Hybrid trust-based optimized virtual machine migration for dynamic load balancing and replica management in heterogeneous cloud

Abstract

Cloud computing is an upcoming technology that has garnered interest from academic as well as commercial domains. Cloud offers the advantage of providing huge computing capability as well as resources that are positioned at multiple locations irrespective of time or location of the user. Cloud utilizes the concept of virtualization to dispatch the multiple tasks encountered simultaneously to the server. However, allocation of tasks to the heterogeneous servers requires that the load is balanced among the servers. To address this issue, a trust based dynamic load balancing algorithm in distributed file system is proposed. Load balancing is performed by predicting the loads in the physical machine with the help of the Rider optimization algorithm-based Neural Network (RideNN). Further, load balancing is carried out using the proposed Fractional Social Deer Optimization (FSDO) algorithm, where the virtual machine migration is performed based on the load condition in the physical machine. Later, replica management is accomplished for managing the replica in distributed file system with the help of the devised FSDO algorithm. Moreover, the proposed FSDO based dynamic load balancing algorithm is evaluated for its performance based on parameters, like predicted load, prediction error, trust, cost and energy consumption with values 0.051, 0.723, 0.390 and 0.431J correspondingly.

Keywords

Cloud computing distributed file system dynamic load balancing virtual machine migration deep fuzzy clustering

1. Introduction

The tremendous development in the field of Information Technology has lead to the rise of a new field of computing named cloud computing replacing the conventional computing methods by offering solutions to users irrespective of their location or time. The users can be charged as per the usage and are provided with access to several resources, such as network applications, storage, servers and other computing services [1, 2]. The cloud services are offered to users by multiple organizations, which are referred as the cloud service providers [1]. Cloud computing is based on the principle of transferring the computation from the conventional desktop to over the internet implying the transfer of data, services and computing to a central system that is present at an external location. The clients utilizing the cloud can scale up or down the resources as required over the internet, thus making a server oriented computing [3]. Cloud computing presents multiple advantages to the clients as well as companies with respect to funds and operation cost. Moreover, it provides a network-oriented setting where users can share resources and computations irrespective of the positions they are based [4].

Clouds available can be basically classified based on the deployment into four types, such as public cloud, private cloud, community cloud and hybrid cloud. The cloud that is present in the premises of the provider is referred as public cloud, whereas those meant for specific firms are private clouds and community clouds are meant for firms that have a common functionality and hybrid clouds are formed by the amalgamation of the private, public and community clouds [5, 27, 28]. The clouds can be classified based on the level of abstraction as, (i) Infrastructure as a Service (IaaS), (ii) Platform as a Service (PaaS) and (iii) Software as a Service (SaaS). IaaS model of cloud provides services to the users as processing, storage, and virtualization, whereas PaaS offers a platform for deploying applications and the SaaS model grants users access to the applications in the cloud [6, 29, 30]. Cloud computing is made possible by the utilization of techniques, such as virtualization, MapReduce programming model, distributed file systems and so on. Distributed file system forms the key constituent in developing any cloud computing based applications [31]. The distributed file system are systems where the nodes perform the operation of the storage and computation simultaneously and the files are divided into multiple chunks of data, which are placed in multiple nodes. As the files in cloud can be created, modified or deleted randomly, addition, replacement or upgrading, of node is possible. Further, the load of a node depends on the quantity of chunks possessed by the node [7].

In cloud computing, allocating virtual machines with specific tasks is called as load and the performance of the cloud can be enhanced by distributing the loads uniformly among the various nodes in order to avoid situation where some nodes remain idle and the other nodes have extra load. The process of allocating the loads uniformly among all the nodes in the cloud is performed by means of load balancing, which helps in maintaining optimal resource allocation and avoidance of the bottle neck issue resulting in load imbalance [8]. There are two kinds of load balancing algorithms, such as static and dynamic. Static algorithms are applicable only for homogeneous system present in stable environments and has less overhead, whereas the dynamic systems can be applied for heterogeneous as well as homogeneous system and is highly adaptable [9]. Load balancing algorithms distributes the produced workload over various virtual machines. Multiple virtual machines may reside on a single physical machine, and to manage the cloud resources, virtual machine migration is performed. Migration of virtual machine can be performed from one physical machine to another effortlessly even when it is running. By performing virtual machine migration, load balancing can be achieved as virtual machines on an over-loaded server can be transferred to one with less load. Thus, virtual machine migration effectively reduces the power consumption and operational expenses. Moreover, virtual machine migration is highly effective in cases when any server has to be upgraded or maintained or in cases when any fault is detected prior to its occurrence. The process of migration can be enhanced by considering optimization for reduction of the downtime, total migration time and total data transferred [10].

Load balancing is the most crucial task in cloud computing that enables distribution of workload among the virtual machines. Many works have focused on developing solutions to address the various problems that occur while distributing the load among the multiple loads in the cloud system. The major issues encountered by the dynamic load balancing in heterogeneous cloud are listed below:

•
Adaptive starvation threshold method was unsuccessful in maintaining a balance among cost minimization and parallel task execution, which remains the major challenge [11].
•
Most of the existing methods of load balancing consider a single factor, such as cost, energy, trust, etc., while distributing tasks among the virtual machine. The focusing on one objective alone tends to degrade the effectiveness of the technique.
•
Another main challenge is that some methods did not considering the training of the tasks from the large data gathered by the cloud.

These conventional techniques drawbacks are inspired in the creation of a novel dynamic load balancing scheme. In this paper, an efficient Fractional Social Deer Optimization (FSDO) algorithm is proposed for performing dynamic load balancing in a distributed file system on heterogeneous cloud. The major contribution of this research is as follows:

A novel FSDO optimization algorithm is developed for performing load balancing and replica management. The FSDO algorithm is developed by incorporating fractional social optimization algorithm and deer hunting optimization. Here, fractional social optimization algorithm is an amalgamation of fractional calculus and social optimization algorithm.

The remaining part of the paper is organized as follows; the existing research works related to load balancing in distributed file system for cloud is detailed in Section 2, Section 3 depicts the system model of cloud, the devised FSDO is elaborated in Section 4, the experimental outcomes of the proposed technique and the comparative evaluation is discussed in Section 5 and finally, the paper is concluded in Section 6 along with future scope.
2. Literature review

Multiple works have concentrated on addressing the various issue encountered in distributing the load among the virtual machines and in developing efficient load balancing schemes. Here, eight of the prevailing load balancing techniques is considered and they are shortly discussed.

Semmoud et al. [11] devised a distributed load balancing technique called adaptive Starvation Threshold based Load Balancing (STLB) for balancing load in a cloud environment. The method checked the load allocated to the virtual machines and if any virtual machine is nearing starvation, load balancing was performed thus effectively reducing the quantity of virtual machine migration required. The technique was highly effective in maintaining stability by reducing the total idle time of the virtual machines. However, it was unsuccessful in reducing the communication cost. A low cost method was proposed by Souravlas et al. in [12] where a fair task distribution scheme was introduced for balancing load in a heterogeneous cloud system. The technique employed load balancer for allocating the tasks to the virtual machines considering their capabilities and present state. The introduced scheme effectively minimized the imbalance among virtual machines, but the completion time consumed was high.

Rajagopal et al. [13] devised an improved efficient scheme to organize the resources for performing dynamic load balancing. The technique allocated tasks considering the makespan time of the nodes between tasks. The improved efficient scheme successfully achieved enhanced workload balancing, although the technique did not consider enhancing the makespan. Nabi and Ahmed [14] proposed a technique, named overall performance-based resource aware dynamic load balancer for distributing the workload based on constraints. The workload was mapped independently and the tasks were computed in balanced way based on the availability of resources. The technique effectively enhanced the overall performance of the cloud by supporting deadline of the tasks; however, it was unsuccessful in considering the latency of the network.

Wu et al. [15] presented a Metadata Dynamic Load Balancing (MDLB) method for balancing load with Reinforcement Learning for load balancing. The technique utilized a Reinforcement Learning and Q learning algorithm for distributing the load and was implemented using three networks, such as the load balancing, policy selection along with the parameter update networks. Though, this model was capable of achieving better adaptability even when a rapid increase in data occurs, it failed to minimize the computational complexity. Kaur et al. [16] developed a deep learning-based Deadline-constrained, Dynamic Virtual Machine Provisioning and Load Balancing (DLD-PLB) for scheduling the workflow and optimizing the process of load balancing. The technique considered the genome workflow task and the deep learning was employed in the scheduling of the virtual machine optimally. The cost of the load balancing algorithm was low, although the time consumption was high.

Kumar et al. [17] proposed a Virtual Cluster Management System (VCMS) for balancing load in cloud. VCMS technique groups multiple virtual machines in a cluster to act as a single system that efficiently handles large data. Applications were managed and run efficiently by the VCMS model. The method effectively reduces the execution time, but the scheme does not guarantee live migration of the virtual machine. The capacity based deadline aware dynamic load balancing scheme was presented by Haidri et al. [18] to balance load in a cloud. Here, virtual machines were selected based on the running cost and deadline to satisfy the customers during task allocation. This method can be effectively scaled for performing load balancing among virtual machines with various speeds and cost. However, the technique does not consider migration of partially executed works.

In the devised scheme, the deep embedded clustering is used for chunk creation, which increases the scalability and offers superior performance. Also, the load prediction is done using the RideNN with better accuracy and less processing time. Moreover, the constrained problems are solved by using the devised FSDO algorithm. Besides, in distributed file system, replicas of the chunks for every file have to be maintained for specific virtual machine to enhance the availability of the data in case of failures and during virtual machine migration. Thus, the devised model is improved than the conventional models.

3. System model

The increase in digitization has resulted in gigantic growth in the data generated daily and with this growing data produces a requisite for huge storage and this can be established by using cloud. Cloud computing enables user to share collective resources, such as servers, storage and computing resources irrespective of location or time. Both hardware and software resources can be made available to the user with the process of virtualization. The flexibility and management of resources can be efficiently enhanced by virtualization of the cloud platform. Virtualization of the cloud is performed by taking into account virtual machines. The Virtual Machines (VMs) are hosted by the Physical Machine (PM). The system model of the cloud is displayed using Fig. 1.

Figure 1.

System model of cloud.

The data stored in the cloud should be always available and accounted for as failure in components in one location can cause loss of data. This is avoided by performing replicating data and storing them at multiple locations. Replica management performs the process of creating instances of data to ensure reliability and also to reduce the delay of access.

Figure 2.

Schematic view of the proposed method.

4. Proposed virtual machine migration for dynamic load balancing and replica management

Load balancing refers to the allocation of workloads among different virtual machines in a cloud so as to ensure that the work is equally distributed and no single virtual machine is inundated with more tasks. This paper is proposed with the aim of the developing an efficient dynamic load balance scheme in heterogeneous cloud using virtual machine migration and replica management. Figure 2 displays the schematic view of the Fractional Social Deer Optimization (FSDO) based dynamic load balancing method. Initially, the file is partitioned into multiple chunks by using the deep embedded clustering, which are then allocated to the virtual machines based on a round-robin manner. After chunk creation, load prediction is performed, where the load in physical machine is predicted using the RideNN. Load balancing is then executed wherein virtual machine migration is carried out based on the load condition in the physical machine using the devised FSDO algorithm. Various fitness parameters, such as resource utilization, migration cost, trust and energy consumption are considered while balancing the load. Later, replica management is accomplished for managing the replica in distributed file system with the help of the devised FSDO algorithm based on the objectives, like predicted load, get cost, storage cost and put cost. Moreover, the proposed FSDO is created by combining the fractional social optimization algorithm and the deer hunting optimization algorithms. The entire process is elaborated in the ensuing sections.

The algorithmic steps of the virtual machine migration and replica management are displayed using Algorithm 1.

Algorithm 1: Pseudocodes for virtual machine migration and replica management
1. Configure heterogeneous cloud model
2. Create chunks using deep embedded clustering
3. Allocate chunks to virtual machine in a round robin manner
4. Perform initialization of the virtual machine utilization factor
5. Initialize iteration
6. Calculate load $L_{i}$ and forecast the load condition using RideNN
7. Compute virtual machine moving factor ( $F_{vm}$ )
8. If $F_{vm}>$ 0.9
Execute virtual machine migration using the devised Fractional Social Deer Optimization (FDSO) algorithm
9. Update virtual machine utilization factor
10. While $F_{vm}>$ 0.9 at last $t$ iteration
Execute replication with the devised FDSO
11. Reiterate Steps 5 to 10

4.1 Chunk creation

Chunk creation is the process of partitioning a file into number of chunks, so as to enable execution of tasks simultaneously in multiple nodes. The number of chunks allocated to a node determines the load of a node. In cloud computing, the files can be randomly produced, updated or deleted; similarly the nodes can also be added, replaced or upgraded in the system. Hence, uniform distribution of chunks among the various nodes is not possible, here, deep embedded clustering [19] is utilized in the process of chunk creation and the chunks are allocated to the virtual machines in a round robin manner.

Deep embedded clustering is an unsupervised clustering technique that offers the advantage of high scalability and superior performance. Moreover, deep embedded clustering is capable of utilizing deep neural networks for learning cluster assignments as well as feature representations. The input file $F$ is subjected to the deep embedded clustering, which performs partitioning the file by transforming the data by using a non-linear mapping. The clustering objective is optimized by considering the mapping the data to a lower-dimensional feature space. The partitioning is performed in two steps, such as initialization of parameters and optimization. In the first step, a deep autoencoder is utilized and in the second step, clustering is carried out, where the process is kept repeated among the reduction of Kullback-Leibler divergence and auxiliary target distribution calculation. Consider a set of $z$ points, $\{{a_{m}\in A}\}_{m=1}^{z}$ be partitioned into $k$ clusters. deep embedded clustering transforms the data into a latent feature space $B$ using a nonlinear mapping $f_{\alpha}:A\to B$ . The clustering can be initiated by considering the estimate of learnable parameter $\alpha$ and $\{{\beta_{l}}\}_{l=1}^{k}$ , where $\beta$ denotes the centroid of the cluster.

Unsupervised clustering is performed by altering among two steps, such as soft assignment and cluster refinement using the original estimate of the $\{{\beta_{l}}\}_{l=1}^{k}$ and non-linear mapping $f_{\alpha}$ . Initially, soft assignment is computed among the cluster centroids and the embedded points. In the next step, the function $f_{\alpha}$ is updated and then an auxiliary target distribution is employed in refining the cluster centroids. The steps are kept altered till the convergence condition is satisfied. The process is briefed in the following steps.

(i) Soft assignment

The similarity exists between the centroid $\beta_{l}$ as well as the embedded point $a_{m}$ is computed using the below equation.

$\displaystyle b_{ml}=\frac{({{1+\|{a_{m}-\beta_{l}}\|^{2}}/\chi})^{-\frac{\chi% +1}{2}}}{\sum{({{1+\|{a_{m}-\beta_{l^{\prime}}}\|^{2}}/\chi})^{{-\frac{\chi+1}% {2}}}}}$ (1)

where $\chi$ is the degree of freedom and after embedding $a_{m}=f_{\alpha}({b_{m}})\in A$ corresponds to $b_{m}\in B$ .

(ii) Kullback-Leibler divergence optimization

The clusters are refined iteratively by utilization of the auxiliary target function by learning high confidence assignments. The target is matched to the soft assignment for training the model and the Kullback-Leibler divergence loss is given by,

$\displaystyle L=KL({Y\|X})=\sum\limits_{m}{\sum\limits_{l}{y_{ml}}}\log\frac{y% _{ml}}{x_{ml}}$ (2)

The above equation gives the loss between the auxiliary distribution, $y_{l}$ and the soft assignment $x_{m}$ and $x_{ml}$ is the probability of allocating a sample $m$ to cluster $l$ .

The value of $y_{l}$ is computed by finding the square of $x_{m}$ and later performing normalization of the value obtained using the frequency per cluster and is given by,

$\displaystyle y_{ml}=\frac{{x_{ml}^{2}}\mathord{\left/{\vphantom{{x_{ml}^{2}}{% f_{m}}}}\right.\kern-1.2pt}{f_{m}}}{\sum\limits_{l^{\prime}}{{x_{ml^{\prime}}^% {2}}\mathord{\left/{\vphantom{{x_{ml^{\prime}}^{2}}{f_{m^{\prime}}}}}\right.% \kern-1.2pt}{f_{m^{\prime}}}}}$ (3)

where $f_{m}=\sum\nolimits_{m}{x_{ml}}$ denote the frequencies of the soft cluster. Thus, deep embedded clustering enhances the initial estimate and effectively improves the low confidence predictions.

The chunks produced by the deep embedded clustering are denoted as $Z$ and are allocated to the virtual machine in a round robin manner, which is then subjected to the load prediction phase.

4.2 Load prediction using rider optimization algorithm based neural network

After the created chunk $Z$ is allocated to the virtual machines and based on the load $L_{i}$ in the virtual machine, load prediction is performed at the physical machines using the RideNN [20]. The RideNN was developed by integrating the rider optimization algorithm with neural network. RideNN is highly efficient in providing better classification accuracy and also the processing time required is very low. RideNN imitates the characteristic of human brain for determining the relationship among large data. RideNN consists of three layers, such as input layer, hidden layer and the output layer. Structure of the RideNN is shown using Fig. 3.

Figure 3.

Structure of Rider optimization algorithm based Neural Network (RideNN).

The outputs of one neuron are connected to the input of other neuron in various layers for creating the RideNN network. The loads in the virtual machine $L_{i}$ are applied as input to the RideNN. Here, the hidden layers comprise neurons that are associated to output layer considering weights. The output layer is integration of output neurons that reveals final network and all these layers are interconnected to establish a network. The input applied to the RideNN can be represented by,

$\displaystyle L_{i}=\{{L_{i1}^{c},L_{i2}^{c},..,L_{i\delta}^{c},..L_{ir}^{c}}% \},1\leqslant\delta\leqslant r$ (4)

where $Z_{\delta}^{c}$ represents the input applied to the $\delta^{\text{th}}$ neuron and $r$ signifies the total number of input neurons. The weights assigned to the hidden layer neuron is represented by,

$\displaystyle W=\{{W_{1},W_{2},\ldots,W_{r}}\}$ (5)

The neurons in the hidden layer are considered to have a bias $W_{r+1}$ . Similarly, the output layer neurons have a weight $W_{r+2}$ and bias $W_{r+3}$ . The output of the RideNN classifier is obtained by considering the transfer function given by,

$\displaystyle O^{c}=W_{r+2}\ast\left[{\textit{logsig}\left({\sum\limits_{% \delta=1}^{r}{Z_{\delta}^{c}\ast W_{\delta}+W_{r+1}}}\right)}\right]+W_{r+3}$ (6)

where $W_{\delta}$ is the weight of the $\delta^{\text{th}}$ hidden neuron and logsig represent the log-sigmoid transfer function, which calculates the output from the input applied. Thus, the predicted load thus calculated using the RideNN is represented as $O^{c}$ .

4.3 Computation of virtual machine moving factor

Virtual machine migration is performed based on the load condition of the virtual machine, the predicted load $O^{c}$ is utilized for computing the virtual machine moving factor $F_{vm}$ [21]. A number of physical machines and virtual machines are configured together to form a cloud, and in order to compute the virtual machine moving factor, for each physical machine, utilization matrix has to be initialized. After initialization, the virtual machine moving factor is computed based on the resource utilization of the virtual machine for monitoring the load condition of the physical machine. The virtual machine moving factor of the physical machine is given by,

$\displaystyle F_{vm}=\left[{\frac{1}{u}\sum\limits_{i=1}^{u}{\frac{1}{N}\left(% {\sum\limits_{j=1}^{N}{U_{j}}}\right)}}\right]$ (7)

where $u$ represent the number of rounds, $N$ denote the number of virtual machines in the physical machine, and $U_{j}$ denote the resource utilization of the $j^{\text{th}}$ virtual machine. Based on the value of $F_{vm}$ , migration is performed as $F_{vm}>$ 0.9 implies that the physical machine is overloaded.

4.4 Virtual machine migration using the proposed Fractional Social Deer Optimization (FSDO) algorithm for load balancing

This section details the proposed Fractional Social Deer Optimization Algorithm (FSDO) algorithm for performing virtual machine migration to balance the load, wherein virtual machine migration is performed based on the virtual machine moving factor $F_{vm}$ . Various objectives, such as migration cost, predicted load, energy consumption and trust are considered before performing virtual machine migration. A novel FSDO algorithm is proposed in this paper for performing virtual machine migration, The devised FSDO algorithm is developed by modifying the fractional social optimization algorithm in accordance with the deer hunting optimization [22] algorithm. The fractional social optimization algorithm has been developed by combining the fractional calculus [23] and social optimization algorithm [24] algorithms. The process of virtual machine migration is detailed in the upcoming sections.

4.4.1 Solution encoding

This section elaborates the solution encoding process of the devised FSDO algorithm for performing virtual machine migration. A number of virtual machines are randomly chosen to form the solution vector. The process of solution encoding is depicted in Fig. 4. In this figure, five virtual machines are selected to form the solution vector. The position of the virtual machines in the solution vector keeps altering as the algorithm progresses depending on the solution generated by the FSDO algorithm. Various fitness parameters, such as migration cost, predicted load, energy consumption and trust are considered while calculating the solutions and the final solution obtained by the FSDO will be the optimal solution.

Figure 4.

Solution encoding of the Fractional Social Deer Optimization (FSDO).

4.4.2 Fitness function

Fitness function developed for the devised FSDO algorithm is detailed in this section. The optimal solution is computed based on the value of fitness function and the solution with the lowest value of fitness is considered as the best solution and the value of fitness is computed using the following expression,

$\displaystyle\textit{Fit}_{1}=\omega(1-D_{k})+\psi(O_{c})+\sigma(E)+\lambda(1-% R_{ij})$ (8)

where $\omega+\psi+\sigma+\lambda=1$ , and $D_{k}$ denotes the cost of migration, $O_{c}$ refers to the predicted load, $E$ represents the energy consumption, and $R_{ij}$ specifies the trust between the demand and the target node to provide security during virtual machine migration. The various fitness objectives are briefed below.

(i) Migration cost

Migration cost is calculated by taking into account the memory, bandwidth and Central Processing Unit (CPU) utilized to the overall memory, bandwidth and CPU available and it gives the cost used by the virtual machine while migrating. The migration cost is given by,

$\displaystyle D_{k}=\frac{1}{n\times o}\left[{\sum\limits_{i=1}^{n}{\sum% \limits_{j=1}^{o}{\frac{1}{3}\left({\frac{Q_{j}^{\textit{CPU}}}{P_{ij}^{% \textit{CPU}}}+\frac{Q_{j}^{\textit{mem}}}{P_{ij}^{\textit{mem}}}+\frac{Q_{j}^% {BW}}{P_{ij}^{BW}}}\right)}}}\right]$ (9)

where $n$ and $o$ denotes the count of the physical machine and virtual machines. $Q_{j}^{\textit{CPU}}$ , $Q_{j}^{\textit{mem}}$ and $Q_{j}^{BW}$ represent the CPU, memory and bandwidth used and $P_{ij}^{\textit{CPU}}$ , $P_{ij}^{\textit{mem}}$ and $P_{ij}^{BW}$ denote the CPU, memory and bandwidth available.

(ii) Predicted load

The load prediction is performed using RideNN and the process is detailed in section 4.2.

(iii) Energy consumption

The energy consumed during virtual machine migration should be always kept low as any increase with impact the performance of the cloud and the following equation is utilized for computing the consumption of energy.

$\displaystyle E=\frac{1}{n\times o}\left[{\sum\limits_{i=1}^{n}{\sum\limits_{j% =1}^{o}{c_{ij}.E_{\max}+\left({1-c_{ij}}\right)\cdot\eta_{ij}\cdot E_{\max}}}}\right]$ (10)

where $c_{ij}=$ 0.1 and $E_{\max}=$ 1. $\eta_{ij}$ is the CPU utilization rate and is given by,

$\displaystyle\eta_{ij}=\frac{1}{3}\left({\frac{Q_{j}^{\textit{CPU}}}{P_{ij}^{% \textit{CPU}}}+\frac{Q_{j}^{\textit{mem}}}{P_{ij}^{\textit{mem}}}+\frac{Q_{j}^% {BW}}{P_{ij}^{BW}}}\right)$ (11)

(iv) Trust

Trust is calculated between the demand and the target nodes, where demand node or source node represents the physical machine on which the virtual machine that has to be migrated is present and the target node is the physical machine to which the virtual machine has to be migrated [25]. The trust of a node is computed to ensure security to the data transmitted [26], and is computed using the following equation

$\displaystyle R_{ij}=h\times\textit{NewR}_{ij}+\lambda\times(1-h)\times\textit% {OldR}_{ij}$ (12)

where $h$ and $\lambda$ denote the parameters of the system, which can be adjusted as per the requirements and has value in the range [0, 1]. The term $i$ represent the demand node and the target node is represented by $j$ . $\textit{OldR}_{ij}$ designates the old value of trust and $\textit{NewR}_{ij}$ represents the new value of trust and is represented as,

$\displaystyle\textit{NewR}_{ij}=u_{ij}+\varepsilon\times v_{ij}$ (13)

where $\varepsilon$ designates the relative atomicity factor. $u_{ij}$ is the belief of node $i$ for forwarding to node $j$ and $v_{ij}$ is the number of uncertainty in reasonable forwarding behaviour. The belief can be represented as,

$\displaystyle u_{ij}=\frac{\kappa_{ij}}{\kappa_{ij}+\gamma_{ij}}\times({1-v_{% ij}})$ (14)

where $\kappa$ and $\gamma$ indicates the count of successful and unsuccessful interactions. The value of belief should be computed such that $u_{ij}+v_{ij}+d_{ij}=1$ , where $d_{ij}$ is the disbelief, and is expressed as,

$\displaystyle d_{ij}=\frac{\gamma_{ij}}{\kappa_{ij}+\gamma_{ij}}\times({1-v_{% ij}})$ (15)

The uncertainty in forwarding $v_{ij}$ is given by,

$\displaystyle v_{ij}=\frac{12\times\kappa_{ij}\times\gamma_{ij}}{({\kappa_{ij}% +\gamma_{ij}})^{2}\times({1+\kappa_{ij}+\gamma_{ij}})}$ (16)

4.4.3 Proposed Fractional Social Deer Optimization (FDSO) algorithm

The proposed FSDO algorithm is employed in the process of virtual machine migration where the FSDO algorithm is developed by combining the fractional social optimization algorithm with the deer hunting optimization [22]. The fractional social optimization algorithm was developed by combining the fractional calculus [23] and the social optimization algorithm [24]. The deer hunting optimization is based on the hunting strategy of human with respect to deer, which mainly focus on the motion of two hunters, namely the leader as well as successor. Although the hunters differ in their activity, the hunting way depends on the technique they create. The deer hunting optimization aims to find the best position for the hunters while hunting the deer and offers high classification accuracy and is highly effective in solving engineering problems. The fractional calculus is highly efficient in enhancing the convergence rate of the algorithm and in improving the performance of the optimization. Fractional calculus is highly efficient in solving derivative and integral equations by the utilization of Laplace transforms. The social optimization algorithm is an algorithm, which is motivated by the societal behavior of humans. The algorithm meticulously follows two principles, such as equality of opportunity along with community for finding the optimal solution. The social optimization algorithm can be effectively handled the unconstrained as well as constrained function and is highly effective in reducing the cost function. By combining all the three methods, a highly efficient optimization algorithm is developed, which can be applied to solve constrained problems. The algorithmic steps of the devised FSDO algorithm are as follows.

Step 1) Initialization: The position of the individuals in the society $M$ is first initialized and is given by,

$\displaystyle M=\{M_{1},M_{2},\ldots,M_{e},\ldots,M_{s}\}$ (17)

where $M_{e}$ denotes the societal position of the $e^{\text{th}}$ individual, hence $1\leqslant e\leqslant s$ .

Step 2) Fitness evaluation: The fitness of the solution is computed using Eq. (8) to find the optimal solution, where the optimal solution is the one with minimum value of fitness.

Step 3) Determine equality of opportunity: The principle of equality of opportunity is based on the fact that the individuals are entitled to get any position in the society irrespective of the social backgrounds and the opportunity is provided only depending on the personal choice and talents. The position of the individual is obtained by using fractional social optimization algorithm as,

$\displaystyle M_{w}^{\textit{new}}=\delta M_{w}^{\textit{old}}+\frac{1}{2}% \delta M_{w}^{t-1}+\frac{1}{6}({1-\delta})M_{w}^{t-2}+\frac{1}{24}\delta({1-% \delta})({2-\delta})M_{w}^{t-3}\!\!{}+rd({bp-pc\times M_{w}^{\textit{old}}})$ (18)

where $r d$ is an arbitrary number with value in the range [0, 1], $\delta$ represents the fractional coefficient. $p c$ denotes the personal choice with value from the set $\{{0,1,2}\}$ and $b p$ indicates the best position. $M_{w}^{\textit{new}}$ , and $M_{w}^{\textit{old}}$ represents the new and old position of the $w^{\text{th}}$ individual in the society. The value of best choice is computed as,

$\displaystyle bp=\textit{rand}\{{b_{s},d_{p}}\}$ (19)

where $b_{s}$ signifies the best solution, and $d_{p}$ is the density point, which is given by

$\displaystyle d_{p}=\frac{M_{1}V_{1}+M_{2}V_{2}+\ldots+M_{s}V_{s}}{V_{1}+V_{2}% +\ldots+V_{s}}$ (20)

where $V_{i}$ indicates the result obtained by the $i^{\text{th}}$ individual in the society.

Equation (18) can be rewritten as,

$\displaystyle M_{w}^{\textit{new}}=M_{w}^{\textit{old}}({\delta-rd\times pc})+% \frac{1}{2}\delta M_{w}^{t-1}+\frac{1}{6}({1-\delta})M_{w}^{t-2}+\frac{1}{24}% \delta({1-\delta})({2-\delta})M_{w}^{t-3}\!\!{}+rd\times bp$ (21)

From the deer hunting optimization, the encircling behavior is given by,

$\displaystyle M_{w}^{\textit{new}}=M^{\textit{lead}}-T\cdot rd_{1}\cdot|{W% \times M^{\textit{lead}}-M_{{}_{w}}}|$ (22)

where $M^{\textit{lead}}$ represents the best position of the hunter, $rd_{1}$ is an arbitrary number having values in the range [0, 2] $T$ and $W$ indicates the coefficient vectors and are expressed by the following equation

$\displaystyle T=\frac{1}{4}\log\left({w+\frac{1}{w_{\max}}}\right)\vartheta$ (23) $\displaystyle W=2.rd_{2}$ (24)

where $\vartheta$ denotes a parameter with value in the range $[-1,1]$ , $w_{\max}$ signifies the maximum iteration and $rd_{2}$ is an arbitrary number with value between 0 and 1.

Now assuming, $M^{\textit{lead}}>M_{w}$ in Eq. (22), the equation can be rewritten as,

$\displaystyle M_{w}^{\textit{new}}=M^{\textit{lead}}-T\cdot rd_{1}\cdot({W% \times M^{\textit{lead}}-M_{{}_{w}}})$ (25) $\displaystyle M_{w}=\frac{M_{w}^{\textit{new}}-M^{\textit{lead}}(1-T\cdot rd_{% 1}\cdot W)}{T\cdot rd_{1}}$ (26)

Substituting Eq. (26) in Eq. (21), we get

$\displaystyle M_{w}^{\textit{new}}=\left({\frac{M_{w}^{\textit{new}}-M^{% \textit{lead}}(1-T\cdot rd_{1}\cdot W)}{T\cdot rd_{1}}}\right)({\delta-rd% \times pc})+\frac{1}{2}\delta M_{w}^{t-1}+\frac{1}{6}({1-\delta})M_{w}^{t-2}{}% +\frac{1}{24}\delta({1-\delta})({2-\delta})M_{w}^{t-3}+rd\times bp$ (27) $\displaystyle M_{w}^{\textit{new}}=\left({\frac{T\cdot rd_{1}}{T\cdot rd_{1}-% \delta+rd\times pc}}\right)\left({\begin{array}[]{l}\frac{1}{2}\delta M_{w}^{t% -1}+\frac{1}{6}({1-\delta})M_{w}^{t-2}+\frac{1}{24}\delta({1-\delta})({2-% \delta})\\ M_{w}^{t-3}+rd\times bp-\frac{M^{\textit{lead}}(1-T\cdot rd_{1}\cdot W)}{T% \cdot rd_{1}}({\delta-rd\times pc})\\ \end{array}}\right)$ (28)

where $M_{w}^{t-1}$ , $M_{w}^{t-2}$ and $M_{w}^{t-3}$ denote the societal position of the person in the $(t-1)^{\text{th}}$ , $(t-2)^{\text{th}}$ and $(t-3)^{\text{th}}$ iterations respectively.

Step 4) Determination of principle of community: Principle of community is based on the idea that two persons of the same capability occupy varying positions in the society which may be due to personal choice. Such conditions if takes place at large scale, an unjust society is created and to avoid this, the persons may have to disregard the profits and give up their choices for the overall advantage of the society. This can be expressed using fractional social optimization algorithm as,

where $b_{s}$ is the best solution and Ept denotes the empty point, computed by,

$\displaystyle\textit{Ept}=\frac{M_{1}\frac{1}{V_{1}}+M_{2}\frac{1}{V_{2}}+% \ldots+M_{s}\frac{1}{V_{s}}}{\frac{1}{V_{1}}+\frac{1}{V_{2}}+\ldots+\frac{1}{V% _{s}}}$ (30)

Substituting Eq. (26) in Eq. (29), we get,

$\displaystyle M_{w}^{\textit{new}}=\delta\left({\frac{M_{w}^{\textit{new}}-M^{% \textit{lead}}(1-T\cdot rd_{1}\cdot W)}{T\cdot rd_{1}}}\right)+\frac{1}{2}% \delta M_{w}^{t-1}+\frac{1}{6}({1-\delta})M_{w}^{t-2}\!\!{}+\frac{1}{24}\delta% ({1-\delta})({2-\delta})M_{w}^{t-3}+rd({b_{s}-\textit{Ept}})$ (31) $\displaystyle M_{w}^{\textit{new}}=\frac{T\cdot rd_{1}}{({T\cdot rd_{1}-\delta% })}\left({\begin{array}[]{l}\frac{1}{2}\delta M_{w}^{t-1}+\frac{1}{6}({1-% \delta})M_{w}^{t-2}+\frac{1}{24}\delta({1-\delta})({2-\delta})M_{w}^{t-3}\\ {}+rd({b_{s}-\textit{Ept}})-\left({\frac{M^{\textit{lead}}(1-T\cdot rd_{1}% \cdot W)}{T\cdot rd_{1}}}\right)\\ \end{array}}\right)$ (32)

The above equation gives the position of the individual based on the equality of community principle.

Step 5) Determination of density point and empty point: Initially the density and empty points are computed based on the assumption that all $V_{i}$ are positive, although it is required to compute the density and empty points for negative values too. The following equations are utilized in computing the density and empty points.

The density point is expressed as,

$\displaystyle d_{p}=\sum\limits_{i=1}^{J}{\frac{V_{i}}{\sum\limits_{j=1}^{J}{V% _{j}}}M_{i}}$ (33)

where $M_{i}$ indicates the position of the individual in the $i^{\text{th}}$ iteration and the term $\frac{V_{i}}{\sum\limits_{j=1}^{J}{V_{j}}}$ is referred to as the relative fitness $\hat{V}_{i}$ and so the above equation can be rewritten as,

$\displaystyle d_{p}=\sum\limits_{i=1}^{J}\hat{V}_{i}M_{i}$ (34)

Empty point can be computed by,

$\displaystyle\textit{Ept}=\sum\limits_{i=1}^{J}{\frac{\frac{1}{V_{i}}}{\sum% \limits_{j=1}^{J}{\frac{1}{V_{j}}}}M_{i}}$ (35) $\displaystyle\textit{Ept}=\sum\limits_{i=1}^{J}{\mathord{\buildrel\lower 3.0pt% \hbox{$\scriptscriptstyle\smile$}\over{V}}_{i}M_{i}}$ (36)

where $\mathord{\buildrel\lower 3.0pt\hbox{$\scriptscriptstyle\smile$}\over{V}}_{i}=% \frac{\frac{1}{V_{i}}}{\sum\limits_{j=1}^{J}{\frac{1}{V_{j}}}}$ .

Step 6) Fitness re-computation: The best solution is obtained by considering the solution with the lowest fitness value, and for this the fitness of the solution obtained is re-calculated. If the fitness of the obtained solution is better than the previous one, then the new solution computed is replaced by the previous one.

Step 7) Terminate: The above steps are reiterated till the maximum iterations are attained for computing the best solution. Algorithm 2 displays the pseudocode of FSDO algorithm.

Algorithm 2: Pseudocodes of developed Fractional Social Deer Optimization (FSDO) algorithm
1. Input: Population size $s$ , maximum iteration $w_{\max}$
2. Output: Best solution $M^{\ast}$
3. Begin
4. Arbitrarily initialize the population $M$
5. Assess the population
6. $b_{s}$ $\leftarrow$ Best solution
7. For $j=$ 1 to $w_{\max}$ do
8. Determine empty point Ept
9. For $j=$ 1 to $J$ do
10. $bp=\textit{rand}\{{b_{s},\textit{Ept}}\}$ ;
11. $pc=\textit{rand}\{{0,1,2}\}$ ;
12. Compute $M_{w}^{\textit{new}}$ using Eq. (31)
13. End
14. If $M_{w}^{\textit{new}}$ better than $M_{w}^{\textit{old}}$ then
15. $M_{w}\leftarrow M_{w}^{\textit{new}}$
16. End
17. Evaluate new population
18. $b_{s}$ $\leftarrow$ Best solution;
19. Compute Density point $d_{p}$
20. For $w=$ 1 to $J$ do
21. Evaluate $M_{w}^{\textit{new}}$ using Eq. (38)
22. End
23. If $M_{w}^{\textit{new}}$ better than $M_{w}^{\textit{old}}$ then
24. $M_{w}\leftarrow M_{w}^{\textit{new}}$
25. End
26. Evaluate new population
27. $b_{s}$ $\leftarrow$ Best solution;
28. End

The best solution obtained is represented by $b_{s}$ .

4.5 Replication Management using the Fractional Social Deer Optimization (FSDO) algorithm

In a distributed file system, replicas of the chunks for every file have to be maintained for specific virtual machine to enhance the availability of the data in case of failures and during virtual machine migration. Replica management is a way of ensuring availability of data. The existing technique of load balancing does not take into account the replica. The random way in which load balancing is performed requires that two or more replica be maintained in virtual machines. In this paper, replica management is carried out using the FSDO algorithm.

4.5.1 Solution encoding

Solution encoding is important to determine the best solution in solving optimization problems. Here, the best solution for replica management is obtained by employing the FSDO algorithm. The solution vector is created randomly and consists of a number of virtual machines to which the resources are allotted. The fitness of the virtual machine is computed and based on the fitness, the optimal count of virtual machines are determined and the solution vector is updated. Figure 5 portrays the solution encoding of the proposed FSDO algorithm for replica management.

Figure 5.

Solution encoding for replica management using proposed Fractional Social Deer Optimization algorithm.

4.5.2 Fitness function

In this section, the fitness function used in replica management using the devised FSDO is detailed. The fitness of the virtual machine is computed by considering objectives, such as predicted load, put cost, get cost, and storage cost. The optimal solution is obtained by finding the solution, which has the maximum value of fitness. The fitness function used in replica management is given by,

$\displaystyle\textit{Fit}_{2}=\frac{1}{\eta}\sum\limits_{c=1}^{\eta}{({(1-L_{i% })+Pt_{i}+Gt_{i}+Sg_{i}})}$ (37)

where $L_{i}$ is predicted load of virtual machine, $Gt_{i}$ indicate the get cost, $Pt_{i}$ denotes the put cost, and $Sg_{i}$ represent the storage cost of $i^{\text{th}}$ virtual machine.

The load in the $i^{\text{th}}$ virtual machine can be found by using the following equation

$\displaystyle L_{i}=({1-U_{i}})$ (38)

where $U_{i}$ is the resource utilization of the virtual machine.

The get cost is given by,

$\displaystyle Gt_{i}=\frac{\left({\sum\limits_{k=1}^{l}{\rho_{k}\ast tr_{g}+}% \sum\limits_{x=1}^{l}{\rho_{k}*q}}\right)}{K}$ (39)

where $\rho_{k}$ is count of get request in $k^{\text{th}}$ region, $q$ is object size and $tr_{g}$ specifies the transaction price of get requests, $l$ indicate the number of regions and $K$ is the normalization factor.

The put cost can be computed by using the following equation

$\displaystyle Pt_{i}=\frac{\left({\sum\limits_{k=1}^{l}{\sigma_{x}-tr_{p}+\sum% \limits_{k=1}^{l}{\sigma_{k}*q}}}\right)}{K}$ (40)

where $\sigma_{k}$ represents the count of put request from $k^{\text{th}}$ region, $T_{p}$ is transaction price of put requests.

The storage cost is given by,

$\displaystyle Sg_{i}=\textit{mem}\ast q$ (41)

where mem denotes the memory.

The FSDO algorithm is detailed in section 4.4.3 and the fitness function utilized for performing replica management is done using Eq. (37).

5. Results and discussion

This section details the experimental evaluation of the devised FSDO technique for performing dynamic load balancing in distributed file system. The performance of the proposed technique is evaluated by considering two cases. In set up 1, the number of physical machines is assumed to be 4 and the number of virtual machines as 15. On the other hand, in set up 2, 6 physical machines and 20 virtual machines are considered. The proposed FSDO is implemented using Java with Cloudsim tool on a PC with the specifications, such as 2 GB RAM, Intel i3 core processor and Windows 10 OS.

5.1 Evaluation metrics

The performance metrics used in the evaluation of the proposed FSDO based dynamic load balancing is briefed in this section. For evaluation, various parameters namely cost, load, trust and energy consumption are utilized.

5.2 Performance evaluation

Load balancing is highly significant in improving the energy efficiency and allocating resources optimally in the cloud. Accurate prediction of load can assist in taking efficient decisions for load balancing and job scheduling. The performance of the FSDO technique can be evaluated by considering the load prediction error, as a low value of error implies the accuracy of the technique. Figure 6 depicts the analysis of the devised FSDO algorithm, by considering various population sizes and by varying the number of rounds. In Fig. 6a, the FSDO method is evaluated with Setup #1. When the number of round is considered to be 10, the load prediction error computed for the population size 5, 10, 15 and 20 is 0.111, 0.079, 0.077 and 0.075. In Fig. 6b, the evaluation of the FSDO technique is displayed. The proposed FSDO technique computed load prediction error of 0.069, 0.069, 0.061 and 0.052 for the population size of 5, 10, 15 and 20, respectively for 20 rounds.

Figure 6.

Performance analysis of the proposed method.

Table 1

Nomenclature

Abbreviation	Definition
DLD-PLB	Deep Learning based Deadline-constrained, Dynamic Virtual Machine Provisioning and Load Balancing
FSDO	Fractional Social Deer Optimization
FSOA	Fractional Social Optimization Algorithm
MDLB	Metadata Dynamic Load Balancing
STLB	Starvation Threshold based Load Balancing
VCMS	Virtual Cluster Management System

Figure 7.

Evaluation using Setup #1.

Figure 8.

Evaluation using Setup #2.

5.3 Comparative methodologies

The devised technique of dynamic load balancing is analyzed for its performance by comparing it with the prevailing techniques, like DLD-PLB [16], VCMS [17], MDLB-based Reinforcement Learning [15], Adaptive starvation threshold [11], and FSOA. The proposed FSDO is evaluated using the several parameters, such as cost, trust and energy consumption by varying the number of rounds taking into consideration the two set ups.

The abbreviations and its definitions are provided in Table 1.

DLD-PLB [16]: In this method, the deep learning was used in the optimal scheduling and the workflow execution was done by the optimization algorithm.

VCMS [17]: In this scheme, the virtual cluster management system scheme was developed for managing and executing the applications and obtains accurate results.

MDLB-based Reinforcement Learning [15]: This approach contains parameter update network, load balancing network, and policy selection network modules for efficient dynamic load balancing.

Adaptive starvation threshold [11]: Here, the dynamic load balancing was performed by adaptive starvation threshold, which minimize the response time and maximize the utilization rate.

FSOA: It is the hybrid optimization algorithm, in which the fractional concept was applied in the Social Optimization Algorithm.

Table 2
Comparative evaluation

Metrics	Adaptive starvation threshold	Dynamic Load Balancing-based Reinforcement Learning
(MDLB-based RL)	Virtual Cluster Management System
(VCMS)	Dynamic Virtual Machine Provisioning
and Load
Balancing
(DLD-PLB)	Fractional Social Optimization Algorithm
(FSOA)	Proposed Fractional Social Deer Optimization (FSDO)
	Setup #1
Cost	0.445	0.444	0.444	0.417	0.406	0.390
Trust	0.447	0.530	0.541	0.554	0.575	0.723
Energy consumption	0.608	0.563	0.562	0.560	0.498	0.456
	Setup #2
Cost	0.516	0.515	0.508	0.506	0.456	0.432
Trust	0.468	0.491	0.532	0.540	0.595	0.634
Energy consumption	0.607	0.560	0.559	0.557	0.468	0.431

5.3.1 Evaluation using Setup #1

Figure 7 displays the evaluation of the technique with Setup #1 by varying the number of rounds. In Fig. 7a, assessment is illustrated with respect to cost. For 10 rounds, the proposed FSDO computed a cost of 0.412, which is less than the values 0.433, 0.465, 0.469, 0.478, and 0.479 computed by the existing methodologies, such as FSOA, DLD-PLB, VCMS, MDLB-based Reinforcement Learning, and Adaptive starvation threshold. The assessment considering the trust is displayed with the help of Fig. 7b. The existing approaches, like FSOA, DLD-PLB, VCMS, MDLB-based Reinforcement Learning, and Adaptive starvation threshold achieved a trust value of 0.615, 0.564, 0.551, 0.551 and 0.502. However, the FSDO attained a higher value of trust at 0.763, for 15 rounds. Figure 7c illustrates the analysis of the approaches based on energy consumption. For 5 rounds, the proposed FSDO approach consumed only an energy of 0.604J, which is less than the energy consumed by other techniques, namely FSOA, DLD-PLB, VCMS, MDLB-based Reinforcement Learning, and Adaptive starvation threshold of 0.610J, 0.643J, 0.652J, 0.654J and 0.654J correspondingly.

5.3.2 Evaluation using Setup #2

Figure 8 illustrates the assessment of the devised FSDO technique in comparison to the prevailing load balancing methods by considering different number of rounds with Setup #2. The evaluation using cost is displaced using Fig. 8a. When the number of round is considered to be 5, the cost achieved by the prevailing techniques, like FSOA, DLD-PLB, VCMS, MDLB-based Reinforcement Learning, and Adaptive starvation threshold is 0.519, 0.526, 0.526, 0.529, and 0.529 correspondingly. However, the devised FSDO technique attained a lower value of cost at 0.495. Figure 8b depicts the evaluation using trust. The devised FSDO method computed a high trust value of 0.881, and the conventional techniques, like FSOA, DLD-PLB, VCMS, MDLB-based Reinforcement Learning, and Adaptive starvation threshold calculated values of 0.795, 0.617, 0.547, 0.534, and0.490. In Fig. 8c, the evaluation with respect to energy consumption is depicted. For 15 rounds, the energy consumed by the prevailing FSOA, DLD-PLB,VCMS, MDLB-based Reinforcement Learning, and Adaptive starvation thresholdis 0.561J, 0.598J, 0.600J, 0.603J, and 0.608J and the FSDO method consumes only energy of 0.470J.

5.4 Comparative discussion

This section deals with the comparative assessment of the devised FSDO technique of dynamic load balancing. Table 2 displays the values obtained by the various techniques using the metrics, such as cost, trust and energy consumption with the multiple set ups considered. The values are computed corresponding to 20 number of round. It can be inferred from the table, that the FSDO method achieved a lower cost and energy consumption values at 0.390 and 0.431J, whereas attaining a high trust value of 0.723. The utilization of RideNN for predicting the load in the physical machine has contributed to the low cost value. The energy consumption is minimized by employing the deep embedded clustering for partitioning the chunks. Moreover, the utilization of trust factor in the load is enhanced by the optimization of the virtual machine migration using the FSDO algorithm.

6. Conclusion

Cloud computing is a power architecture that allows handling of huge data generated globally with the incremental growth of digitization. The cloud offers the benefit of utilizing shared computing resources and servers that can be utilized from any location irrespective of the time. The growing workload in cloud requires efficient techniques for balancing the load in virtual machine. In this paper a trust based dynamic load balancing algorithm in distributed file system is proposed. Initially, deep embedded clustering is utilized for partitioning the files into chunks and then RideNN is utilized in the prediction of load in the physical machine. A novel FSDO algorithm is developed for performing virtual machine migration and replica management. The FSDO algorithm is devised by incorporating the deer hunting optimization algorithm with the fractional calculus and social optimization algorithm. Moreover, the proposed FSDO based dynamic load balancing algorithm is evaluated for its performance based on parameters, like predicted load prediction error, trust, cost and energy consumption with values 0.051, 0.723, 0.390 and 0.431J correspondingly which reveals superior performance. In future works, other deep learning networks will be employed for load prediction. Moreover, in future other parameters will be consider for virtual machine migration, like makespan and average response time in addition to the ones used in the existing works.

Footnotes

Author’s Bios

Manjula H. Nebagiri received B.E. degree in computer science and engineering in 2001 from Karanataka University and M.Tech degree in Software engineering from Rashtreeya Vidyalaya College of Engineering (RVCE). She is currently pursuing Ph.D in Computer Science and Engineering (in Cloud Computing) from Visvesvaraya Technological University, Belagavi. She is working as an Assistant Professor in Atria Institute of Technology and has an experience of 14

+

years. She has taken the Sambhram Institute of Technology as Research center. Her main area of research interests are Cloud Computing, IoT, Machine Learning and Networks.

Latha P.H. received B.E. degree in Information Technology from B.M.S college of Engineering, Bangalore University in 1992 and M.Tech degree in Computer Network and Engineering from AMC and PhD from Visvesvaraya Technological University, Belagavi 2017. She is working as a Head of the Department in Sambhram Institute of Technology and has an experience of 21 years. Her main area of research interests are Embedded System, IoT, and Network Security.

References

Kumar

and Kumar

, Issues and challenges of load balancing techniques in cloud computing: A survey, ACM Computing Surveys (CSUR) 51(6) (2019), 1–35.

Mell

and Grance

, The NIST definition of cloud computing, NIST Special Publication, 2011.

Haidri

R.A.

Katti

C.P.

and Saxena

P.C.

, A load balancing strategy for cloud Computing environment, in: proceedings of 2014 International Conference on Signal Propagation and Computer Technology (ICSPCT 2014), IEEE, Ajmer, India, 2014, pp. 636–641.

Subramanian

and Jeyaraj

, Recent security challenges in cloud computing, Computers & Electrical Engineering 71 (2018), 28–42.

Rashid

and Chaturvedi

, Cloud computing characteristics and services: A brief review, International Journal of Computer Sciences and Engineering 7(2) (2019), 421–426.

Prajapati

A.G.

Sharma

S.J.

and Badgujar

V.S.

, All about cloud: A systematic survey, in: Proceedings of 2018 International Conference on Smart City and Emerging Technology (ICSCET), IEEE, Mumbai, India, 2018, pp. 1–6.

Hsiao

H.C.

Chung

H.Y.

Shen

and Chao

Y.C.

, Load rebalancing for distributed file systems in clouds, IEEE Transactions on Parallel and Distributed Systems 24(5) (2012), 951–962.

Kansal

N.J.

and Chana

, Cloud load balancing techniques: A step towards green computing, IJCSI International Journal of Computer Science Issues 9(1) (2012), 238–246.

Mishra

S.K.

Sahoo

and Parida

P.P.

, Load balancing in cloud computing: A big picture, Journal of King Saud University-Computer and Information Sciences 32(2) (2020), 149–158.

10.

Noshy

Ibrahim

and Ali

H.A.

, Optimization of live virtual machine migration in cloud computing: A survey and future directions, Journal of Network and Computer Applications 110 (2018), 1–10.

11.

Semmoud

Hakem

Benmammar

and Charr

J.C.

, Load balancing in cloud computing environments based on adaptive starvation threshold, Concurrency and Computation: Practice and Experience 32(11) (June 2020), 1–14.

12.

Souravlas

Anastasiadou

S.D.

Tantalaki

Katsavounis

and Fair

, Dynamic load balanced task distribution strategy for heterogeneous cloud platforms based on markov process modeling, IEEE Access 10 (2022), 26149–26162.

13.

Rajagopal

T.K.P.

Venkatesan

and Rajivkannan

, An improved efficient dynamic load balancing scheme under heterogeneous networks in hybrid cloud environment, Wireless Personal Communications 111(3) (2020), 1837–1851.

14.

Nabi

and Ahmed

, OG-RADL: Overall performance-based resource-aware dynamic load-balancer for deadline constrained cloud tasks, The Journal of Supercomputing 77(7) (2021), 7476–7508.

15.

Z.Q.

Wei

Zhang

Guo

and Xie

G.W.

, MDLB: A metadata dynamic load balancing mechanism based on reinforcement learning, Frontiers of Information Technology & Electronic Engineering 21(7) (2020), 1034–1046.

16.

Kaur

Singh

Devgan

M.S.

and Toor

H.K.

, Load balancing optimization based on deep learning approach in cloud environment, International Journal of Information Technology and Computer Science 12(3) (2020), 8–18.

17.

Kumar

M.B.

Reddy

A.P.

Kumar

M.P.

and Reddy

M.D.

, Improved virtual cluster management system for best performance in cloud computing, Annals of the Romanian Society for Cell Biology 25(6) (2021), 3929–3937.

18.

Haidri

R.A.

Katti

C.P.

and Saxena

P.C.

, Capacity based deadline aware dynamic load balancing (CPDALB) model in cloud computing environment, International Journal of Computers and Applications 43(10) (2021), 987–1001.

19.

Xie

Girshick

and Farhadi

, Unsupervised deep embedding for clustering analysis, in: International Conference on Machine Learning, 2016, pp. 478–487.

20.

Binu

and Kariyappa

B.S.

, RideNN: A new rider optimization algorithm-based neural network for fault diagnosis in analog circuits, IEEE Transactions on Instrumentation and Measurement 68(1) (2018), 2–26.

21.

Krishna

J.V.

Naidu

G.A.

and Upadhayaya

, A Lion-Whale optimization-based migration of virtual machines for data centers in cloud computing, International Journal of Communication Systems 31(8) (2018), 1–18.

22.

Brammya

Praveena

Ninu Preetha

N.S.

Ramya

Rajakumar

B.R.

and Binu

, Deer hunting optimization algorithm: a new nature-inspired meta-heuristic paradigm, The Computer Journal, May 2019. doi: 10.1093/comjnl/bxy133.

23.

Bhaladhare

P.R.

and Jinwala

D.C.

, A clustering approach for the-diversity model in privacy preserving data mining using fractional calculus-bacterial foraging optimization algorithm, Advances in Computer Engineering 1 (September 2014). doi: 10.1155/2014/396529.

24.

Karimi

and Khandani

, Social optimization algorithm with application to economic dispatch problem, International Transactions on Electrical Energy Systems 30(11) (2020), e12593.

25.

Xie

Yuan

Zhou

and Cheng

, Research on trust model in container-based cloud service, Computers, Materials and Continua 56(2) (2018), 273–283.

26.

Salehi

Boukerche

Darehshoorzadeh

and Mammeri

, Towards a novel trust-based opportunistic routing protocol for wireless networks, Wireless Networks 22(3) (2016), 927–943.

27.

Addesso

Barni

Mauro

M.D.

and Matta

, Adversarial Kendall’s model towards containment of distributed cyber-threats, IEEE Transactions on Information Forensics and Security 16 (2021), 3604–3619.

28.

Mauro

M.D.

Galatro

Fortino

and Liotta

, Supervised feature selection techniques in network intrusion detection: A critical review, Engineering Applications of Artificial Intelligence 101 (2021).

29.

Alazeez

M.A.

Abdalla

H.B.

and Lin

, NoSQL injection: Data security on web vulnerability, International Journal of Security and Its Applications 10 (2016), 55–64.

30.

Devagnanam

and Elango

N.M.

, Optimal resource allocation of cluster using hybrid grey wolf and cuckoo search algorithm in cloud computing, Journal of Networking and Communication Systems 3(1) (2021), 31–40.

31.

Thorat

V.A.

, Cloud intrusion detection using modified crow search optimized based neural network, Journal of Networking and Communication Systems 4(2) (2021), 9–15.

Hybrid trust-based optimized virtual machine migration for dynamic load balancing and replica management in heterogeneous cloud

Abstract

Keywords

1. Introduction

3. System model

4.1 Chunk creation

(i) Soft assignment

(ii) Kullback-Leibler divergence optimization

4.4.1 Solution encoding

(i) Migration cost

(ii) Predicted load

(iii) Energy consumption

(iv) Trust

4.5.1 Solution encoding

5.1 Evaluation metrics

5.2 Performance evaluation

Table 2 Comparative evaluation

5.3.2 Evaluation using Setup #2

5.4 Comparative discussion

6. Conclusion

Footnotes

Author’s Bios

References

Table 2
Comparative evaluation