cloud applications from experiments. We exploited t-tests and dependent and independent random variables using the one way ANOVA method in order to assess the proposed method's efficiency.

Result Discussion
This section presents the evaluation performances of different methods to solve the workload assignment problem of applications. The results are compared via different methods and the performances of the workloads are compared with reference to their constraints during execution. The component performances are presented in their different subsections.

Communication Delay
Initially, the study considered the full offloading technique to migrate the workload from E-Transport devices to a cloudlet-based cloud network for execution. Due to the implementation of cloudlet and cloud in different networks, there is a communication delay during the offloading of the workload to the computing nodes. MLAWAS allows the workload when there is a shorter communication delay in the cloudlet or the cloud for execution. LEAN and conventional methods also focused on the communication delay during offloading; however, they did not focus on the delay between the wireless network and computing nodes, such as the cloudlet and cloud network-in this way, the communication increases in the different time intervals. Figure 7 shows that MLAWAS gained a shorter communication delay during the offloading of the workload as compared to the LEAN and the conventional method. The round-trip time (RTT) is shown in seconds during offloading and execution in the cloudlet-based cloud network. The E-Transport determines the round-trip from the wireless to the cloudlet or cloud computing nodes. Figure 8 shows that MLAWAS gained a smaller round-trip delay during the offloading and execution of the workload in the network. There is a migration delay between E-Transport and cloudlet, and cloudlet and cloud computing during the offloading and execution processes of the workload. This controls the load balancing situation. When the cloudlet becomes overloaded, it will migrate some workload to the cloud for further execution. However, LEAN and conventional methods did not consider the migration delay in their studies. Figure 9 shows that MLAWAS gained a shorter migration delay as compared to existing baseline methods during the migration of the workload from cloudlet to cloud for further processing. There is a migration delay between E-Transport and cloudlet, and between cloudlet and cloud computing during the offloading and execution processes of the workload. In these two ways, we migrate the workload in the study. Figure 9 shows the performance of workload migration from cloudlet to cloud. However, the study also considered the virtual machine migration between cloudlet and cloud. Figure 10 shows that MLAWAS gained a shorter migration delay as compared to existing baseline methods during the migration of workload from cloudlet to cloud for further processing.

Processing Delay
The cloudlet delay means the computing delay of the workload at different computing nodes during execution. The cloudlet based cloud offers different computing nodes to run the workload of the applications. For instance, a cloudlet is a small data center consisting of different servers to run the application in proximity to E-Transports. It is a latency optimal data center where the application has small end-to-end latency. Due to the limited resource capability of the cloudlet servers, they can further migrate the workload to cloud computing for further execution. In this way, the load balancing situation and the overhead situation will be overcome during the processing of the workload. Figures 11 and 12 show that MLAWAS outperformed existing LEAN and conventional baseline approaches in terms of the execution of workload when considering load balancing and the overhead situation of nodes. However, existing approaches only focus on computing delay based on their computing speed rather than their overloading and overhead situation. Therefore, they had a longer delay as compared to MLAWAS during the execution of the workload.

Average Response Time
The average response time of applications during offloading and execution is the study's objective and that of existing studies. The average response time (ms) combines communication time and execution time during the offloading and execution of the workload in the cloudlet-based cloud network. Communication delay takes different forms such as round-trip delay, migration delay, and offloading in the network. Figure 13 shows the performances of the average response time of the workload in the cloudlet-based cloud network. This performance indicates that MLAWAS outperforms existing LEAN and conventional baseline approaches in terms of the average response time of applications. The main reason is that MLAWAS adopts dynamic changes in the network whereas MLAWAS exploits the Q-Learning approach compared to the greedy method, which is implemented by the existing method. The limitation of greedy and heuristic methods is that they can verify the optimal performance of applications at the end or initial stages of the workload assignment. However, Q-Learning in MLAWAS calculates the performance of the application at each step during offloading and execution. That is why MLAWAS performs better than the greedy methods, which are implemented in the existing method during offloading and scheduling.  Figure 14 shows the optimal performance of the average response time of applications. The main reason is that MLAWAS adopts dynamic changes in the network whereas MLAWAS exploits the iterative approach compared to the greedy method, which is implemented by the existing method. The limitation of greedy and heuristic methods is that they can verify the optimal performance of applications at the end or initial stages of the workload assignment. However, the iterative approach in MLAWAS calculates the application's performance at each step during offloading and execution. That is why MLAWAS performs better than the greedy methods, which are implemented in the existing method during offloading and scheduling.  Figure 15 shows the optimal performance of the average response time of applications. The main reason is that MLAWAS adopts dynamic changes in the network where MLAWAS exploits the iterative approach as compared to the greedy method, which is implemented by the existing method. The limitation of greedy and heuristic methods is that they can verify the optimal performance of applications at the end or initial stages of the assignment of the workload. However, the iterative approach in MLAWAS calculates the performance of the application at each step during offloading and execution. That is why MLAWAS performs better as compared to the greedy methods, which are implemented in the existing method during offloading and scheduling.  Figure 16 shows the optimal performance of the average response time of applications. The main reason is that MLAWAS adopts dynamic changes in the network where MLAWAS exploits the iterative approach as compared to the greedy method, which is implemented by the existing method. The limitation of greedy and heuristic methods is that they can verify the optimal performance of applications at the end or initial stages of the assignment of the workload. However, the iterative approach in MLAWAS calculates the performance of the application at each step during offloading and execution. That is why MLAWAS performs better as compared to the greedy methods, which are implemented in the existing method during offloading and scheduling.

Conclusions and Future Work
Mobile cloudlet based cloud computing is an emerging hybrid computing architecture in which the workload of applications is executed at different nodes to gain lower endto-end latency. Nonetheless, many delays are present in cloudlet-based cloud networks, such as communication delay, round-trip delay, and migration delay during cloudlet-based cloud network workload. The distributed execution of workloads at different computing nodes during an assignment is a challenging task. This paper proposes a novel Multilayer Latency (e.g., communication delay, round-trip delay, and migration delay) Aware Workload Assignment Strategy (MLAWAS) to allocate the application's workload into optimal cloudlets and clouds. MLAWAS consists of different components, such as the Q-Learning aware assignment and Iterative methods, which distribute workload in a dynamic environment where runtime changes of overloading and overheating remain controlled. The migration of workload and VM migration are also part of MLAWAS. The goal is to minimize the average response time of the applications. Simulation results demonstrate that MLAWAS earns the minimum average response time as compared with two other existing strategies. In the future, we will extend our work from the security perspective, where blockchain-enabled technology and mobility of the Internet of Things (IoT) will be considered more in the current network. The work will introduce a new framework based on the serverless model-instead of a virtual machine based on container technology-which is lightweight and cost-optimal.