RWWO: an effective strategy for workflow scheduling in cloud computing with predicted energy using Deep Maxout Network

Introduction

Cloud computing is the latest technology to improve virtualized resources, developed for consumers or end-users in a dynamic core to offer better and trustworthy services (Dillon, Wu, and Chang 2010). Cloud computing is a metered technology at different standards over the virtualized system to several consumers (Stephanakis et al. 2013). Also, it is a rapidly developing advancement that improves the computerization of both Information Technology (IT) infrastructure and IT resources. Cloud computing has had tremendous growth in research fields for the last few years and it is considered a system for computation that provides the resource to cloud users. Cloud computing satisfies the heterogeneous requirements utilizing Google mail, Amazon cloud, Google File System (GFS), and Hadoop framework are the more eminent reservoir of cloud computing resources (Jana, Chakraborty, and Mandal 2019; Netaji and Bhole 2020). The cloud computing services charge the users for entering the valuable resources over the network. The enterprises receive the resources as services rather than investing the expenses towards the resources, and thus, consumers can minimize their funding cost over the resources. Service Level Agreement (SLAs) is developed among service providers and consumers to point out guarantees of service providers to users (Kumar and Venkatesan 2019; Michael Mahesh 2020). The information available in the social media (Reddy Bojja et al. 2020) motivates the recommendation system (Hung and Chang 2019) in business (Hien Bui et al. 2021) and education (Hung 2020). The primary intention of the cloud is to offer an effective strategy for accurate manipulation of computational characteristics impelled in an organization and to assist the enterprise in exploiting consumer requirements (Al-Maytami et al. 2019). Due to the massive expansion and increasing demand for cloud computing resources (Singh and Chana 2016; Tsai, Lin, and Ke 2016), energy consumption is considered a major constraint in complicated cloud data centers. High energy utilization leads to heavy expenses and generates unrestricted amounts of heat emissions that frequently cause performance degradation and cannot operate appropriately (Ding et al. 2020).

The limitation of Task scheduling in the cloud environment is an eminent optimization issue. Cloud users always need an efficient load balancing model depending on distributed computing work for their effective deliveries. A highly-established cloud can optimize parameters like mobility, scalability, elasticity, availability, and minimum expense and improve throughput and storage capability (Beegom and Rajasree 2014; Jana, Chakraborty, and Mandal 2019). Task scheduling dispenses user works over resources to increase the consumption rate and reduce the work implementation time. Optimal task scheduling is a significant concept in the virtualization of the cloud (Tsai et al. 2014). The scheduling strategy is an NP-complete limitation, where the duration required for locating the solution changes concerning the size of the issue. The process of workflow scheduling (Juarez, Ejarque, and Badia 2018) is the most significant problem in the cloud environment, where efficient searching and resolutions are included in determining the best virtual machine (VM). Different computation-based performance metrics are utilized in the scheduling process, like network communication cost, traffic volume, and round trip (Abd Elaziz et al. 2019). Workflow scheduling directly influences the consumption of both cloud resources and Quality of Service (QoS) in response to consumer requests (Ding et al. 2020). The task scheduling mechanism in the cloud environment is outlined using the following phases: resource selection, resource searching, and task submission. The fundamental aim of cloud-based scheduling techniques is to reduce the task implementation time, reduce the transmission time of the task, minimize the execution expenses, and maximize balancing of load and resource availability (Karunakaran 2019; Masdari et al. 2017).

Task scheduling is broadly categorized into two categories, such as static and dynamic scheduling. In the former one, the scheduler learns about the task and resource details, such that the tasks are scheduled to the cloud service providers. However, in the latter one, task and resource particulars are unclear and possess more overhead than the former one. Therefore, the scheduler distributes the dynamic scheduling plans to consumer tasks with effective resources (Kumar and Venkatesan 2019). Optimal task scheduling can be categorized into three kinds: heuristic, meta-heuristic, and hybrid techniques. The heuristic task scheduling techniques deliver the best solutions to schedule the task, whereas in meta-heuristic task scheduling techniques can effectively handle search space to find the best result for workflow scheduling issues within the polynomial duration. The hybrid task technique is the combination of both heuristic and meta-heuristic task scheduling approaches (Abd Elaziz et al. 2019). So far, several algorithms have been developed to meet the requirements of scheduling the tasks, but they all face an NP-complete limitation. Different optimization methods have been utilized in the workflow scheduling process, like electro search, ant colony optimization, genetic algorithm, simulated annealing, tabu search, and particle swarm optimization. Due to the requirement of massive computing energy needed to implement the real-time task, it is essential to utilize an effective algorithm by considering all the functions required to improve workflow scheduling (Velliangiri et al. 2021).

The ultimate goal of this research is to design an effective strategy for workflow scheduling in a cloud environment employing a developed RWWO algorithm. The number of user requests in cloud computing is satisfied according to their needs. However, the problem arises when simultaneously many user requests arrive at the cloud requesting similar resources causes a lack of resources. In order to overcome such issues, a task scheduling technique using the proposed RWWO algorithm is implemented. This scheduling mechanism allocates the resources to the users depending on their implementation time, capacity, and energy. After that, the fitness function is evaluated according to three parameters, like predicted energy, QoS, and resource utilization. However, the Deep Maxout Network is utilized to compute the predicted energy. Moreover, the proposed RWWO algorithm is derived by combining Regressive Whale Optimization (RWO) and Water Cycle Algorithm (WCA).

The primary contribution of this research is given as below:

1. Proposed RWWO + Deep Maxout Network: An effective task scheduling mechanism is developed by the proposed RWWO algorithm for cloud computing. Due to a shortage of resources, it seems difficult to allocate the user’s tasks with the same resources. However, this limitation can be completely solved through the proposed RWWO algorithm by allocating the resources to the user’s tasks depending on their capacity, time consumed by the execution process, etc. Moreover, the predicted energy is determined using the Deep Maxout network.

The paper’s organization is organized as follows: Section 2 elaborates motivation for developing the proposed approach, and a literature survey of the conventional techniques is also explained. Section 3 describes the system model of cloud computing. Section 4 elaborates the developed RWWO + Deep Maxout Network, and the results and discussion are explained in Section 5, and the research comes to an end in Section 6.

Motivation

This section elaborates the literature review of the conventional approaches related to task scheduling mechanisms and their merits and limitations that motivate the researchers to design and establish an effective strategy for task scheduling.

System model

This section describes the system framework of the cloud environment. Typically, the environment of the cloud comprises a huge number of infrastructures and service providers to offer effective services to the consumers or end users according to their demands. The cloud environment consists of different physical machines, and it includes several virtual machines. The main purpose of utilizing the physical machines is to offer servers to the users based on their requests. However, an infinite number of requests is claimed for the same resource; there raises a huge problem in assigning the resources to the users due to a lack of sufficient resources. Each virtual machine consists of different configurations, such as memory, size of CPU, and cost for implementing the tasks. Hence, assigning an effective virtual machine that consumes a small amount of time and cost for implementing the task remains challenging in the cloud environment. In order to overcome such issues, it is essential to establish a workflow schedule. The scheduling mechanism sorts the request for offering service to end-users. Moreover, workflow is completely relying on the energy and capacity requires while implementing a task. The cloud environment comprises with n number of physical machines, and it is expressed as,

where, |M| specifies the total count of physical machines available in cloud, the individual physical machines are represented as M1,M2,...Mi,...Mn and Mi indicates the ith physical machine in cloud. The total virtual machines with respect to ith physical machine is represented as,

Here, Ii specifies the total number of virtual machines with respect to ith physical machine, h denotes total count of virtual machines and Ihi denotes h count of virtual machines similar to ith physical machine. However, the jth virtual machine of ith physical machine is expressed as Iji. Sji implies the resource cost of jth virtual machine exist at ith physical machine, whereas the capacity of jth virtual machine available at ith physical machine is indicated as Cji. Let us consider that there are h count of virtual machines and it is denoted as I. The capacity and task length are the significant factors to be considered while allocating resources according to user requests. Let us assume the total tasks as G and it is given by,

Here, G denotes the total task for execution purpose and l specifies the total count of tasks present in cloud environment. Moreover, gc indicates the cth task. However, each task includes a number of sub-tasks and sub-tasks in cth task is expressed using the following equation,

where, g1c,g2c,....,gbc,...gmc specifies the sub-tasks of cth task and m represents the total count of sub-tasks in cth task. The decision of a virtual machine to implement the task depends upon the task length. Moreover, task length signifies the waiting time of a task for execution.

Proposed RWWO for task scheduling and energy prediction based on Deep Maxout Network

This section elaborates on the proposed RWWO algorithm for the effective scheduling of work in cloud computing. In the cloud environment, an infinite count of user requests is satisfied according to the user demands. If several user requests reach the cloud environment, seeking similar resources is challenging because there are limited resources, and proper scheduling is difficult. There exist different techniques related to scheduling mechanisms. Still, they all faced many hurdles while transmitting the data between the data centers due to the high expense of resources, communication overhead, etc. However, some conventional workflow scheduling mechanisms reduced the cost of resources and the execution time, but they failed to effectively handle the heavy workflow’s complexity. In order to encounter the issues mentioned above, the developed RWWO algorithm is established to effectively handle the task scheduling mechanism based on priority. The task schedule mainly depends on three parameters, like QoS, resource utilization, and predicted energy. However, predicted energy is determined using Deep Maxout Network. The proposed RWWO algorithm is derived by incorporating RWO (Narendrababu Reddy and Phani Kumar 2019) and WCA (Eskandar et al. 2012). In addition, the proposed mechanism provides higher performance with low execution costs. Figure 1 represents the multi-objective scheduling utilizing the developed RWWO algorithm.

Figure 1:

Schematic view of the proposed method of multi-objective task scheduling using RWWO algorithm.

Energy prediction using Deep Maxout Network

Predicted energy is considered one of the fitness parameters computed using the Deep Maxout Network. Thus, the predicted energy enhances the overall performance of the system.

Architecture of Deep Maxout Network

Deep Maxout Network (Sun, Su, and Wang 2018) is an integral part of a Rectified Linear unit (ReLU), which employs the max operation on trainable linear functions. The main advantage of deep networks is that it effectively performs the complex structures of a huge set of activation functions. In the Deep Maxout Network, the activation function is substituted with a Maxout unit. The feature maps are achieved, and it is then applied to the higher layers. Moreover, each hidden unit in a Deep Maxout Network is considered a maxout unit, which enhances the potential in designing different distributions of various concepts. Deep Maxout Network is a type of trainable activation function comprised of a multi-layer structure. Figure 2 depicts the architecture of the Deep Maxout Network. The input applied to the Maxout Network is termed as E, the activation of a hidden unit is computed as follows:

(14)dq,r1=maxr∈[1,α1] Ev K....qr+aqr

(15)dq,r2=maxr∈[1,α2] dq,r1v k...qr+aqr

(16)⋮dq,rx=maxj∈[1,αx] dq,rx−1v K....qr+aqr

(17)⋮dq,ry=maxr∈[1,αy] dq,ry−1v K....qr+aqr

(18)Ep=maxr∈[1,αy] dq,ry

Figure 2:

Architecture of Deep Maxout Network.

Multi-objective task scheduling utilizing proposed RWWO

This section describes the mechanism of multi-objective task scheduling employing the developed RWWO algorithm. The works are assigned to a specific virtual machine according to its execution cost, energy, and implementation time. In order to execute a work effectively, the RWWO algorithm is utilized as it achieves fast convergence for the optimal location with reduced time. Moreover, it prevents the cost of execution tasks, thereby enhancing the effectiveness of the task scheduling mechanism. During the advent of various user requests, the proposed RWWO algorithm achieves optimized results.

Solution encoding

Figure 3 portrays solution encoding. Initially, the solution is selected and then it is subjected for determining the fitness evaluation. The randomly chosen solution is sort out and correct solution is computed exploiting function m mod (5)+1. Here, the virtual machine I5 executed the task g1, task g2 is executed by the virtual machine I2, the virtual machine I1 executed the task g3 and it goes on. However, the stored solution undergoes an operation using m mod(5)+1 and the achieved solution identifies the virtual machine, which is suitable for a specific task. Let us assume there are two physical machines, which is illustrated as M={M1,M2}. Here, two virtual machines is allocated to M1, whereas three virtual machines is assigned to M2. The total count of virtual machines is specified as I={I1,I2,I3,I4,I5}. A0 specifies the randomly selected solution, which is used to evaluate the fitness function. The randomly selected solution is sorted out depending on virtual machines’ potential and As denotes the organized solution. After that, the approximation is performed using the function m mod (B)+1 is represented as Ac. Then, the solution Ac is transformed into a matrix form, which is termed as A. Ajc is solution value of cth work executed in jth virtual machine. However, set matrix is determined based on two functions, such as capacity of the task, and length of the task. J implies the set matrix. Moreover, m implies the total count of tasks and five tasks are considered here.

Figure 3:

Solution encoding.

Algorithmic procedure of RWWO

This section elaborates the proposed RWWO algorithm, which is derived by combining the RWO (Narendrababu Reddy and Phani Kumar 2019) algorithm and WCA (Eskandar et al. 2012). The WCA effectively solves the optimization issues based on the cost function. The RWO is an algorithm that is a modified version of WOA, which is easy to implement and has the fast convergence rate. The major benefit of utilizing this RWWO algorithm is that it finds an optimal position depending on the global search mechanism and provides minimum search cost. RWO algorithm is obtained by the incorporation of the regressive parameter of CAViaR’s model (Engle and Manganelli 2004) with Whale Optimization Algorithm (WOA) (Chen et al. 2020). Thus, the incorporation of the three algorithms efficient performance enhancement is achieved in term of cost and time.

Step 1:

Initialize the population

This is the first step of this algorithm, in which the size of the population is initialized. Let us consider the population of the whale as V, and it is expressed as,

(19)V={V1,V2,…Vo,…Vtp}

Here, tp denotes the total population with Vo, which represents the oth population. The whales search for their food depending upon the optimal position based on the global search mechanism.

Step 2:

Evaluation of fitness function

The fitness function is utilized to determine the best optimal solution. The fitness value is determined utilizing three functions, namely QoS, resource utilization, and predicted energy.

Step 3:

Localization of prey

The next step is to determine the location of prey depending on the current best solution. The localization process is generally based on the encircling phase. The whales encircle the prey at this phase that paves a way to the location of the prey. Thus, the prey has a fixed position, and the searching process is repeated concerning the optimal solution. The current optimal solution is updated utilizing the following equation,

(20)f→=|N→.z‾∗ (k)−z→(k)|

where, z‾∗ (k) represents the current best solution, position vector is specified as z (k) and f→ depicts the distance between the whale and prey. The coefficient vector is represented as N→. The above equation is rewritten as,

(21)z∗(k)=f‾+z→(k)N→

By considering the autoregressive model, position vector is expressed as,

(22)z→(k)=β0+β1 zt−1+β2 zt−2+β1 P (zt−1)+β2 P (zt−2)

Substituting z→(k) in Eq. (21), the equation becomes,

(23)z∗(k)=f→+β0+β1 zt−1+β2 zt−2+β1 P (zt−1)+β2 P (zt−2)N→

where, f→ implies the distance vector, the regressive constant are denoted as β0,β1,β2. z∗(k) is the best position of prey. Hence, the best position is improved based on the below-expressed equation,

(24)z→(k+1)=z∗(k)−W→.f→

Substituting the value of z∗(k) in Eq. (24), then the equation becomes,

(25)z→(k+1)=f→+β0+β1 zt−1+β2 zt−2+β1 P (zt−1)+β2 P (zt−2)N→−W→.f→

(26)z→(k+1)=f→+β0+β1 zt−1+β2 zt−2+β1 P (zt−1)+β2 P (zt−2)−(W→.f→) N‾N→

(27)z→(k+1)=f→ (1−W→.N→N)+β0+β1 zt−1+β2 zt−2+β1 P (zt−1)+β2 P (zt−2)N→

Therefore, Eq. (27) is the newly updated position of prey by employing RWO algorithm. Now, substituting the Eq. (20) in Eq. (27), then the final update equation of RWO algorithm is,

(28)z→(k+1)=(N→.z→∗(k)−z→(k))(1−W→ N→N)

where, z→(k+1) represents the final update equation, z∗(k) is the best position of prey, z‾∗ (k) represents the current best solution, coefficient vector is represented as N→ and W→.

At this step, the WCA algorithm is incorporated with the RWO algorithm to provide an optimal global solution. WCA is a nature-inspired algorithm that depends on the mechanism of water cycle mechanism and how streams and rivers flow into the sea. The complexities in computational performance and then the convergence rate is enhanced using the below equation,

(29)z→∗(k)=z→(k+1)−ω rand w (1,Xvar)

As z→(k+1) is considered as the optimal solution in WCA, then the Eq. (28) is substituted with Eq. (29),

(30)z→(k+1)=(N→ (z→ (k+1)−ω rand w (1,Xvar))−z→(k)) (1−W→ N→N→)+β0+β1 zt−1+β2 zt−2+β1 P (zt−1)+β2 P (zt−2)N→

(31)z→(k+1)=(N→ (z→ (k+1)−(N→ω rand w (1,Xvar))−z→(k)) (1−W→ N→N→)+β0+β1 zt−1+β2 zt−2+β1 P (zt−1)+β2 P (zt−2)N→

(32)z→(k+1)=(N→ (z→ (k+1)(1−W→ N→N→)−N→ω rand w (1,Xvar))+z→(k)) (1−W→ N→N→)+β0+β1 zt−1+β2 zt−2+β1 P (zt−1)+β2 P (zt−2)N→

(33)z→(k+1)−(N→ (z→ (k+1)(1−W→ N→N→)=β0+β1 zt−1+β2 zt−2+β1 P (zt−1)+β2 P (zt−2)N→−(N→ω rand w (1,Xvar))+z→(k)) (1−W→ N→N→)

(34)z→(k+1)=1W→ N→ β0+β1 zt−1+β2 zt−2+β1 P (zt−1)+β2 P (zt−2)N→−(N→ω rand w (1,Xvar))+z→(k)) (1−W→ N→N→)

Step 4:

Update the location of prey

After finding the location of the prey, the location of the prey is identified. A parameter known as the probability factor is utilized to determine the position of the prey. If the probability factor is less than that of 0.5, the process of localization is carried out. If the probability factor is greater than that of 0.5, then the process of exploitation is performed. The exploitation phase consists of two steps: encircling the prey with the help of bubbles and position update of spiral movement. In position update of the spiral movement, the distance between the location of the prey and the whale is determined using the following equation,

(35)f→=z⇀∗(k)−z⇀(k)

Rearranging the Eq. (35),

(36)z⇀∗(k)=f→∗ z→ (k)

Substituting the value of z→(k) expressed in Eq. (22) in Eq. (36), the obtained equation is represented as,

(37)z⇀∗(k)=f⇀∗+β0+β1 (zt−1)+β2 (zt−2)+β1 P (zt−1)+β2 P (zt−2)

The position update in the exploration step is based on the following equation,

(38)z→(k+1)=f→∗⋅TkZ⋅cos (2πZ)+z∗(k)

Substitute the value of z⇀∗(k) in z⇀(k+1),

(39)z⇀(k+1)=f⇀∗(1+TkZ.cos (2πZ))+β0+β1 (zt−1)+β2 (zt−2)+β1 P (zt−1)+β2 P (zt−2)

Substitute Eq. (35) in Eq. (39),

(40)z→(k+1)=(z⇀∗(k)−z⇀(k)) (1+TkZ.cos (2πZ))+β0+β1 (zt−1)+β2 (zt−2)+β1 P (zt−1)+β2 P (zt−2)

Now, incorporating the Eq. (29) of WCA algorithm,

z⇀∗(k+1)=Z⇀ (k+1)−ω rand w (1,Xvar)

(41)z⇀(k+1)=(z⇀(k+1)−ω rand w (1,Xvar)−z⇀(k))(1+TkZ.cos (2πZ))+β0+β1 (zt−1)+β2 (zt−2)+β1P (zt−1)+β2 P (zt−2)

(42)z⇀(k+1)=z⇀(k+1) (1+TkZ.cos (2πZ))−(ω rand w (1,Xvar)+z⇀(k)) (1+TkZ.cos (2πZ))+β0+β1 (zt−1)+β2 (zt−2)+β1 P (zt−1)+β2 P (zt−2)

(43)z⇀(k+1)−z⇀(k+1) (1+TkZ.cos (2πZ))=β0+β1 (zt−1)+β2 (zt−2)+β1 P (zt−1)+β2 P (zt−2)−(ω rand w (1,Xvar)+z⇀(k)) (1+TkZ.cos (2πZ))

(44)z⇀(k+1) (1−1−TkZ cos (2πZ))=β0+β1 (zt−1)+β2 (zt−2)+β1 P (zt−1)+β2 P (zt−2)−(ω rand w (1,Xvar)+z⇀(k)) (1+TkZ.cos (2πZ))

(45)z⇀(k+1)=1TkZ cos (2πZ)[(ω rand w (1,Xvar)+z⇀(k)) (1+TkZ.cos (2πZ))−β0−β1 (zt−1)−β2(zt−2)−β1 P (zt−1)−β2 P (zt−2)]

The above equation is the final updated equation of the RWWO algorithm.

However, in the exploration step, the position update is depending on the global search mechanism, and it is represented as,

(46)z→(k+1)=z→δ W→.f⇀

(47)z→(k+1)=f→ (1−W⇀.f⇀)N⇀+β0+β1 (zt−1)+β2 (zt−2)+β1 P (zt−1)+β2 P (zt−2)N⇀

Step 5:

Termination

The process is continued until the best optimal solution is reached and the condition is satisfied. Then, the updated equation is considered as the best solution in the global search space. Finally, Algorithm 1 portrays the pseudo-code of the developed RWWO algorithm.

Algorithm 1.

Pseudo code of RWWO algorithm

Sl. No	Pseudo code of developed RWWO algorithm
1	Input: V={V1,V2,...,Vo,...,Vtp}
2	Output:z⇀(k+1)– Best position
3	Parameters: N⇀,W⇀,γ
4	Read
5	Update the parameters
6	If (γ < 0.5)
7	And if (W→<1)
8	Update the position using Eq. (27).
9	Else
10	Update the position using Eq. (45).
11	Else if
12	Update the position based on Eq. (46).
13	Compute the fitness parameter
14	Return the best position, z⇀(k+1)
15	End

Results and discussion

This section describes the results and discussion of the developed RWWO + Deep Maxout Network concerning the evaluation metrics.

Comparative analysis

This section elaborates the developed RWWO + Deep Maxout Network analysis using three experimental setups concerning the evaluation metrics, like QoS, resource utilization, and predicted energy. The setup-1 includes 10 physical machines (PM), 41 Virtual machines (VM), five tasks, and 17 sub-tasks. In experimental setup-2, the system is comprised of 10 PM, 43 VM, 10 tasks, and 23 sub-tasks. However, setup-3 includes 5 PM and 20 VM for scheduling five tasks with eight sub-tasks.

Analysis based on setup-1

Figure 4 depicts the analysis of the proposed approach using setup-1 concerning the evaluation metrics by changing the number of iterations. Figure 4a represents the analysis of task scheduling time. If the number of iteration = 10, the time attained by the developed RWWO + Deep Maxout Network is 0.185. However, the conventional schemes achieved the scheduling of 0.5132 for ACO, 0.5132 for PSO, 0.2149 for WOA, and 0.4654 for RWO. By varying the number of iteration to 40, the scheduling time achieved by the traditional techniques, like ACO is 0.5132, PSO is 0.3237, WOA is 0.2149, and RWO is 0.1217, respectively, and the proposed RWWO + Deep Maxout Network attained the scheduling time of 0.0208.

Figure 4:

Analysis using experimental setup-1, a) task scheduling time, b) task scheduling cost, c) resource utilization, d) predicted energy.

The analysis of task scheduling cost is depicted in Figure 4b. If the iteration = 10, the scheduling cost obtained by developed RWWO + Deep Maxout Network is 0.00815, while the conventional techniques, such as ACO, PSO, WOA, and RWO attained scheduling cost 0.0094, 0.0098, 0.0098, and 0.0118, respectively. Similarly, if the number of iteration = 40, cost achieved by the developed RWWO + Deep Maxout Network is 0.0017. However, the scheduling cost obtained by the conventional approaches is 0.0082 for ACO, 0.0091 for PSO, 0.0098 for WOA, and 0.0046 for RWO.

The analysis of predicted energy is represented in Figure 4c. When the number of iteration = 10, the predicted energy obtained by the proposed scheme is 0.290, whereas the existing methods achieved the predicted energy of 0.5003 for ACO, 0.4792 for PSO, 0.3638 for WOA, and 0.4315 for RWO. By increasing the iteration number to 40, predicted energy obtained by the traditional schemes, such as ACO is 0.4177, PSO is 0.4725, WOA is 0.3638, and RWO is 0.2859. However, the proposed RWWO + Deep Maxout Network attained the predicted energy is 0.197.

Figure 4d represents the analysis of resource utilization. When the number of iteration = 10, the resource utilization achieved by the proposed RWWO + Deep Maxout Network is 0.998, whereas the existing methods, like ACO, is 0.9971, PSO is 0.9971, WOA is 0.9979, and RWO is 0.9974. Correspondingly, if the number of iteration = 40, the resource utilization achieved by the proposed RWWO + Deep Maxout Network, conventional schemes ACO, PSO, WOA, and RWO is 0.999, 0.9973, 0.9978, 0.9979, and 0.9989, respectively.

Analysis based on setup-2

Figure 5 illustrates the analysis using setup-2 concerning the performance metrics by changing the count of iterations. Figure 5a portrays the analysis of task scheduling time by changing the number of iterations. If the number of iteration = 10, the task scheduling time obtained by the proposed RWWO + Deep Maxout Network is 0.247. However, the time achieved by the conventional approaches, like ACO is 0.2701, PSO is 0.3408, WOA is 0.2771, and RWO is 0.5837. Similarly, if the number of iteration = 40, the task scheduling time attained by the developed scheme and the traditional schemes, such as ACO is 0.0218, PSO is 0.2701, WOA is 0.2583, and RWO is 0.1619.

Figure 5:

Analysis using experimental setup-2, a) task scheduling time, b) task scheduling cost, c) resource utilization, d) predicted energy.

Figure 5b depicts the proposed RWWO + Deep Maxout Network analysis concerning the task scheduling cost. If the number of iteration = 10, the cost attained by the traditional techniques, like ACO is 0.0111, PSO is 0.0141, WOA is 0.0129, and RWO is 0.0251. However, the proposed RWWO + Deep Maxout Network obtained the task scheduling cost is 0.0102. Likewise, the number of iteration = 40, the task scheduling cost achieved by developed RWWO + Deep Maxout Network is 0.0045, while the traditional techniques, such as ACO is 0.0111, PSO is 0.0105, WOA is 0.0121, and RWO is 0.0064, attained the scheduling cost.

Figure 5c represents the proposed RWWO + Deep Maxout Network analysis concerning the predicted energy. If the number of iteration = 10, the proposed RWWO + Deep Maxout Network’s predicted energy is 0.336, while the existing methods like ACO are 0.3688, PSO is 0.5580, WOA is 0.5700, and RWO is 0.5141. By varying the number of iteration = 40, the predicted energy attained by the proposed RWWO + Deep Maxout Network is 0.177.

The proposed RWWO + Deep Maxout Network analysis concerning resource utilization is represented in Figure 5d. If the number of iteration = 10, the resource utilization consumed by the proposed RWWO + Deep Maxout Network is 0.998. However, the conventional schemes, such as ACO, PSO, WOA, and RWO, utilized 0.9981, 0.9978, 0.9980, and 0.9966. Similarly, if the number of iteration = 40, the resource utilization attained by the conventional approaches, like ACO is 0.9981, PSO is 0.9984, WOA is 0.9983, and RWO is 0.9992. However, the proposed RWWO + Deep Maxout Network attained the resource utilization as 0.9995.

Analysis based on setup-3

Figure 6 illustrates the proposed RWWO + Deep Maxout Network analysis using setup-3 concerning the performance metrics. The analysis of task scheduling time is represented in Figure 6a. When number of iteration = 10, the scheduling time achieved by developed RWWO + Deep Maxout Network is 0.0173, whereas the conventional schemes, like ACO, PSO, WOA, and RWO attained the task scheduling time of 0.1324, 0.1228, 0.0242, and 0.0182. Similarly, if the number of iteration = 40, the task scheduling time attained by existing schemes, like ACO is 0.0238, PSO is 0.0224, WOA is 0.0224, and RWO is 0.0182. The proposed RWWO + Deep Maxout Network attained the task scheduling time as 0.0035.

Figure 6:

Analysis using experimental setup-3, a) task scheduling time, b) task scheduling cost, c) resource utilization, d) predicted energy.

Figure 6b represents the analysis of task scheduling cost. If the number of iteration = 10, the cost achieved by the proposed RWWO + Deep Maxout Network is 0.001061. However, the conventional schemes achieved the task scheduling cost of 0.0032 for ACO, 0.0076 for PSO, 0.0015 for WOA, and 0.0011 for RWO. By varying the number of iteration = 40, the scheduling cost obtained by the proposed RWWO + Deep Maxout Network, existing ACO, PSO, WOA, and RWO is 0.000557, 0.0012, 0.0007, 0.0013, and 0.0011, respectively.

The analysis of predicted energy concerning the number of iterations is illustrated in Figure 6c. If the number of iteration = 10, the proposed RWWO + Deep Maxout Network’s predicted energy is 0.3314. However, the conventional schemes, like ACO, PSO, WOA, and RWO attained the predicted energy of 0.5623, 0.3319, 0.3849, and 0.3710. Similarly, if the number of iteration = 40, the proposed RWWO + Deep Maxout Network’s predicted energy is 0.2065, whereas the traditional approaches attained predicted energy of 0.3460 for ACO, 0.2292 for PSO, 0.2665 for WOA, and 0.3710 for RWO.

Figure 6d represents the analysis of resource utilization. If the number of iteration = 50, the proposed RWWO + Deep Maxout Network resource is 0.9998, whereas 0.99997 for iteration = 40. However, the existing schemed consumed the resource utilization of 0.9997 for ACO, 0.9997 for PSO, 0.9999 for WOA, and 0.9999 for RWO for the number of iteration = 10.

Conclusions

Cloud computing has tremendously increased its growth, and this technology has a rapid evolution over the past few years because of its simplicity and easy way of accessing resources. This technology plays a vital role in many applications, like software, computing, storage, network, and various heterogeneous needs. Task scheduling is an essential mechanism in cloud computing in allocating the appropriate resources to the users. However, there is a shortage of resources in serving the users when multiple users request similar resources simultaneously. Therefore, it is necessary to provide cost-effective executions with proper scheduling of tasks. Hence, this research presents an effective mechanism for task scheduling using the RWWO algorithm. The proposed approach provides hassle-free executions in the task scheduling mechanism by allocating suitable resources to the machines according to the user requirements. The fitness parameters utilized in this mechanism are QoS, resource utilization, and predicted energy. However, the predicted energy is determined using Deep Maxout Network. Moreover, the proposed RWWO + Deep Maxout Network achieved a minimum task scheduling time of 0.0208, minimum task scheduling cost of 0.0017, minimum predicted energy of 0.1971, and maximum resource utilization of 0.9999. There is still room for enhancement in enhancing the performance of the workflow scheduling mechanism by adding some other optimization algorithms. The implications of the proposed method is efficient for the scheduling the work flow for the selection of appropriate resource. In the future, the new optimization technique based on deep learning will be implemented for the efficient allocation of resource based on the workflow.