1 Introduction

To provide communications, cloud computing has received enormous attention in recent years as well as it needs massive resources with services for executing large-scale applications [1]. Therefore, it joins similar ideas also knowledge toward giving mutual assets, tools, software, and information meant for individual computers with other machines [2]. Several computer frameworks are suggested for the enormous quantity of data storage as well as the computer needs of cloud computing [3]. Virtual Machine (VM) is one of the main software in cloud computing, it supports all windows in single software [4]. Also, through cloud infrastructure users can run large-scale workloads on VMs hosted. It improves application completion time and enables parallel processing of application tasks [5]. Utilizing accessible computing resources as effectively as possible is the primary goal of a cloud computing system. Scheduling tasks in suitable order so they can be completed under problem-specific limitations is the primary goal of scheduling [6].

Workflow scheduling problem on resources is the most essential issue for the uses of cloud environment [7]. Single cloud computing is flat to resource problems like hardware breakdown, software breakdown as well as power breakdown similar to other distributed computing [8]. Single cloud computing meets some disadvantages given user needs in scheduling tasks [9]. While operating a complex workflow program, these issues are expected and fixed develop task failures and workflow system faults [10]. Thus, the method considers the multi-cloud technique, it satisfies the customers with a variety of options and toward greatest assure their purpose necessities, particularly intended for persons scientific computing request [11]. Between various cloud contributors, this might be the better answer for the exchange of resources [12]. The cloud is a platform for large dispersed computing that is available for streaming for various process applications [13].

A crucial performance metric for cloud-based process scheduling is durability [14]. However, the reliability criteria used by various cloud providers vary. Users must therefore pay particular consideration to the dependability restriction of the workflow when choosing computer resources in a multi-cloud environment [15]. We have suggested multi-objective scheduling in this paper, which is taken into account for scientific processes in a multi-cloud context. The primary goal is to reduce makespan time and cost while also following dependability constraints. An optimization algorithm like the Adaptive Dingo Optimization Algorithm (ADO) is proposed to achieve the multi-objective scheduling with optimal values.

2 Related works

Many of the researchers had developed workflow scheduling on the cloud. Among them some of the works are analyzed here.

Hu et al. [16] had explained a multi-objective scheduling (MOS) algorithm for scientific workflow in a multi-cloud environment. The MOS algorithm is based on Particle Swarm Optimization (PSO), as well as the equivalent code strategy to get the tasks finishing place, as well as tasks, categorize information transmission toward concern. Widespread experimental explained the important multi-objective presentation development of MOS algorithm better result compared with CMOHEFT algorithm as well as the RANDOM algorithm.

Adhikari et al. [17] had discovered via allowing for numerous contradictory objectives utilizing the Firefly Algorithm (FA). The goal of FA is to match each workflow with a relevant cloud platform that can accommodate its requirements for load balancing and resource usage of the cloud servers. Additionally, a rule-based technique is built to distribute the tasks to the appropriate VM instances to shorten the duration of the workflow as well as gather the deadline. Using the approach, parameters like makespan, reliability, resource utilization and loads of the cloud servers were well defined.

For scheduling dependent tasks toward VMs Abed-Alguni et al. [18] introduce a discrete variation of the Distributed Grey wolf Optimizer (DGWO). The computation, as well as data transmission costs, are the two objectives for the scheduling process in DGWO. Compared with existing algorithms like PSO and GWO, DGWO gives better results and it was experimentally tested. Also, compared with other tested algorithms, the experimental outcome suggests that DGWO distributes tasks to VMs faster and it gives the makespan time. Based on Adaptive resource Allocation as well as Consolidation an Online Workflow Scheduling algorithm was proposed by Chen et al. [19]. This algorithm performs better compared to other methods.

Alaei et al. [20] develop an adaptive fault detector strategy based on the Improved Differential Evolution (IDE) algorithm. The study utilizes an Adaptive Network-Based Fuzzy Inference System (ANFIS) forecasting framework to prevent resource load fluctuation, which improves fault predictive performance before fault development. Compared with existing techniques, the method considerably enhances the whole arrangement presentation, attains a superior quantity of error acceptance with tall Hyper Volume (HV) contrasted by way of ICFWS, IDE, also ACO algorithms, reduces the makespan, the power expenditure as well as job error fraction, and minimizes the overall cost (The summary of literature survey is presented in Table 1).

Table 1 Summary of literature survey
Full size table

Chakravarthi et al. [21] had proposed a scheduling algorithm called as Multi-Criteria Decision Making (MCDM) approach. According to the task requirement, a weighted total of run time cost, as well as the transfer of data instances, is utilized to determine the optimal resource among the available resources. The investigational outcome revealed that T-BDMWS provides present modern heuristics among the condition of attaining the user-specified resources otherwise deadline constriction and resource competence.

Medara et al. [22] presented an energy-aware algorithm in cloud computing called EASVMC with VM consolidation. It takes the multi-objective like consumption of energy, resource utilization as well as VM migrations. Algorithm-like inspired meta-heuristic approach called the Water Wave Optimization (WWO) toward lessening the energy consumption which finds an appropriate migration also lever off redundant hosts after migrating its VMs toward a suitable target swarm.

3 Workflow model using DAG

A scientific workflow is consist of several tasks which is modeled as a DAG. A workflow can be modeled as W = (A, E), where A = {T1, T2, …, Tn} denotes the set of n tasks in the workflow as well as E = {(Ti, Tj)|Ti, Tj ∈ A} denotes the set of task dependencies. All tasks in the process must be completed within the deadline, which is a deadline constraint D attached to the workflow W. The simple representation of the DAG workflow [8] is shown in Fig. 1.

Fig. 1
figure 1

Simple DAG model

Full size image

T1 and T10 represent the incoming and outgoing tasks, respectively, in Fig. 1. Tasks T2 and T3 are offspring of task T1. Only once the parent’s job has been completed may the child’s tasks be run. To be discovered is the ideal task–VM (task, VM) pair.

4 Multi-objective function design

To run extensive experimental operations, an IaaS platform offers processing resources in the form of VMs. The proposed approach is taken into account in this scenario as a multi-cloud system for process scheduling. Two or more cloud processes connecting over the Internet make up a multi-cloud architecture. Cloud customers have the availability of VM resources from all cloud providers in our multi-cloud ecosystem. The features of each cloud, however, and the specific pricing structures of cloud platforms vary. A global cloud manager receives an application, divides the related workflow into numerous tasks, and presents them to the cloud scheduler. The tasks are scheduled by the cloud scheduler using a multi-objective scheduling technique. The tasks are then distributed to the accessible VMs using the local scheduler installed in each data center. The three characteristics of makespan, cost, and dependability were taken into account by the proposed method for multi-objective scheduling. The main goals of the suggested scheduling model are to decrease each task’s makespan, cost, and dependability restriction. Control characteristics are established to outline this problem. The equation can be used to describe the issue (1).

$$\begin{aligned} & Minimize{:}\,O(S) = (makespan,\cos t) \hfill \\ & Subject\,to{:}\,R \ge r_{con} \,;\,\,\,\,\,r_{con} \in [r_{\min } ,r_{\max } ] \\ \end{aligned}$$
(1)

where

$${\text{Re}} liability = R = \Pi_{a_i \in A} r(a_i )$$
(2)
$$r(a_i ) = \exp ( - \xi \cdot T_{rent} (a_i ,VM(n,m)));\,\,\,\,\,\xi > 0$$
(3)
$${\text{Cos}} t = C = \sum_{a_i \in A} {C(a_i ,VM(n,m))}$$
(4)
$$makespan = A_{end} (a_{exit} )$$
(5)
$$A_{end} (a_i ) = A_{start} (a_i ) + A_{recev} (a_i ) + A_{exeu} (a_i ,VM(n,m))$$
(6)
$$A_{exeu} (a_i ,VM(n,m)) = \frac{W(a_i )}{{P(n,m)}}$$
(7)

where \(P(n,m)\) represent the processing capacity.

The objective function mentioned above will be minimized in this study utilizing the proposed approach. For that, the proposed method considered the adaptive dingo optimization (ADO) algorithm.

5 Proposed workflow scheduling methodology

A framework for emerging capabilities and the capacity to provide reliable cloud services in cloud computing. The availability of “unlimited” computing resources to consumers on-demand is one of the key components of cloud computing. Single cloud holding resources, nevertheless, are typically constrained and could not be sufficient to handle the unexpected spike in user demands. To allow resource sharing amongst clouds, the multi-cloud concept is proposed. Offering resources and services from several clouds is a paradigm that is getting more and more popular today. This research proposes a multi-objective scheduling method for scientific workflow in a multi-cloud environment, to concurrently reduce workflow duration and cost while satisfying the reliability requirement. To achieve this concept adaptive dingo optimization (ADO) algorithm is designed. The overall flow diagram of the proposed model is shown in Fig. 2.

Fig. 2
figure 2

Overall flow diagram of the proposed model

Full size image

The main aim of the proposed model is,

  • To improve the efficiency of cloud scheduling, a large number of tasks are mapped onto cloud resources using a workflow model. Here DAG (directed acyclic graph) is used for the workflow model.

  • To support resource sharing between the clouds, the multi-cloud concept is introduced here.

  • Multi-objective function-based scheduling is presented using adaptive dingo optimization (ADO). This ADO helps to diminish the makespan and cost and improve the reliability constraint of the proposed workflow model.

  • The efficiency of recommended technique is analyzed based on different metrics namely, makespan, cost, and reliability. A detailed explanation of each process is described in a further section.

5.1 Multi-objective scheduling strategy using ADO

The scheduling parameters such as makespan, cost, and reliability are optimized by using ADO in the scheduling process. The DOA is a brand-new bio-inspired global optimization algorithm that imitates dingoes’ hunting tactics. Group hunting is an intriguing aspect of dingoes’ social activity, which furthers the social behavior of dingoes. Predatory strategies are classified into their stages as follows: chasing and approaching, encircling and harassing, and attacking. Here the traditional dingo optimization is modified using updation process [23]. The step-by-step process of ADO is explained in beneath,

Step 1: Solution encoding

Solutions are representing the flow of tasks. In each optimization algorithm, encoding solution is a significant step. In this proposed work, the solution is considered as the task workflow. A workflow can be modeled as W = (A, E), where A = {a1, a2, …, an} represents the set of n tasks in the workflow as well as E = {(ai, aj)|ai, aj ∈ A} denotes the set of task dependencies.

Step 2: Fitness evaluation

Once the initial population is completed, the fitness function is computed. The fitness function is evaluated with the consideration of makespan, cost and reliability constrained. The objective of the proposed model is to minimize the makespan and cost value and improve the reliability constrained of each task. The fitness function is mathematically formulated in Eq. (1).

Step 3: Encircling process

After fitness evaluations, solutions are updated using dingo optimization. Dingoes are intelligent sufficient to locate their prey. The pack, led by the alpha, circles the prey after locating it. The following analytical solutions simulate these dingoes’ behavior,

$$\vec{D}_d = \left| {\vec{A}\cdot{\mathop{P}\limits^{\rightharpoonup}}_p \left( X \right) - \vec{P}\left( i \right)} \right|$$
(8)
$${\mathop{P}\limits^{\rightharpoonup}} \left( {i + 1} \right) = \mathop{P}\limits^{\rightharpoonup} _p \left( X \right) - {\mathop{B}\limits^{\rightharpoonup}} .\vec{D}\left( d \right)$$
(9)
$$\mathop{A}\limits^{\rightharpoonup} = 2.\mathop{a}\limits^{\rightharpoonup} _1$$
(10)
$$\mathop{B}\limits^{\rightharpoonup} \, = \,2\mathop{b}\limits^{\rightharpoonup} .\mathop{a}\limits^{\rightharpoonup} _2 \, - \mathop{b}\limits^{\rightharpoonup}$$
(11)
$$\vec{b}\, = \,3 - \left( {1 \ast \left( {\frac{3}{{l_{\max } }}} \right)} \right)$$
(12)

positions of neighboring dingoes are addressed using a two-tier level vector given in Fig. 3. As mentioned by the location of the prey \(({\text{P}}*, {\text{Q}}*),\) a dingo (P, Q) can update its position in place. By changing the value of the vectors \(\overrightarrow{A}\) and \(\overrightarrow{B}\) for the current area, each of the possible areas is arranged individually on the map around the best expert. The position vector of dingoes [23] is presented in Fig. 3.

Fig. 3
figure 3

Structure of workflows a Montage, b CyberShake, and c LIGO

Full size image

Step 4: Hunting process

The position of possible prey is well known to all of the pack members, particularly alpha, beta, and others. The hunt is always led by the dominant dingo. Beta and other dingoes, though, could occasionally join in on the hunts. Other dingoes must change their positions by the location of the best search agent.

$$\overrightarrow {D_\alpha } = \left| {\overrightarrow {a_1 } . \overrightarrow {P_\alpha } - \vec{p} } \right|$$
(13)
$$\overrightarrow {D_\beta } = \left| {\overrightarrow {a_1 } . \overrightarrow {P_\beta } - \vec{p} } \right|$$
(14)
$$\overrightarrow {D_0 } = \left| {\overrightarrow {a_1 } . \overrightarrow {P_0 } - \vec{p} } \right|$$
(15)
$$\overrightarrow {P_1 } = \left| {\overrightarrow {P_\alpha } . \vec{b} - \overrightarrow {D_\alpha } } \right|$$
(16)
$$\overrightarrow {P_2 } = \left| {\overrightarrow {P_\beta } . \vec{b} - \overrightarrow {D_\beta } } \right|$$
(17)
$$\overrightarrow {P_3 } = \left| {\overrightarrow {P_0 } . \vec{b} - \overrightarrow {D_0 } } \right|$$
(18)

In the dingo optimizer, the intensity of every dingo is computed based on the below equations,

$$\overrightarrow {I_\alpha } = log\left( {\frac{1}{f_\alpha - (1E - 100)} + 1} \right)$$
(19)
$$\overrightarrow {I_\beta } = log\left( {\frac{1}{f_\beta - (1E - 100)} + 1} \right)$$
(20)
$$\overrightarrow {I_0 } = log\left( {\frac{1}{f_0 - (1E - 100)} + 1} \right)$$
(21)

In this, we can undoubtedly portray the situation updated by alpha, beta, and remaining dingo. Dingos (alpha, beta, and others) may also feel updated. Their positions are arbitrary and work in space prey in the hunting ground.

Step 5: Attacking prey process

Dingo concluded the hunt by assaulting the target if there is no location update.

Step 6: Searching process

Dingoes search for prey primarily based on the position of the pack. They continuously move forward to pursue and attack predators.

Step 7: Termination criteria

When the best fitness value is chosen, the algorithm stops working. The allocation procedure uses the chosen ideal task. The algorithm of scheduling is presented in Table 2.

Table 2 Algorithm for ADO based workflow scheduling
Full size table

6 Result and discussion

This section analyses the actual outcomes of the suggested workflow scheduling. Java is used to execute the suggested scheduling technique. The device is powered by an Intel Core i5 CPU and runs Windows 10 on a device with 6 GB of RAM. Java is used to simulate the suggested approach. Four different types of workflows, including Montage, CyberShake, Sipht, and LIGO, are employed for experimentation analysis [24, 25]. Figure 3, shows the workflow’s organizational structure.

6.1 Experimental results

Here, the experimental analysis of various workflow models has been taken also simulation is done with the help of a cloud simulator in the Java domain to minimize the makespan time, and cost and satisfy the reliability constraint. The proposed algorithm is compared with existing algorithms like FA, GWO, and PSO. The dataset details are given in Tables 3 and 4.

Table 3 Amazon EC2 VM instance specification
Full size table
Table 4 Google compute engine VM instance specification
Full size table

Table 5 explains the best score value for various tasks that correspond with virtual machines. For algorithm iterations, we have obtained distinct values. And it is calculated for our proposed and existing algorithms. In task T1, the best score obtained for ADO is 245, FA is 243and GWO is 241. Similarly, other algorithms give their best values. Our proposed model ADO gives the best score attained compared with existing algorithms. Also, we have gotten the fitness value best for the proposed ADO algorithm.

Table 5 Various tasks based on algorithm iterations
Full size table

Figure 4, discusses to the minimum makespan time based on various iterations. Here the iterations are changed from 20 to 40. By utilizing the three algorithms the makespan time differs from 1200 to1550. The makespan time is contrasted with three algorithms FA, GWO, and ADO. We have acquired the greatest influence traverse to time got for FA and GWO algorithm. In our proposed ADO algorithm, we have accomplished the least makespan time.

Fig. 4
figure 4

Convergence graph for makespan time

Full size image

Figure 5, explains a comparative chart for various tasks done with different scientific workflow models. The models are a montage, SIPHT, LIGO, and Cybershake. The X-axis mentions various tasks from the user and the y-axis mentions makespan time represented in h. Here, our existing algorithms like FA, GWO, and GSO. Here, we have gotten less time using our proposed ADO algorithm contrasted with all other algorithms.

Fig. 5
figure 5

Comparative analysis for time

Full size image

Figure 6, explains the cost analysis compared with various optimization algorithms. Different workflow models have been taken for this analysis. We have achieved low cost in our proposed ADO algorithm. Compared with the existing algorithm our method gives the best solutions.

Fig. 6
figure 6

Comparative analysis of Cost

Full size image

Figure 7, demonstrates the failure coefficient analysis for reliability constraint. In this, we have considered the failure coefficient 0.001, 0.002, and 0.003 compared with existing algorithms. The X-axis represents various tasks analyzed with failure coefficient and the y-axis represents the reliability constraint for our proposed and existing algorithms. Compared to existing techniques, ADO gives better reliability.

Fig. 7
figure 7

Failure coefficient

Full size image

7 Conclusion

In this paper, we have proposed a multi-objective scheduling task taken for scientific workflow in a multi-cloud environment. Here, our main aim is to reduce the multi-objective parameters like reliability, makespan time, and cost. It is obtained by utilizing the fitness value of the algorithm our proposed algorithm. Further, the better execution of the ADO algorithm gives the least time, cost, and reliability as well as gives optimal value. The execution of the proposed scientific workflow scheduling strategy was analyzed and the test comes about to demonstrate that the proposed scheduling procedure has accomplished high exactness and effectiveness than the existing methods. Compared with existing algorithms like FA and GWO, our proposed algorithm gives better performance. In future work we will take on new algorithms on another platform also we will analyze faults that occurred during transmitting data which affects the data path.