Cloud computing environment provides several on-demand services and resource sharing for clients. Business processes are managed using the workflow technology over the cloud, which represents one of the challenges in using the resources in an efficient manner due to the dependencies between the tasks. In this paper, a Hybrid GA-PSO algorithm is proposed to allocate tasks to the resources efficiently. The Hybrid GA-PSO algorithm aims to reduce the makespan and the cost and balance the load of the dependent tasks over the heterogonous resources in cloud computing environments. The experiment results show that the GA-PSO algorithm decreases the total execution time of the workflow tasks, in comparison with GA, PSO, HSGA, WSGA, and MTCT algorithms. Furthermore, it reduces the execution cost. In addition, it improves the load balancing of the workflow application over the available resources. Finally, the obtained results also proved that the proposed algorithm converges to optimal solutions faster and with higher quality compared to other algorithms.
1. Introduction
The needs for computing and huge storage resources are fast growing. Therefore, cloud computing gets the attention due to the high performance computing services and facilities that are provided to the users as Software as a Service (SaaS), Infrastructure as a Service (IaaS), and Platform as a Service (PaaS) [1–3]. Various applications can be modeled as workflow applications of a set of tasks with dependencies between them in the sense that before one task can execute, dependant tasks have to complete their execution first. Workflow applications are being used in a range of domains, such as astrophysics, bioinformatics, and disaster modeling and prediction. Moreover, complicated problems like complex scientific applications are emerging recently through combining various methods and techniques in a single solution. For such a need, this type of applications has been executed on supercomputers, clusters, and grids [4]. Fortunately, with the advent of clouds, such workflow applications are executed in the cloud. The workflow applications are the mechanism of a large-scale business process execution, consisting of a set of events or tasks in which information is distributed from one task to another based on some technical rules, to achieve a general goal [5]. The workflow application tasks are dependent on each other, where the output of some tasks is the input to another. Therefore, the order of their execution must be considered when assigning the tasks to VM processors in a multiprocessor environment. Assigning the dependent tasks to the most appropriate VM processors is known to be an NP-complete problem as discussed by Verma and Kaushal [6]. The scheduling processes of the workflow applications are a multiobjective optimization problem (also known as Pareto optimization), where users might wish to minimize the money cost and the execution time for the whole workflow application with efficient load balancing over the VMs in the cloud environment. The optimal decision for the multiobjective workflow optimization is the trade-off between the three objectives; therefore, the objectives must be rated based on their importance to the users to select the best Pareto solutions because, for instance, minimizing the overall cost may lead to maximizing the execution time and the load over a specific VM [7, 8]. The workflow scheduling problem is an inherited problem from the heterogeneous computing environments, for which different research efforts were made to address the scheduling problem [9–11]. However, heterogeneous computing environments are not easy to set up, and their capability of giving more uniform performance with less failure is quite limited in comparison to the cloud environments [12, 13]. Moreover, the main objective of the various previous efforts in addressing the workflow scheduling problem in heterogeneous environments is to only minimize the finish time. Therefore, with the wide adoption of the cloud environments and their services as a pay-per-use model, there is a need to consider both the total monetary cost and the execution makespan. As a result, several metaheuristic algorithms were proposed to solve the scheduling problem of the workflow tasks and to get an efficient solution for tasks distribution over the different VMs in the cloud environment. For instance, Genetic Algorithm (GA) [14], Ant Colony Optimization [15], Swarm Intelligence [16], and Artificial Bee Colony (ABC) [17] are few examples of the various proposed solutions of workflow scheduling problem that addresses the total monetary cost and the execution makespan.
The main objective of this paper is to propose an algorithm that addresses the workflow scheduling problem. The proposed algorithm should also reduce the total makespan execution time and balances the load over the VMs with minimum total monetary cost. Therefore, this paper proposes a Hybrid GA-PSO algorithm through combining the strengths of both algorithms to address the workflow scheduling problem. The efficiency of the proposed algorithm is evaluated against other algorithms to prove its effectiveness in solving the workflow scheduling problem in the cloud environment.
The remainder of the paper is organized as follows. The problem description and the state-of-the-art in workflow scheduling are described along with the challenges when applying the existing common scheduling algorithms on IaaS platforms which are also highlighted in Section 2. This is followed in Section 3 by the design of the workflow scheduling algorithm and definitions of the proposed algorithm. Section 4 provides details of the performance evaluation of the multiobjective scheduling problem in cloud along with the experimental results and their discussion, and the paper is concluded and the future work is summarized in Section 5.
2. Related Work
Workflow scheduling problems are considered one of the main challenges in cloud environments. Many heuristic algorithms were proposed to solve the tasks scheduling problem using different strategies. However, the problem becomes obvious when the tasks are dependent on each other (i.e., workflow application). The dependent tasks require a specific execution order due to the relationship between them. There are two types of workflow scheduling: the best-effort workflow scheduling and the quality of services (QoS) constraint workflow scheduling [5, 18]. However, the best-effort workflow scheduling focuses on reducing the execution time of the whole workflow tasks regardless of other factors. Many types of research were based on the best-effort workflow scheduling to reduce the execution time, such as Braun et al. [16] who use the min-min algorithm for workflow scheduling. Their proposed approach executes the small tasks first and delays the larger tasks for a longer time. On the other hand, Mao et al. [19] use the max–min algorithm for task scheduling to execute the large tasks first and the small tasks are delayed for a longer time. In an attempt to resolve the aforementioned issues, Kumar and Verma [20] combined the min-min and max–min algorithms along with the Genetic Algorithm to improve the scheduling of multiple jobs over multiple virtual machines efficiently. The authors employ the min-min and the max–min algorithms to generate the GA individual and to provide better initial population rather than randomly chosen initial population. The achieved results were better than GA-based algorithms; however, it requires a lot of computation steps that consume time. This makes it unsuitable for cloud computing pay-per-use models. Guo et al. [21] proposed a Particle Swarm Optimization (PSO) based algorithm for solving the task-scheduling problem with an objective of reducing the total execution and transfer time. The optimization process is based on a heuristic scheduling combined with the PSO, to allocate the tasks to the different available resources. They practically proved that the PSO could run faster and give a better solution than GA. However, the PSO algorithm might get trapped in the a local optimal solution [22].
Different types of research, based on the QoS constraint for workflow scheduling, were considered to reduce the execution time under different predefined constraints, such as the following: user’s predefined budget constraints, user predefined deadline constraints, or workflow scheduling considering the reliability, time, cost, load balance, and fault recovery constraints. In this regard, Pandey et al. [23] presented a heuristic algorithm based on Particle Swarm Optimization to solve the workflow tasks scheduling over cloud resources. The conducted experiment shows that the computation cost using the PSO algorithm is three times better than the “Best Resource Selection” (BRS) algorithm under user predefined time constraints. However, the obtained result was not completely accurate due to the fast convergence towards the solution, which may cause PSO to get stuck in the local optimal solutions, and even the results cannot reflect the real performance of PSO. Arabnejad and Barbosa [24] presented a Heterogeneous Budget-Constrained Scheduling (HBCS) algorithm. The algorithm computes two possible schedules for the DAG (Directed Acyclic Graph) of the workflow. One schedule produces the minimum execution time with the maximum cost, while the other produces the minimum cost. The user, therefore, is able to decide which schedule to use to execute his task before the required deadline and within the cost range. The HBCS algorithm reduces the makespan by 30% and the cost within the user’s specified budget constraint. Furthermore, it reduces the time complexity compared to other budget-constrained algorithms.
Researchers such as Verma and Kaushal [6] realize that the priority of the tasks determined their execution order. Consequently, they presented a Bicriteria Priority Based Particle Swarm Optimization (BPSO) algorithm, to schedule the workflow tasks over the available cloud resources. The BPSO algorithm represents the trade-off between the execution time and the execution cost under the user’s predefined budget and deadline constraints. The proposed scheduling algorithm significantly reduces the execution cost and the makespan through selecting the best-known scheduling solution from the heuristic solutions under the predefined deadline and budget constraints compared to BHEFT (Budget-constrained Heterogeneous Earliest Finish Time) [31] and PSO algorithms [22, 26]. However, the BPSO algorithm does not consider the various loads of the available resources. Consequently, Xu et al. [25] developed a multiobjective heuristic algorithm based on the min-min algorithm. The proposed algorithm uses four real-world scientific workflows to evaluate its performance. The conducted experiments evaluate the performance of the makespan and the execution cost with fault recovery procedure. The heuristic algorithm, based on the min-min algorithm, is considered a better choice only when both the cost and the makespan are considered.
The multiobjective optimization is a very promising direction to tackle the problem of workflow scheduling. In this regard, Ge and Wei [27] used a Genetic Algorithm to optimize the tasks scheduling in the job queue. They used a centralized scheduler (i.e., master node) to distribute the waiting tasks to the different available resources (i.e., slave nodes) based on the resources status messages. Their results show that the proposed schedule was better than the First-In-First-Out (FIFO) and the Delay scheduling that distributes the load over all resources in the cloud. However, the proposed algorithm requires a lot of processing time to reach the optimal solution. Later, Fard et al. [28] suggested a heuristic static multiobjective scheduling algorithm for scientific workflows in heterogeneous environments. The proposed algorithm adopted the strategy of maximizing and minimizing the distance between the constraints for each of the four objectives (i.e., makespan, economic cost, energy consumption, and reliability). The researchers analyzed and categorized the different objectives based on their impact on the optimization process. The results showed that most of the generated solutions are within the predefined deadline and budget constraints. However, the proposed algorithm is not efficient with a small number of tasks and processors. Wu et al. [29], therefore, suggested a Revised Discrete Particle Swarm Optimization (RDPSO) algorithm to schedule the workflow applications over the different available resources. The experiments were conducted over a set of workflow applications with different data communication and computation costs. The result showed that the proposed RDPSO algorithm reduces the cost and yields better makespan compared to the standard PSO and BRS (Best Resource Selection) algorithm. However, the proposed algorithm is not efficient with large search space. Continuously, Chitra et al. [26] proposed a local minima jump solution using PSO (i.e., JPSO) for workflow scheduling in the cloud to schedule the tasks and load balance the workflow applications, to reduce the makespan. The JPSO algorithm overcomes getting trapped in the local minimal solution problem through making a jump in the gbest value to avoid the poor convergence of the gbest values. The results show that the proposed algorithm is more efficient compared to the GA algorithm by 3.8% with a small number of tasks. However, the GA algorithm shows the better result with a large number of tasks.
Many researchers attempted to solve the multiobjective optimization problem of the workflow applications using a different number of objectives. In this paper, a Hybrid GA-PSO algorithm is proposed to schedule the workflow tasks over the available resources. The proposed algorithm aims to achieve three objectives: reducing the makespan, reducing the cost, and balancing the load of the workflow tasks on heterogeneous VMs in the selected cloud DC. In summary, the GA-based algorithms provide better results than other algorithms when the number of iterations is large. However, increasing the number of iterations means that the GA algorithm will consume more time to reach the optimal solution. On the other hand, the PSO-based algorithms provide better results than the other algorithms and in less time. However, the results may not be accurate due to the fast convergence of the PSO-based algorithms to the solution, which may cause being stuck in the local optimal solution. Therefore, the proposed Hybrid GA-PSO algorithm is distinguished by the characteristics of the GA and the PSO algorithms. The Hybrid GA-PSO algorithm is expected to work faster with different sizes of workflow applications compared to other algorithms with the same objectives. Moreover, the Hybrid GA-PSO algorithm may not get trapped in the local optimal solution, because of the use of the GA mutation operator that enhances the accuracy of the solutions. Table 1 summarizes the review of the literature works along with their pros and cons.
Literature review summary.
Author
Name of Algorithm
Objective
Advantages
Limitation
Braun et al. [16]
min-min algorithm
Time
12% better than GA
Delayed large tasks for long time
Kumar and Verma [20]
Combination of min-min and max–min strategies in Genetic Algorithm
Time
Faster than the GA
Time consuming
Guo et al. [21]
Particle Swarm Optimization (PSO) algorithm
Execution and transfer time
Faster than the M-PSO and L-PSO algorithms in a large scale
Stuck in local optimal solution
Pandey et al. [23]
Heuristic algorithm based on particle swarm optimization
Time and cost
Three times better cost compared to BRS, good load distribution over resources
Reduction of 30% in execution time whilemaintaining the same budget
Not considering the load over resources
Verma and Kaushal [6]
Bicriteria Priority Based Particle Swarm Optimization (BPSO) algorithm
Time and execution cost
Decreasing the execution cost compared to BHEFT and PSO
Not considering the load over resources
Xu et al. [25]
Heuristic algorithm based on the min-min algorithm
The fault recovery, the time, and the cost
Fault recovery has a significant impact on the two performance criteria
Better choice only when both cost and makespan are considered
Chitra et al. [26]
The PSO algorithm
Load balance and the makespan
Better than GA and PSO
Time consuming
Ge and Wei [27]
The Genetic Algorithm
Load balance and makespan
Better than FIFO
Time consuming to reach to optimal solution
Fard et al. [28]
The heuristic algorithm
Makespan, economic cost, energy consumption, and reliability
Improve all four objectives
Not efficient with small number of tasks and processors
Wu et al. [29]
The Revised Discrete Particle Swarm Optimization (RDPSO) algorithm
Makespan, communication costs, and computation costs
Better than the standard PSO and BRS (Best Resource Selection) algorithm
Not efficient with large search space
The proposed algorithm
Genetic and particle swarm optimization algorithm
Makespan, communication costs, load balance, and execution and transfer time
Faster convergence to the solution in comparison with other approaches
Supports one data center without considering the dynamic workflow
3. The Proposed Algorithm
Many researchers used random workflows graph or real-world workflows graph to represent the workflow applications using the Pegasus framework [7]. Pegasus framework provides the DAG of different real workflow applications and defines the number of the workflow tasks, the sizes of transmission data between the tasks, and the execution time of each task. These workflows will be used for measuring the performance of the proposed Hybrid GA-PSO algorithms. There are five real workflow applications that are used in the scientific domains, namely, Montage [32], CyberShake [33], Epigenomics [34], LIGO Inspiral Analysis Workflow [35], and SIPHT [36], as portrayed in Figure 1. The Montage application created by NASA/IPAC closes together multiple input images to form custom mosaics of the sky [32]. The CyberShake workflow is used by the Southern California Earthquake Center to distinguish earthquake threatening a region [33]. The Epigenomics workflow created by the USC Epigenome Center and the Pegasus framework is used to automate the different operations in genome sequence processing [34]. LIGO’s Inspiral Analysis workflow is used to create and analyze gravitational waveforms from data gathered during the coalescing of compact binary systems [35]. The SIPHT workflow, from the bioinformatics project at Harvard, is used to automate the search for small untranslated RNAs (sRNAs) for bacterial replicons in the NCBI database [36].
Known scientific workflows.
Montage
CyberShake
Epigenomics
SIPHT
LIGO
The main steps of the GA-PSO algorithm are shown in Figure 2. The GA-PSO algorithm starts with generating a random population and defines a specific number of iterations as a parameter to the algorithm. The population represents several solutions to the workflow tasks problem and each solution is a distribution of the whole workflow tasks over the available VMs. The initialized population is passed through the GA algorithm with the first half of the defined iterations; that is, if the number of the iterations is (n), then the GA algorithm will be repeated (n/2) times. The reason behind using (n/2) iteration is to reduce the complexity of the proposed algorithm, as the performance of the GA algorithm depends mainly on the method used to encode solutions into chromosomes and particles and what the fitness function is measuring, as well as the size of the population, that is, the number of iterations, as proved by Alajmi and Wright [37]. These parameter values can be adjusted after evaluating the algorithm’s performance on a few trial runs. Through experiments, the GA-PSO algorithm’s performance was the best, when the defined number of iterations is divided equally between the GA and PSO algorithms. This also agrees with the concept of the divide and conquer that divides one problem into two subproblems to produce a complexity equal to T(n)=2xT(n/2)+F(n), where F(n) is the divide and conquer time. The solution for such equation depends on F(n) and the complexity is O(F(n)) if F(n)≥n (by master theory). Moreover, it is also known that both the GA and the PSO require many function evaluations because each needs to evaluate the objective of every population member in the current sample. Therefore, decreasing the population size in a GA or PSO (i.e., decreasing the number of iterations) is a common practice to avoid degrading the GA or PSO performance in terms of the accurate results and reduction rate.
The flowchart of the GA-PSO algorithm [30].
In the GA algorithm, the solutions are called chromosomes; the chromosomes are enhanced gradually at each iteration through the GA operators (i.e., selection, crossover, and mutation). The resulting chromosomes are passed to the PSO algorithm at the second half of the defined iterations. In the PSO algorithm, the chromosomes are called particles; the particles are enhanced gradually at each iteration through the PSO algorithm. The particle with the minimum fitness value is selected to represent the solution of the workflow task problem.
3.1. Initializing Population
The Hybrid GA-PSO algorithm is initialized to a specific number of iterations. A solution is initiated randomly at the first iteration. After the first iteration, a sequence of new populations are created and recursively enhanced using the previous solutions to form a set of suggested solutions as illustrated in Figure 3. The population in the GA algorithm is called chromosome. The length of the chromosomes is equal to the number of the workflow tasks, and the genes of each chromosome represent the different VMs. The randomly generated chromosomes represent the input to the proposed GA-PSO algorithm. The GA algorithm represents the first part of the proposed GA-PSO algorithm which will be used to generate different solutions to the workflow scheduling problem.
An example of a randomly initiated population.
3.2. Applying the GA Algorithm
At the first phase, the GA is applied to the whole generated population for (n/2) of the determined iterations, to generate the optimal solution from existing solutions, which is required to solve the scheduling problem. The PSO is applied to the whole generated population for the following (n/2) of the determined iterations, which is generated by the GA algorithm. The PSO algorithm keeps in memory the best and the worst solutions, which can be useful for the fast solution convergence when GA generates bad solutions. The solutions in GA that define the scheduling solution of our problem are represented by several chromosomes with length equal to the number of the whole workflow tasks. Each chromosome consists of several genes representing the hosts’ VMs. In each iteration, the GA passes the chromosomes between three different operators: the selection, crossover, and mutation operator. The first operator of the GA algorithm is the selection operator that is used to select different solutions from the existing chromosomes and used to build the next generation of the chromosomes for the next iteration.
3.2.1. Selection Operator
In the GA algorithm, not all generated chromosomes are evolved through the GA operators in each iteration. Therefore, the chromosomes are passed through the tournament selection to select the best chromosome from a group of chromosomes. The function selects a random (id) after running several tournaments between few chromosomes. The selected ids represent the index of the selected chromosome from a set of chromosomes. The best chromosome in the group is selected for crossover operator based on its fitness value as shown in Figure 4 and Algorithm 1.
id = Math.random()∗ chromosome.size() // select chromosome randomly
tournament[i] = get_chromosome (id)
End For
fitness ← tournament.getFitnest() //return the fitness value in tournament group.
The schema of the fittest value.
3.2.2. The Crossover Operator
The crossover operator aims to generate new chromosomes through changing the position of the genes inside every two chromosomes. In the crossover, a random number is selected in the range of the number of the chromosome genes, to represent the division point of each chromosome into two parts. The crossover returns an offspring chromosome of two parts that contains both chromosomes genes, that is, VMs. The first group of VMs takes the first chromosome until the index, which is determined by the random number. The second chromosome has the second group of the VMs starting from the index, which is determined by the random number, until the end of the chromosome. The implementation of the crossover method is illustrated in Algorithm 2.
<bold>Algorithm 2: </bold>The crossover method.
Input: two chromosomes
Output: offspring chromosome
r = (Math.random()∗ chromosome.length)
For i=0, j=0 to r
offspring _chromosome[j] = chromosome1[i]
End For
For i=r to chromosome.length
offspring_chromosome[j] = chromosome2[r]
End For
3.2.3. The Mutation Operator
The mutation operator aims to make unusual modifications in the new chromosomes that are generated from the previous crossover operator with better fitness value than the existing chromosomes. The mutation operator operates over the returned chromosome from the selection method, and the occurrence of the mutation is based on the mutation rate variable. The mutation process starts with a number that is randomly generated to be less than or equal to the mutation rate. Two genes, that is, VMs, are selected randomly from the same chromosome and checked to be different. If they are the same, their places are swapped to generate new chromosome, which represents a different distribution of the tasks over the available VMs. The generated chromosome is then passed to the next stage of the algorithm. The implementation of the mutation method is illustrated in Algorithm 3.
<bold>Algorithm 3: </bold>The mutation method.
Input: offspring_chromosome //returned from crossover operator
Output: Newchromosome
Set mutationRate = 0.5
If (Math.random()≤ mutationRate)
t1= Rand 0,1∗ offspring_chromosome.length // select a random number t1
t2= Rand 0,1∗ offspring_chromosome.length // select a random number t2
If offspring_chromosome [t1] !=offspring_chromosome[t2]
The solutions that are returned from the GA algorithm are fed into the PSO algorithm with the rest of the determined iterations, to find the optimal solution from the GA generated solutions. In the PSO algorithm, the solutions are called particles, the individuals of each particle represent the VMs of the DC, and the index of each VM represents a workflow task. The PSO algorithm consists of three stages as follows.
3.3.1. Evolve <inline-formula><mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" id="M46"><mml:mo stretchy="false">(</mml:mo><mml:msub><mml:mrow><mml:mi mathvariant="bold">g</mml:mi></mml:mrow><mml:mrow><mml:mi mathvariant="normal">b</mml:mi><mml:mi mathvariant="normal">e</mml:mi><mml:mi mathvariant="normal">s</mml:mi><mml:mi mathvariant="normal">t</mml:mi></mml:mrow></mml:msub><mml:mo mathvariant="bold">)</mml:mo></mml:math></inline-formula> and <inline-formula><mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" id="M47"><mml:msub><mml:mrow><mml:mo mathvariant="bold">(</mml:mo><mml:mi mathvariant="bold">p</mml:mi></mml:mrow><mml:mrow><mml:mi mathvariant="normal">b</mml:mi><mml:mi mathvariant="normal">e</mml:mi><mml:mi mathvariant="normal">s</mml:mi><mml:mi mathvariant="normal">t</mml:mi></mml:mrow></mml:msub><mml:mo mathvariant="bold">)</mml:mo></mml:math></inline-formula> of the Particles
In each iteration, a new generation of the particles is produced based on their velocity and position in the previous iteration. The changes in the velocity and position of the particles are based on the values of (gbest) and (pbest) which are evolved in each iteration. The implementation of evolving (gbest) and (pbest) values of the particles is illustrated in Algorithm 4.
The progress of the particles in the PSO algorithm is based on the values of (gbest) and (pbest) that keep changing with each iteration. At the first iteration, pbest[k] is equal to the solutions that are generated from the GA algorithm, where k is to distinguish each solution from the other. The (gbest) is equal to the solution with the smallest fitness value. Furthermore, at each iteration, the comparison between the previously generated particles and the newly generated particles is taking place based on the fitness value. The particle with the best fitness value is stored in (pbest). The (gbest) store the best particle from the whole generation of the particles at each iteration by comparing their fitness value and the one with the best value in (Pbest). At each iteration, the comparison process ensures that all particles are moving towards the best solution, to reach the optimal solution.
3.3.2. Update the Velocity and Position Matrix
After generating the initial particles velocity and position values randomly and calculating both (pbest) and (gbest), the velocity of each particle, in each iteration, is updated accordingly. The implementation of the update process of the velocity matrix is illustrated in Algorithm 5.
<bold>Algorithm 5: </bold>Update the velocity matrix.
Input: velocity values
Output: updated velocity values
Setk=0; c1=1; c2=1.1; r1,r2=rand[0,1]; //k is the index of the particles individuals
While not reach max particles.length do
If Particle[k] == pbest[k]
velocity[k] −= c1∗r1; //r1,r2 are random numbers
Else
velocity[k] += c1∗r1; //c1,c2 are acceleration coefficient
End If
If Particle[k] == gbest[k]
velocity[k] −= c2∗r2;
Else
velocity[k] +=c2∗r2;
End If
k=k +1;
Repeat //until the last particle
The process of updating the velocity of the particles aims to generate a new generation from the different VMs locations that have better fitness value than the previous one. Each individual in the particles is compared with its (pbest) value, which was generated at the previous iteration. The velocity value for each individual is decreased when both individuals in (pbest) and the particle are equal; otherwise, the velocity value is increased. Similarly, a comparison of each individual in the particles and their gbest values, from the previous iteration, is taking place. The velocity value for each individual is decreased when both individuals in gbest and the particle are equal; otherwise, the velocity value is increased. Accordingly, the position of the VMs of each particle is changed based on the updated values of the velocity. The implementation of the update position matrix is illustrated in Algorithm 6.
<bold>Algorithm 6: </bold>Update the position matrix.
Input: updated velocity values
Output: updated particles position
Setj=0; // j is the index of the particle individuals
Two VMs that have the maximum velocity values are swapped within each particle of all particles within the produced population. The termination criteria of the GA-PSO algorithm are represented by reaching the maximum number of iterations. When the termination criteria are satisfied, the solution that has the smallest fitness value within the population, which was generated at the last iteration, is presented as the scheduling solution of the workflow application. Otherwise, the (gbest) and (pbest) values evolve repeatedly until the termination condition is achieved. The complete GA-PSO algorithm is illustrated in Algorithm 7.
<bold>Algorithm 7: </bold>The proposed algorithm.
Input: workflow WN,E and set of resources VM1,VM2,VM3,…,VMj
Output: gbest// the best solution to allocate W over VMj
For i=0 to p
population ← randomize() // initialize population, P is the population size
End For
While not Reach n/2 do // n is number of iterations
SetNewchromosomejasparticlej // j is the index of the particles
Initialize particles position and velocity randomly
Calculate the (gbest) and (pbest) values
While not Reach ndo
velocitymatrix←updateparticlejvelocity
positionmatrix←updateparticlejposition
Repeat
The algorithm is bounded by the GA operations (i.e., mutation, crossover, and selection). However, calculating the complexity of the GA or PSO algorithms is unlikely to be useful and worse probably deceptive. Moreover, because of the complexity (i.e., NP-complete) of the workflow scheduling problem, it is very challenging to develop an optimized workflow scheduling algorithm for workflow tasks distribution to the available resources within a reasonable overhead, that is, CPU time. However, since the main goal of the proposed scheduling algorithm is to optimize the overall cost (i.e., may not be optimal), it is, therefore, a practical trade-off between the overhead of the task-scheduling algorithm and the optimization on the running cost of the data center. Therefore, as will be demonstrated in our simulation experiments, we will evaluate the time complexity by measuring and averaging the runtime with a different number of tasks, as discussed in Section 4.2.
4. Performance Evaluation
For the purpose of evaluating the proposed algorithm, the proposed GA-PSO algorithm was implemented using the WorkflowSim [38]. The WorkflowSim extends the existing CloudSim simulator [39] by providing a higher layer of workflow management, through providing a suitable environment for applying different scheduling algorithms. Furthermore, to evaluate the performance of the proposed GA-PSO algorithm, the obtained results of the proposed GA-PSO algorithm have been compared with existing work scheduling algorithms, such as GA proposed in [27] and PSO proposed in [21]. In addition, the performance of the proposed GA-PSO algorithm was also compared with other related works, as discussed in Section 4.3.
4.1. Environment Setup
To evaluate the impact of the proposed algorithm on the workflow scheduling problem in comparison with other algorithms, we ran extensive experiments on real workflow applications using the simulation parameters in Table 2.
Simulation parameters.
Parameter
Value
Number of tasks in application
25–1000
The number of VMs
16
MIPS
250–1500
RAM
256–1024 (MB)
BW
250–1500 (mbps)
Processor speed
10,000
Number of processors
4
VM policy
TIME_SHARED
These parameters were used to identify the characteristics of the VMs and the workflow applications in the experiments. A real workflow application—Montage workflows application—was created with different numbers of tasks to evaluate three objectives: (1) reducing the makespan of the application, (2) optimizing the processing cost, and (3) balancing the load on the different resources with respect to the different heterogeneous resources characteristics. The parameters defined in Table 3 were used throughout the GA-PSO evaluation experiments.
GA-PSO algorithm parameters.
Parameter
Value
Population size
100
Mutation rate
0.05
Crossover
Single point
Number of iterations
100
Number of executions
500
C1
1
C2
1.1
r1, r2
[0,1]
α1, α2
0.4
α3
0.2
The algorithm starts with 100 random solutions, called population. The single point crossover method was chosen in the GA phase (Section 3.2). The mutation operator rate was defined as 0.05 in the mutation stage. In the PSO algorithm phase (Section 3.3), the acceleration coefficients (C1)and(C2), as well as the random number which is used in the update velocity and position equation, were also defined, as in Table 3. Furthermore, the degree of importance of each objective in the fitness function was defined as “α1,α2 = 0.4 and α3=0.2” for the makespan, execution cost, and the load balance rate, respectively. The Montage workflow application is used in the evaluation with a different number of tasks (25–1000), to enlarge the size of the workflow and evaluate the algorithm under these different cases. The number of iterations for the GA-PSO algorithm was defined to 100 iterations to reach the optimal solution. The experiments were repeated 500 times, and the average results were compared with other algorithms. Four experiments were conducted based on the characteristic of the VMs, as in Table 2. The size of the workflow will be changed to examine the ability of the proposed algorithm in reducing the makespan, execution cost, and load balance for the small and large size workflow applications. For this purpose, four test scenarios using Montage workflow application with a different number of tasks and different number of edges and data sizes were used. The characteristics of the Montage workflow application that were used in the experiments are summarized in Table 4.
The characteristics of the Montage workflows.
Scenarios
Number of tasks
Number of edges
Average data size (MB)
Scenario One
25
95
3.43
Scenario Two
50
206
3.36
Scenario Three
100
433
3.23
Scenario Four
1000
4485
3.21
4.2. Performance Analysis
All the four scenarios were executed to evaluate the reduction in the makespan, the execution cost, and the load balance using the proposed GA-PSO algorithm in comparison with the GA and PSO algorithms. The results of the executed experiments for the four scenarios are reported in Table 5.
The result of the executed experiments.
Algorithm
Makespan (sec)
Execution cost ($)
Load balance (rate)
Scenario One
GA-PSO
95.09
16.85
9.76
GA
197.65
52.68
52.58
PSO
101.21
18.16
21.33
Scenario Two
GA-PSO
116.01
49.89
13.81
GA
250.89
86.34
61.93
PSO
155.31
62.86
18.23
Scenario Three
GA-PSO
233.78
127.74
33.03
GA
345.72
137.09
49.2
PSO
253.44
133.55
41.82
Scenario Four
GA-PSO
1585.6
1021.42
73.83
GA
2402.28
1529.23
134.67
PSO
1802.31
1200.41
90.15
For each scenario, the number of tasks in the Montage workflow was increased. Scenario One, for instance, represents a small search space that makes the process of reaching the optimal solution fast and straightforward (i.e., the simplest case). On the other hand, Scenario Four represents a large number of tasks in the Montage workflow to expand the search space (i.e., the worst case). The large search space makes the process of finding the optimal solution a challenging task for the optimization algorithm. The results in Table 5 show minor differences in the makespan, the execution cost, and the load balance between the GA and PSO algorithms with Scenario One. However, there is a slight improvement for the GA-PSO algorithm compared with GA and PSO algorithms. On the other hand, the results show significant differences for the GA-PSO algorithm compared with the GA algorithm with Scenario Two and Three. These significant differences could be due to the unnecessary diversity caused by an inappropriate mutation rate. There is also a slight difference in the result between the GA-PSO and the PSO algorithm with Scenarios Two and Three as well. This slight difference is due to the fact that the GA-PSO algorithm depends mainly on the PSO algorithm in converging the solutions towards the optimal solution. In Scenario Four, the large number of tasks expands the search space to represent the worst case scenario. The GA-PSO algorithm still achieves a better result compared to the GA algorithm. This result is due to the fast solution convergence that avoids the unnecessary diversity of the solutions. In addition, GA-PSO algorithm showed a significant enhancement compared to the PSO algorithm, because the PSO algorithm normally gets trapped in the local optimal solution. The above experiment was repeated several times, and the average results in terms of makespan, execution cost, and the load balance for the proposed GA-PSO, GA, and PSO algorithms with the four scenarios were calculated and consolidated in Table 6. The results in Table 6 also demonstrate the proposed algorithm ability in resolving the workflow task-scheduling problem in comparison with the GA and the PSO algorithms.
Average results in makespan, execution cost, and load balance for the different algorithms.
Methods
Avg.makespan
Avg.execution cost
Avg.load balance
Hybrid GA-PSO
507.62
303.975
32.6075
GA
799.135
451.335
74.595
PSO
578.0675
353.745
42.8825
When comparing the improvement in the makespan using the proposed GA-PSO algorithm with the GA and PSO algorithm, one can notice that the GA-PSO algorithm achieves a significant enhancement of 16% better than the GA algorithm and 4% better than the PSO algorithm as illustrated in Figure 5. This is because the proposed GA-PSO algorithm always chooses the most appropriate VMs to execute the tasks without focusing only on fast VMs, which actually may overload one VM over the other and slow down the overall execution of the workflow application (i.e., increase the execution time).
Comparison of the average makespan over different number tasks.
In terms of execution cost, Figure 6 shows that the proposed GA-PSO algorithm is 13% better than the GA algorithm and 4% better than the PSO algorithm.
Comparison of the average execution cost over different number tasks.
The improved result of the proposed GA-PSO algorithm is because the proposed algorithm chooses VMs to achieve a minimum execution cost to execute the selected tasks. Finally, the proposed GA-PSO algorithm balances the load over the resources compared with GA and PSO algorithms as shown in Figure 7.
Comparison of the average load balancing rate over different number of tasks.
The average result of the load balancing obtained by the proposed GA-PSO algorithm is better than the GA algorithm by 28%, and the load balance was reduced by 4% compared to the PSO algorithm. This result is because the proposed GA-PSO algorithm converges to the solutions in a better way using the GA algorithm with avoiding the unnecessary diversity that may degrade the quality of the algorithm.
Finally, the CPU time is defined as the average running time of the proposed GA-PSO algorithm in comparison with the GA and the PSO algorithms running on hardware of the characteristics defined in Table 2. The result of the average running time for each algorithm using a different number of tasks is consolidated in Table 7. It can be noticed that the GA algorithm consumes more CPU time compared to the other algorithms. When the workflow size increases, the CPU time of the GA and PSO is also increased. For instance, with 3000 tasks, the GA took about 28.3 seconds and the PSO consumed 23.9 seconds while the proposed algorithm only took 22.4 seconds to reach the final solution.
The running time of the executed algorithms in seconds.
Method/number of Tasks
25 tasks
50 tasks
100 tasks
1000 tasks
2000 tasks
3000 tasks
GA
0.869465
0.888796
1.093582
21.321738
23.4338
28.28996
PSO
0.761534
0.871797
1.037802
18.515041
18.65318
23.99583
Hybrid GA-PSO
0.764333
0.873796
1.025266
17.576722
18.00189
22.44825
The increase in the CPU time is actually because of the (gbest) and (pbest) update process. The update process calculates the fitness value for every particle with calculating the makespan and the cost at the same time. This is true because the proposed GA-PSO algorithm builds the solutions task by task, and hence the CPU time increases as the workflow size increases.
4.3. Comparison of Related Approaches
For the comparison purposes, three algorithms were evaluated for workflow tasks scheduling, namely, HSGA algorithm proposed in [40], WSGA algorithm proposed in [41], and MTCT algorithm proposed in [25] with the proposed GA-PSO algorithm. The comparison was carried out over two objectives: the makespan and the execution cost. The reason behind the selected objectives is that WSGA and MTCT algorithms optimize only the makespan and the execution cost, while HSGA optimizes only the load balancing and the makespan. The algorithms were implemented according to their description in the literature. The results show that the proposed GA-PSO algorithm converges to the optimal solution faster than the other algorithms and with higher quality in terms of load balancing as discussed in Section 4.2. All performance analyses were carried out over a workflow with different numbers of tasks, 25, 50, and 100 along with specific parameters, as defined in Table 8. The size of tasks, price, and the speed of resources are generated randomly to simulate a heterogeneous environment.
GA-PSO versus HSGA simulation parameters.
Parameter
Value
Number of tasks in application
20–100
Task lengths
12–72 (×105 MI)
Number of resources
30
Resource speeds
500–1000 (MIPS)
Bandwidth between resources
10–100 (mbps)
The workflow application was evaluated with a different number of tasks, to illustrate the impact of the proposed GA-PSO on the makespan and the load balancing rate in comparison with the HSGA algorithm. Table 9 and Figure 8 illustrate the average results of the experiment.
GA-PSO versus HSGA experiment results.
Methods
Avg. makespan
Avg. load balance
GA-PSO
28191.96
2.23
HSGA
35000
2.63
GA-PSO versus HSGA analysis of results.
The result illustrated in Figure 8 shows that the proposed GA-PSO algorithm is able to solve the workflow problem with better makespan and load balancing than the HSGA by 11% and 9%, respectively. The improvement of the proposed GA-PSO algorithm is due to the fast convergence to the solution, as an advantage of employing the PSO algorithm, which avoids the unnecessary diversity that may occur in the HSGA algorithm and leads to reaching the best solution. Similarly, we compared the proposed GA-PSO algorithm with WSGA algorithm based on the simulation parameters in Table 10.
GA-PSO versus WSGA simulation parameters.
Parameter
Value
Population size
20
Selection method
Roulette wheel
Crossover method
Single point crossover
Mutation rate
0.1
The number of resources
3–14
Number of tasks in application
50–100
Number of iterations
200
The different values of the workflow size and the resources configurations illustrate the impact of the proposed GA-PSO and the WSGA algorithms on the makespan and the execution cost. The average results of the experiment are shown in Table 11 and Figure 9.
GA-PSO versus WSGA experiment results.
Methods
Avg. makespan
Avg. execution cost
Hybrid GA-PSO
84.875
5.195
WSGA
93
7.695
GA-PSO versus HSGA analysis of results.
The proposed GA-PSO algorithm obtained a solution for the workflow problem by 5% better value for the makespan and 9% better value for the execution cost in comparison with WSGA. It is worth mentioning that both the proposed GA-PSO and the WSGA algorithms are based on GA technique. However, the proposed GA-PSO algorithm uses the PSO algorithm to avoid the unnecessary diversity in the solution and enhances the obtained solutions, which might be scattered due to GA technique. Finally, the proposed GA-PSO algorithm was also compared with the MTCT algorithm, based on the simulation parameters in Table 12.
GA-PSO versus MTCT simulation parameters.
Parameter
Value
The number of resources
20
Resource speeds
500–1000 (MIPS)
Bandwidth between resources
20 (mbps)
For the evaluation purposes, four different types of workflow applications were used to show the impact on the makespan and the execution cost of the proposed GA-PSO and the MTCT algorithm. The details of the workflow applications are illustrated in Table 13.
Workflows details.
Workflow
The number of tasks in different workflow sizes
Small
Medium
Large
XLarge
Montage
25
50
100
1000
CyberShake
30
50
100
1000
Epigenomics
24
46
100
1000
LIGO
30
50
100
977
The makespan and the execution cost results, of the proposed GA-PSO and the MTCT algorithms, with the four types of workflow applications, are summarized in Table 14 and Figure 10.
GA-PSO versus MTCT experiment results.
Methods
The makespan (sec)
The execution cost ($)
Montage
Hybrid GA-PSO
1.12
1.04
MTCT
1.4
1.4075
CyberShake
Hybrid GA-PSO
0.9875
1.12
MTCT
1.365
1.3725
Epigenomics
Hybrid GA-PSO
1.23
1.112
MTCT
1.3525
1.36
LIGO
Hybrid GA-PSO
1.1075
1.132
MTCT
1.4975
1.4225
GA-PSO versus MTCT analysis of results.
The obtained results of the GA-PSO and the MTCT algorithms show that the GA-PSO algorithms enhance the makespan by 11% with 15% less in execution cost, in comparison with the MTCT algorithm, using the Montage workflow, whereas the proposed GA-PSO algorithm achieved an improvement by 17% in terms of makespan and 11% less in the execution cost, compared with the MTCT algorithm, using the CyberShake workflow. Furthermore, the proposed GA-PSO algorithm schedules the Epigenomics workflow with 5% better makespan and 9% less execution cost than the MTCT algorithm. Finally, the results of the makespan and the execution cost of the LIGO workflow were better by 15% and 11% compared with the MTCT algorithm, respectively.
The results and the enhancements that were obtained by the proposed GA-PSO algorithm are because the proposed algorithm always selects the best solution for distributing the workflow tasks over the most suitable VMs regardless of the number of the workflow tasks. The proposed GA-PSO algorithm combines the suitable diversity and the fast convergence to optimal solutions, to find the optimal solution faster than any other algorithm.
5. Conclusion and Future Work
In this paper, a GA-PSO algorithm was proposed and implemented using the WorkflowSim simulator, for workflow task scheduling in cloud environments. The performance of the proposed algorithm was also compared with some known algorithms such as GA, PSO, HSGA, WSGA, and MTCT. The purpose of the proposed algorithm is to ensure a fair distribution of the workload among the available VMs, considering the order of the execution of the workflow tasks to reduce the makespan and the processing cost of the workflow applications in cloud computing environments. The GA-PSO algorithm selects the VMs to execute the workflow tasks in the minimum time based on the execution speed of the VMs and the size of the workflow tasks. The design of the GA-PSO algorithm tends to allow executing the tasks over the VMs with a balanced load distribution over the fast and slow VMs, without overloading some VMs over the others. This technique reduces the makespan through a fair utilization of the slow VMs instead of overloading the fast VMs and slowing down the overall execution of the tasks. The GA-PSO algorithm yields an optimal solution of the workflow task scheduling in terms of makespan compared with GA, PSO, HSGA, and WSGA algorithms by 16%, 4%, 11%, and 5%, respectively. In addition, the enhancements in the makespan using the Montage, CyberShake, Epigenomics, and LIGO workflow were averaged as 11%, 17%, 5%, and 15%, respectively, in comparison to MTCT algorithm. Moreover, the results prove that the GA-PSO algorithm minimizes the total execution cost of the workflow tasks compared to GA, PSO, and WSGA algorithms by 13%, 4%, and 9%, respectively. The GA-PSO algorithm also enhances the execution cost in comparison to MTCT algorithm using the Montage, CyberShake, Epigenomics, and LIGO workflow which are averaged at 15%, 11%, 9%, and 11%, respectively. The significance of the results, from the GA-PSO algorithm, are affected by the appropriate selection of the VM with a balance between cost and time through the fitness function of the GA-PSO algorithm. This goal was achieved by using the same weights for both the makespan and the execution cost in the fitness function. The proposed GA-PSO algorithm improves the load balancing of the workflow applications over the available resources, in contrast with GA, PSO, and HSGA algorithms, by allocating the tasks based on the VMs ability and the task sizes. The enhancements of the load balance in comparison with GA, PSO, and HSGA algorithms are averaged at 28%, 4%, and 9%, respectively. The design of the GA-PSO algorithm uses the standard deviation to select the best solution that keeps the variance of the distributed load, over the VMs, as low as possible taking into account the size of the tasks and the speed of each VM during the distribution of the tasks.
In the future, the work can be extended to more than one data center in a heterogeneous environment. Furthermore, the distribution of the workflow application can be extended into two levels: when workflow tasks reach the service broker and when the workflow tasks are distributed to the available VMs of each DC based on the size of the tasks and the speed of each VM. The justification can be verified over real-time cloud environment. In addition, the work can be improved through using dynamic workflow that allows more flexibility for the users to change the characteristics of the workflow tasks during the runtime.
Conflicts of Interest
The authors declare that there are no conflicts of interest regarding the publication of this article.
AljammalA. H.ManasrahA. M.AbdallahA. E.TahatN. M.A new architecture of cloud computing to enhance the load balancinggLiJ.LiuZ.ChenX.XhafaF.TanX.WongD. S.L-EncDB: A lightweight framework for privacy-preserving data queries in cloud computingManasrahA. M.SmadiT.ALmomaniA.A Variable Service Broker Routing Policy for data center selection in cloud analystGuptaB. B.AkhtarT.A survey on smart power grid: frameworks, tools, security issues, and solutionsYuJ.BuyyaR.RamamohanaraoK.Workflow scheduling algorithms for grid computingVermaA.KaushalS.Cost-Time Efficient Scheduling Plan for Executing Workflows in the CloudJiH.BaoW.ZhuX.Adaptive workflow scheduling for diverse objectives in cloud environmentsManasrahA. M.Dynamic weighted VM load balancing for cloud-analystChenW.-N.ZhangJ.An ant colony optimization approach to a grid workflow scheduling problem with various QoS requirementsTalukderA. K. M. K. A.KirleyM.BuyyaR.Multiobjective differential evolution for scheduling workflow applications on global GridsWieczorekM.HoheiselA.ProdanR.Towards a general model of the multi-criteria workflow scheduling on the gridLiP.LiJ.HuangZ.GaoC.-Z.ChenW.-B.ChenK.Privacy-preserving outsourced classification in cloud computingStergiouC.PsannisK. E.KimB.-G.GuptaB.Secure integration of IoT and Cloud ComputingDasguptaK.MandalB.DuttaP.MandalJ. K.DamS.A genetic algorithm (GA) based load balancing strategy for cloud computingZhangZ.ZhangX.A load balancing mechanism based on ant colony and complex network theory in open cloud computing federation2Proceedings of the 2nd International Conference on Industrial Mechatronics and Automation (ICIMA '10)May 201024024310.1109/icindma.2010.55383852-s2.0-77956217072BraunT. D.SiegelH. J.BeckN.BölöniL. L.MaheswaranM.ReutherA. I.RobertsonJ. P.TheysM. D.YaoB.HensgenD.FreundR. F.A comparison of eleven static heuristics for mapping a class of independent tasks onto heterogeneous distributed computing systemsRanaM.BilgaiyanS.KarU.A study on load balancing in cloud computing environment using evolutionary and swarm based algorithmsProceedings of the 2014 International Conference on Control, Instrumentation, Communication and Computational Technologies, ICCICCT 2014July 2014India2452502-s2.0-8492165423610.1109/ICCICCT.2014.6992964ZhuZ.ZhangG.LiM.LiuX.Evolutionary multi-objective workflow scheduling in cloudMaoY.ChenX.LiX.PatnaikS.LiX.Max–Min task scheduling algorithm for load balance in cloud computingKumarP.VermaA.Scheduling using improved genetic algorithm in cloud computing for independent tasksProceedings of the 2012 International Conference on Advances in Computing, Communications and Informatics, ICACCI 2012August 2012India1371422-s2.0-8486612557110.1145/2345396.2345420GuoL.ZhaoS.ShenS.JiangC.Task scheduling optimization in cloud computing based on heuristic algorithmZhangL.ChenY.SunR.JingS.YangB.A task scheduling algorithm based on PSO for grid computingPandeyS.WuL.GuruS. M.BuyyaR.A particle swarm optimization-based heuristic for scheduling workflow applications in cloud computing environmentsProceedings of the 24th IEEE International Conference on Advanced Information Networking and Applications, AINA2010April 2010Australia4004072-s2.0-7795433657110.1109/AINA.2010.31ArabnejadH.BarbosaJ. G.A Budget Constrained Scheduling Algorithm for Workflow ApplicationsXuH.YangB.QiW.AheneE.A multi-objective optimization approach to workflow scheduling in clouds considering fault recoveryChitraS.MadhusudhananB.SakthidharanG. R.SaravananP.Local minima jump PSO for workflow scheduling in cloud computing environmentsGeY.WeiG.GA-based task scheduler for the cloud computing systems2Proceedings of the International Conference on Web Information Systems and Mining (WISM '10)October 2010IEEE18118610.1109/wism.2010.872-s2.0-84871490850FardH. M.ProdanR.BarrionuevoJ. J. D.FahringerT.A multi-objective approach for workflow scheduling in heterogeneous environmentsProceedings of the 12th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing, CCGrid 2012May 2012Canada3003092-s2.0-8486367414110.1109/CCGrid.2012.114WuZ.LiuX.NiZ.YuanD.YangY.A market-oriented hierarchical scheduling strategy in cloud workflow systemsBa_aliH.ZhengW.SakellariouR.Budget-Deadline Constrained Workflow Planning for Admission ControlJacobJ. C.KatzD. S.PrinceT.BerrimanG. B.GoodJ. C.LaityA. C.DeelmanE.SinghG.SuM.-H.MagistraleH.DayS.ClaytonR. W.GravesR.The SCEC southern California reference three-dimensional seismic velocity model version 2DeelmanE.VahiK.JuveG.RyngeM.CallaghanS.MaechlingP. J.MayaniR.ChenW.Ferreira Da SilvaR.LivnyM.WengerK.Pegasus, a workflow management system for science automationBrownD. A.BradyP. R.DietzA.CaoJ.JohnsonB.McNabbJ.A case study on the use of workflow technologies for scientific analysis: Gravitational wave data analysisLivnyJ.TeonadiH.LivnyM.WaldorM. K.High-throughput, kingdom-wide prediction and annotation of bacterial non-coding RNAsAlajmiA.WrightJ.Selecting the most efficient genetic algorithm sets in solving unconstrained building optimization problemChenW.DeelmanE.WorkflowSim: A toolkit for simulating scientific workflows in distributed environmentsProceedings of the 2012 IEEE 8th International Conference on E-Science, e-Science 2012October 2012USA2-s2.0-8487366864110.1109/eScience.2012.6404430CalheirosR. N.RanjanR.BeloglazovA.de RoseC. A. F.BuyyaR.CloudSim: a toolkit for modeling and simulation of cloud computing environments and evaluation of resource provisioning algorithmsGhorbannia DelavarA.AryanY.HSGA: A hybrid heuristic algorithm for workflow scheduling in cloud systemsAmalarethinamD. G.BeenaT. L. A.Workflow Scheduling for Public Cloud Using Genetic Algorithm (WSGA)