Future Generation Computer Systems 83 (2018) 14–26

Contents lists available at ScienceDirect

Future Generation Computer Systems

journal homepage: www.elsevier.com/locate/fgcs

A GSA based hybrid algorithm for bi-objective workflow scheduling in

cloud computing
Anubhav Choudhary, Indrajeet Gupta, Vishakha Singh, Prasanta K. Jana *,1
Department of Computer Science and Engineering, Indian Institute of Technology (ISM), Dhanbad, India

highlights

• Proposed an efficient hybrid scheme of GSA and HEFT, called HGSA for workflow scheduling.
• Systematic derivation of fitness function based on makespan and cost.
• Novelty in introducing a proficient elimination strategy of inferior agents.
• Demonstration of better performance through simulation results and statistical test ANOVA.

article info a b s t r a c t
Article history: Workflow Scheduling in cloud computing has drawn enormous attention due to its wide application
Received 28 February 2017 in both scientific and business areas. This is particularly an NP-complete problem. Therefore, many
Received in revised form 12 October 2017 researchers have proposed a number of heuristics as well as meta-heuristic techniques by considering
Accepted 3 January 2018
several issues, such as energy conservation, cost and makespan. However, it is still an open area of research
Available online 8 January 2018
as most of the heuristics or meta-heuristics may not fulfill certain optimum criterion and produce near
optimal solution. In this paper, we propose a meta-heuristic based algorithm for workflow scheduling that
Keywords:
Gravitational Search Algorithm considers minimization of makespan and cost. The proposed algorithm is a hybridization of the popular
Workflow scheduling meta-heuristic, Gravitational Search Algorithm (GSA) and equally popular heuristic, Heterogeneous
Cost Earliest Finish Time (HEFT) to schedule workflow applications. We introduce a new factor called cost time
Makespan equivalence to make the bi-objective optimization more realistic. We consider monetary cost ratio (MCR)
Cost time equivalence and schedule length ratio (SLR) as the performance metrics to compare the performance of the proposed
algorithm with existing algorithms. With rigorous experiments over different scientific workflows, we
show the effectiveness of the proposed algorithm over standard GSA, Hybrid Genetic Algorithm (HGA)
and the HEFT. We validate the results by well-known statistical test, Analysis of Variance (ANOVA). In all
the cases, simulation results show that the proposed approach outperforms these algorithms.
© 2018 Elsevier B.V. All rights reserved.

1. Introduction resources which are provisioned dynamically. However, the allo-

Workflow has wide applications in business as well as in the
scientific areas such as astronomy, weather forecasting, medical
and bio-informatics. Generally, these workflows are vast in size
as they consist of a large number of independent and/or depen-
dent tasks and thus they demand huge infrastructure for their
computation, communication, and storage. Clouds [1] provide such
an infrastructure in order to execute the workflow on virtualized
* Corresponding author.
E-mail addresses: anubhav.choudhary@live.com (A. Choudhary),
indrajeet7830@gmail.com (I. Gupta), vs.make.a.vish@gmail.com (V. Singh),
prasantajana@yahoo.com (P.K. Jana).
1 IEEE Senior Member.
cation of the resources and the order in which the execution of
tasks of a given workflow will be performed, are matters of great
importance. This is commonly referred as workflow scheduling
problem. In fact, workflow scheduling is an NP-complete problem
which has been extensively studied for other paradigms, such
as grid and cluster computing. It is noteworthy that if there are
n tasks in a workflow and m available virtual machines (VMs),
then there exist mn different ways in which the tasks can be
mapped to the VM pool. For a large value of n and m, finding
an optimal solution by brute force approach is computationally
very expensive. Therefore, a meta-heuristic approach can be very
effective for solving this problem. However, every meta-heuristic
algorithm has its own merits and demerits. Hybridization of such
meta-heuristic approaches has evidence to produce better results

https://doi.org/10.1016/j.future.2018.01.005
0167-739X/© 2018 Elsevier B.V. All rights reserved.
 A. Choudhary et al. / Future Generation Computer Systems 83 (2018) 14–26
[2,3] and therefore this has become the recent trend of research in
cloud computing.
Heterogeneous Earliest Finish Time (HEFT) [4] is an efficient
heuristic proposed for task scheduling in heterogeneous multipro-
cessor which is also used for cloud computing [5–7]. This algorithm
maps each task arranged in a priority order to a VM for which
the earliest finish time is minimum. It should be noted that it is
essentially a single objective algorithm which can only optimize
the makespan. The Gravitational Search Algorithm (GSA) [8] is
a popular meta-heuristic approach which utilizes the concept of
the law of gravitation to find the near-optimal solution. The algo-
rithm starts with a set of random particles, where each particle
represents a solution and the mass of each particle is calculated
using a fitness function based on the application. Particles with
higher fitness value have the higher mass and, hence, it can exert
more force to attract other particles towards it. Eventually, all
the particles converge towards an optimal point. It is capable
of obtaining a global optimum faster than other meta-heuristic
algorithms and, hence, has a higher convergence rate. Moreover, it
provides better results than the Central Force Optimization (CFO)
and Particle Swarm Optimization (PSO) as demonstrated in [8].
In this paper, we propose a meta-heuristic based algorithm
for workflow scheduling problem which is a hybrid of the HEFT
and the GSA. Specifically, we address the following workflow
scheduling problem. Given a workflow consisting of a set of tasks
T = {t1 , t2 , . . . , tn } with their computational load and precedence
constraints, and also given a set of VMs, V = {v1 , v2 , . . . , vm }, our
objective is to map all the tasks to the available VMs so that the
entire workflow can be executed in minimum time and minimum
computational cost. The proposed algorithm is presented with an
efficient agent representation and systematic derivation of fitness
function. The algorithm is extensively simulated using the scien-
tific workflows of different sizes and is shown to produce better
results as compared to other related algorithms such as Hybrid
Genetic Algorithm (HGA), GSA and HEFT. We use ANOVA [9], a sta-
tistical test to validate the simulation results. This test determines
if a given result set has significant difference statistically with other
set of results.
Many algorithms have been proposed for workflow scheduling
in cloud computing which are based on meta-heuristic approaches.
For instance, Rodriguez et al. [10] have proposed a PSO based
algorithm with the objective of minimizing the execution cost
while meeting the deadline constraints. Similarly, HGA has been
presented in [3] that also has a single objective, i.e., minimizing
the makespan. Many other meta-heuristic approaches have been
developed for workflow scheduling, the survey of which can be
found in [11]. However, our approach is different from all such
approaches and has the following novelty. We consider parameters
such as communication bandwidth, output data size of each task,
VM boot time, VM shutdown time and performance variability of
VMs in order to create a more realistic environment for scheduling.
Most of these features are absent in the existing works. Moreover,
our approach deals with two objectives in contrast to single objec-
tive in many existing algorithms. Hybridization of GSA with HEFT is
also novel in the sense that there is full exploitation of the benefits
of both these algorithms.
Our contribution can be summarized as follows.
• A hybrid algorithm based on GSA and HEFT to minimize
makespan and total computational cost.
• An efficient agent representation and derivation of system-
atic fitness function.
• Introduction of cost time equivalence and a procedure for
eradicating the inferior agents.
• Demonstration of better performance of the algorithm
through extensive simulation and comparison with other
heuristic/meta-heuristic based approaches.
15
• Validation of the performance through the statistical test
ANOVA.
The rest of the paper is organized as follows. Related works are
stated in Section 2. Section 3 explains the application and cloud
model. Section 4 describes the terminologies used in the paper
and the problem statement. Section 5 presents the proposed work
with an illustration. Performance metrics, experimental results,
and comparison are discussed in Section 6 followed by Section 7
which concludes the paper.
2. Related works
Many heuristic and meta-heuristic based algorithms have been
proposed for workflow scheduling in cloud computing. In this
section, we present a short review of some of the works that are
relevant to our proposed scheme.
HEFT [4] is a popular heuristic which was initially developed
for task scheduling in heterogeneous multiprocessor systems. It is
well-known that HEFT performs better than many other heuris-
tics such as [12,13] for task scheduling. However, it considers
the minimization of makespan only. An extension of HEFT called
Pareto Optimal Scheduling Heuristic (POSH) was proposed by Su
et al. [5] for workflow scheduling in cloud, to minimize makespan
and cost of execution. The POSH produces acceptable solution.
Nevertheless, the solution is derived from a constricted search
space and thus, it may miss on the better solutions. An energy
efficient scheduling with deadline constraint for heterogeneous
cloud environment was proposed in [14]. In this work, a new VM
scheduler is developed which is shown to reduce energy consump-
tion in the execution of workflows. They claimed to achieve up
to 20% reduction in energy requirement and improvement in the
processing capacity by 8%. Fard et al. [15] proposed another heuris-
tic called multi-objective list scheduling (MOLS), which provides a
general framework for multi-objective static workflow scheduling.
It supports four objectives namely, makespan, cost, reliability, and
energy. Based on selected objectives it provides the execution plan.
Abrishami et al. [16] adopted the Partial Critical Path (PCP) for
workflow scheduling and designed two algorithms, a one-phase
algorithm which is called IaaS Cloud Partial Critical Paths (IC-PCP)
and a two-phase algorithm called IaaS Cloud Partial Critical Paths
with Deadline Distribution (IC-PCPD2). Here, homogeneous cloud
environment is assumed.
Recently, Casas et al. [17] have proposed balanced and
file reuse-replication scheduling (BaRRS) algorithm to schedule
workflows that are based on two optimization constraints, i.e.,
makespan and cost. They have also focused on finding an optimal
number of VMs required for a given workflow. However, it has
large computational overhead. Panda et al. [18] have developed a
normalization based task scheduling for a heterogeneous multi-
cloud environment. This technique provides a way to schedule
tasks over multiple cloud providers. In another work [19], they
have proposed a modification of min–min algorithm with uncer-
tainty parameter for scheduling tasks in a heterogeneous multi-
cloud environment. Gupta et al. [20] have also reported a work-
flow scheduling algorithm for the multi-cloud environment. How-
ever, this work focuses more on compute intensive workflows for
scheduling.
Meta-heuristics are well-known techniques to obtain near op-
timal solution. For instance, Pandey et al. [21] proposed a PSO
based workflow scheduling algorithm for cost optimization. It is
designed to consider computational cost and data transmission
cost to provide an execution plan such that overall cost is min-
imized. However, this approach has been tested only on limited
workflow applications. Jena et al. [22] proposed a multi-objective
nested Particle Swarm Optimization (TSPSO) algorithm for work-
flow scheduling to optimize energy as well as processing time.
16 A. Choudhary et al. / Future Generation Computer Systems 83 (2018) 14–26
Fig. 1. An example of workflow: Nodes represent the task and the edges represent
precedence relation. The numeric value 7 inside the node t2 is the computational
load of the task t2 and the label 11 is the size of the data generated by t2 .
However, a crucial scheduling factor, i.e, bandwidth between the
VMs is not considered in this work. Moreover, no experiment was
reported on large-scale workflows.
Many GA based solutions have also been proposed. Garg
et al. [23] proposed a hybrid GA driven by linear programming (LP)
for cost optimization in the grid computing. This work combines
the capabilities of both LP and GA to find a schedule such that
it minimizes the combined cost of all users of the grid. Wang
et al. [24] proposed a look-ahead genetic algorithm (LAGA) to
optimize both reliability and makespan. But this work focused
only on the compute intensive workflows and did not consider the
communication time, which is a very vital factor for scheduling
workflows.
3. Models
This section contains a detailed description about workflow
application model and the cloud server model assumed for the
development of the proposed algorithm. The important terminolo-
gies used throughout the paper are also described in this section.
3.1. Workflow application model
The workflow application is represented by a Direct Acyclic
Graph (DAG) [25,26], G = (T , E) as shown in Fig. 1, where T =
{t1 , t2 , . . . , tn } is the set of tasks and E is the set of edges. An edge
ti → tj indicates the precedence relation between the predecessor
ti and the successor tj . Thus, task tj cannot start unless task ti is
complete. Each task ti is labeled with its computational load in
million instructions (MI). Also, the label on each edge ti → tj
indicates the size of the output data generated by ti . This data is
required to start the execution of the task tj . The task without any
predecessor is termed as entry task (tentry ) and task without any
successor is termed as exit task (texit ). If there are more than one
entry tasks in the workflow, a new pseudo entry task is created
with zero computational load as well as no output data. Then all
the entry tasks are connected with the pseudo task so formed.
Similarly, pseudo exit task can also be created, if required.
3.2. Cloud server model
We assume a cloud server which contains set of m VMs, rep-
resented by V = {v1 , v2 , v3 , . . . , vm }. Each VM has its own com-
putational power measured in terms of million instructions per
second (MIPS). All VMs are fully connected to each other and may
reside in one or more physical cloud server. The time required to
transfer output data from task ti to tj is described as communica-
tion overhead time. Note that combination overhead time is the
ratio of output data size of a task ti to the bandwidth between the
VMs. If both ti and tj execute on the same VM, then communication
overhead is assumed to be zero.
4. Terminologies and problem statement
4.1. Constraints and assumptions
The notations used in the proposed work are given in Table 1.
The proposed algorithm considers the following constraints and
assumptions similar to the work as presented in [10].
1. We consider performance variance for the VMs to calcu-
late effective cpu cycles for the execution of the tasks. The
reasons behind this variability is due to heterogeneity and
shared nature of infrastructure of the underlying hardware.
Based on a survey [27], it is found that an overall perfor-
mance variability of Amazon’s EC2 cloud is 24%. For the VM
vj , performance variance is represented as degvj . Thus, using
degvj , execution time of task ti on VM vj can be written as
vj Load(ti )
ETti = (1)
Capacity(vj ) × (1 − degvj )
vj
where ETti is execution time of task ti on VM vj , Load(ti ) is
computational load of task ti and Capacity(vj ) is computa-
tional capacity of VM vj .
2. The unit chargeable time τ is considered to calculate the cost
of execution. If we utilize the leased VM for time less than τ ,
then also it is charged for a full time period.
3. An initial boot time is always required when a VM is
leased. So, we consider VM boot time for calculation of the
makespan. We also consider VM shutdown time as it is
required to release the provisioned VM.
4. Each VM is assumed to be roughly connected with the same
bandwidth.
4.2. Problem formulation
For the sake of bi-objective problem formulation, we first de-
scribe the two important parameters, i.e., makespan and cost as
follows.
Makespan Calculation:
Makespan is also referred as the total execution time for the
entire workflow. It includes the boot time of the leased VM, execu-
tion time, data transfer time between two VMs and the shutdown
time of VM. While computing makespan, it is assumed that a VM
cannot execute a task while data is being transferred from it to
another VM. The makespan is equal to the summation of VM boot
time, maximum of VM-time for all VM and the VM shutdown time.
VM-time[i] denotes the last timestamp (starting from zero for each
workflow) up to which the VM vi executes its assigned task. We
need to add VM boot time and VM shutdown time only once because
only the boot time of first VM and the shutdown time of last
VM will contribute to makespan and the rest will be overlapped
 A. Choudhary et al. / Future Generation Computer Systems 83 (2018) 14–26 17

Table 1
Notations and definitions.
Notation Definition
N Population size.
n Number of the tasks in a given workflow.
ti ith task.
m Number of available VMs.
vj jth VM.
Xi It represents the ith agent.
Xbest It represents the best agent known so far.
α It determines the weight of makespan and cost for calculation fitness.
β It is the cost makespan equivalence factor. It is a part of SLA and its
value depends on the priority and urgency of the application.
fiti Fitness value of the ith agent based on its makespan and cost.
Mi Mass of the ith agent.
σ A random variable used in the pricing model.
Vcbase Base price based on slowest VM
digvj Performance degradation of VM vj .
γ A small constant, which regulates the declination of gravitational
constant.
δ Threshold mass for replacing the inferior agents.

with other events. Therefore, makespan can be mathematically
formulated as follows.
m follows.
Makespan = VM-boot-time + max(VM-time[i])
i=1
+ VM-shutdown-time
Cost Calculation:
In our assumed model, a cloud server consists of VMs with
varying computational capacity for different types of workload.
vj
ETti is the execution time of the task ti on VM vj as defined in Eq.
(1). Let τ be the unit chargeable time for which charge of execution
of any task will be accounted. Let Vcbase be the base price charged
for the slowest VM, then as per the exponential pricing model [5],
the cost of execution of the task ti on VM vj is denoted as, cost(ti , vj )
and is formulated as follows.
⌈ vj ⌉
ETti
( )
cost(ti , vj ) = σ × × Vcbase × exp
τ slow est CPU cycle
where, σ is a random variable used to generate different combina-
tion of VM pricing and capacity. Let Bi,j be a boolean variable, such
that
1, if task ti is assigned to vj
{ }
Bi,j =
0, otherwise
Therefore, the total cost of execution for a workflow is defined as
n m
∑ ∑
Total Cost = Bi,j × cost(ti , vj )
i=1 j=1
However, the values of makespan and cost may have different
scales. So, the value of one attribute may overwhelm the value
of the other attribute, in the case of agent evaluation. Thus, it
is not valid to perform a linear formulation directly using the
actual values. One of the institutive approaches is to normalize
both makespan and cost in order to scale these values in the same
range. Normalization can be done using any of the well-known
methods such as min–max normalization, Z-score normalization,
etc. However, this solution has a major drawback, i.e, if there is a
change in global minimum and maximum values of makespan and
cost, it may lead to the variation in the relative rank of agents in
two consecutive iterations. To resolve this issue, we use makespan
equivalent of total cost (MEcost ) calculated using Eq. (6) instead of
total cost.
MEcost = β × Total Cost
Problem Statement:
The objective of the proposed work is to minimize makespan
and makespan equivalent of total cost as given in Eqs. (2) and
(6) respectively. Therefore, it is wise to minimize their linear com-
bination. The workflow scheduling problem can be formulated as
Minimize z = α × Makespan + (1 − α ) × MEcost
(2) m
∑
subject to (i) Bi,j = 1, i = 1, 2, 3, . . . , n (7)
j=1
(ii) 0 ≤ α ≤ 1
The constraint (i) indicates that any task of the workflow can be
assigned to one and only one VM and the constraint (ii) limits the
range of α which balances makespan and total cost.
5. Proposed work
CPU cycles of vj
(3) As our algorithm is a hybrid of GSA and HEFT, we first provide
a brief description about both of these algorithms as follows.
5.1. Overview of heterogeneous earliest finish time (HEFT)
(4) The main idea behind HEFT [4] is to schedule tasks in such a way
that earliest finish time (EFT ) is minimized for all the tasks. HEFT
is executed in two phases which is described as follows.
Phase 1: Calculating priority of tasks
(5)
In this phase, the priority of each task is calculated
using average execution time and average communi-
cation time. The priorities are calculated in a bottom -
up approach. The sequence of tasks generated based on
higher to lower priority value, satisfying the precedence
constraints of the given workflow. The priority of task ti
is given by
pri(ti ) = wi + max
( )
ci,j + pri(tj ) (8)
tj ∈succ(ti )
where, wi is the average execution time of task ti on
available VMs and ci,j is the average communication
time between task ti and task tj .
Phase 2: Mapping tasks to VMs
This is the main phase of the algorithm, where actual
mapping of task to VM is performed according to the
(6)
priority of the tasks. The task with the highest priority
is scheduled first, by calculating earliest start time (EST )
and EFT on all available VMs.
18 A. Choudhary et al. / Future Generation Computer Systems 83 (2018) 14–26

Table 2
Example of a Mapping / Agent.
Task t1 t2 t3 t4 t5 t6 t7 t8
VM 1 3 2 1 5 1 3 1

5.3. Agent representation

In the proposed algorithm, a mapping of the tasks of a given

workflow to the VMs, is considered as an agent. For workflow
having n tasks and cloud server having m number of VMs, the ith
agent can be represented as a vector of dimension 1 × n, i.e.,

Xi = x1i , x2i , x3i , . . . , xni

[ ]
(9)

xdi
where, represents the VM assigned to the task td . Note that is xdi
an integer which lies in interval [1, m]. Table 2 shows an example
of an agent for 8 tasks on 5 VMs. Here, task t1 is mapped with VM
v1 , i.e., task t1 will be executed on VM v1 , while preserving the
precedence constraints.

5.4. Fitness evaluation

For a given agent, we can always calculate its makespan and

makespan equivalent of total cost using Eqs. (2) and (6) respec-
tively. Let Makespani and MEcosti
be the makespan and makespan
equivalent of total cost for the ith agent in the population. Now,
the fitness value of ith agent can be computed as
1
Fitness(fiti ) = (10)
1 + α × Makespani + (1 − α ) × MEcost
i

The fitness value calculated using Eq. (10) is an absolute fit-

ness value. As its calculation does not require any information
about the population, the relative difference in fitness value of
two agents will remain same irrespective of the iteration in which
they are present. Agents with lower makespan and lower cost, will
have higher fitness value and they can be considered as potential
Fig. 2. Working of GSA.
candidates for a final solution instead of the agents with higher
makespan or higher cost.

5.2. Overview of gravitational search algorithm (GSA) Remark. The problem of workflow scheduling is to minimize the
linear combination of the makespan and the makespan equivalent
of total cost as described in Section 4.2. As our fitness value is
The GSA is based on the law of gravity [28] introduced by the reciprocal of the same, so the higher fitness value would be
Rashedi et al. [8]. It is a population-based search algorithm where desirable.
each agent is considered as a particle (so, we use agent and particle
In the proposed algorithm, we apply min–max normalization in
interchangeably throughout the paper) and its fitness value is
order scale the fitness values of all the agents in population in the
considered as its mass. Each particle represents a solution of a given range [0, 1] to get the mass of each agent. So, mass of the ith agent
problem. The main idea is that the heavier particles, i.e., superior is given by
solutions do not move much as compared to lighter particles, i.e., ( )
fiti − min fitj
inferior solutions. All particles apply force on every other particle. j=1 to N
As each particle has some mass, its acceleration and velocity can be Mass(Mi ) = ( ) ( ) (11)
max fitj − min fitj
calculated using the net force. Using calculated velocity, the new j=1 to N j=1 to N

position of the particle can be found. When algorithm terminates,
the particle with highest mass provides the near optimal solution.
The working of the GSA is depicted in Fig. 2.
To use this algorithm in scheduling, we first identify the search
space and particle representation. Then we initialize the popu-
lation randomly. Now in each iteration, we calculate the fitness
value of all particles using the fitness function defined as per the
optimization constraints. Based on the best and worst particles
identified, we calculate the mass to update the position of each
particle. The gravity constant that is used to calculate the velocity
and the position, is also updated in each iteration. We repeat all the
steps till the algorithm attains a certain termination criterion.
We use Eq. (11) as the fitness for our proposed work which is
used in the simulations.
5.5. Proposed scheduling technique
The proposed HGSA algorithm works in two phases. The first
phase focuses only on optimization of makespan. In the second
phase, it attempts to optimize the cost while trying to minimize
the fitness value which is calculated from both makespan and cost.
The result from the first phase guides the particle movement in
GSA, which is considered in the second phase of the proposed work.
This improves the result as compared to GSA with random initial
particles. We use GSA by incorporating HEFT with following steps.
 A. Choudhary et al. / Future Generation Computer Systems 83 (2018) 14–26 19

Step 1: The initial population is seeded with the output of the
HEFT algorithm. The HEFT heuristic provides guidance
to GSA algorithm that improves the overall performance
of the proposed algorithm. It helps to generate better
solutions in fewer number of iterations.
Step 2: The best particle identified from current generation based
on the fitness function is preserved. This is done to ensure
that the best agent does not get degraded in the future
generation.
Step 3: The agents having mass less than threshold mass (δ ) is
removed from the current population as they have very
little or no contribution in updating the population. In
place of all the removed agents, new agents are added to
the population generated with the help of the best agent
identified so far. This improves the overall fitness of the
population.
5.6. Position update of particle
Let us consider a system of N agents. We define the position of
the ith agent as follows.
Xi = x1i , x2i , x3i , . . . , xni for i = 1, 2, 3, . . . , N
[ ]
where, xdi
shows the position of ith agent in the dth dimension.
Let Mi (k) and G(k) be the mass of ith agent and the gravitational
constant respectively in kth iteration. We can define the force
acting on the ith agent by jth agent for the kth iteration as follows.
Mi (k) × Mj (k)
Fid,j (k) = G(k) × × xdj (k) − xdi (k)
( )
Ri,j (k) + ϵ
where, ϵ is a very small constant and Ri,j (k) is the Euclidean dis-
tance between ith and jth agent in the kth iteration. Ri,j (k) is defined
as
 
Ri,j (k) = Xi (k) · Xj (k))2
We suppose that the total force that acts on the ith agent in the
dth dimension be a randomly weighted sum of the forces exerted
in the dth dimension from the other agents. Then,
N
∑ 12
Fid (k) = randj × Fid,j (k)
j=1,j̸ =i
where randj is a random number that lies in the interval [0, 1]. By
the law of motion [28], the acceleration of the ith agent in the dth
dimension in kth iteration is given by,
Fid (k)
adi (k) =
Mi (k)
Furthermore, the next velocity of the agent is considered as a
fraction of its current velocity added to its acceleration. Therefore,
its position and velocity can be calculated as follows.
v eldi (k + 1) = randi × v eldi (k) + adi (k)
5.7. Algorithm
We start by generating initial population by random mapping
of the tasks on VMs in step 1 of the proposed algorithm, followed
by seeding of the results of HEFT into the population in step 2. Once
the initial population is ready, a set of iterative steps is applied
to each agent of the population to get the final result as per step
3 through step 16. The first step of iteration is to calculate the
gravitational constant in step 4. Then in step 5, we compute the
fitness value of each agent using Eq. (10). Note that Eq. (10) requires
the values of both cost and makespan. Algorithms 2 and 3 can be
utilized for calculating these values. Based on fitness value, we
identify the best and the worst agents in step 6 for calculating the
mass of all the agents in step 7. In step 8, we update the position
of each agent by calculating the net force, net acceleration and
velocity. In remaining steps, we replace the inferior agents by the
new agents generated with the help of best agent known so far. The
new agent is generated by mapping one of the tasks to a randomly
selected VM and rest of the mapping remains same as the best
agent.
Algorithm 1 : Proposed Workflow Scheduling Algorithm
(12) Input: Workflow Application (W ) and Cloud Server Specification
(CSS)
Output: Task mapping with VMs (M)
1 Initialize population X with N randomly generated agents.
2 Replace one of the agent by mapping generated by HEFT
(13) 3 for k = 1 to MAX_ITERATION do
4 Compute gravitational constant G(k) using Eq. (19)
5 Compute fitness value fiti for i = 1, 2, 3, ..., N using Eq. (10)
6 Identify best and worst agent based on calculated fitness
value.
7 Compute mass Mi for i = 1, 2, 3, ..., N using Eq. (11)
(14) 8 Update velocity and position of each agent using Eq. (13) to
(18)
9 for i = 1 to N do
10 if Mi < δ then
11 Pos = a random integer from interval [1, n]
xdi = xdbest for d = 1, 2, 3, ..., n.
(15)
13 xPos
i = a random integer from interval [1, m]
14 end if
15 end for
16 end for
17 Find M corresponding to the best agent based on the fiti for i =
1, 2, 3, ..., N
(16) 18 return M
Algorithm 2 : Cost-Calculation
Input: Workflow Application (W ), Mapping (M), Cloud Server
Specification (CSS)
(17) Output: Cost value (Cost)

xdi (k + 1) = xdi (k) + v eldi (k + 1) (18) 1 Set Cost = 0

2 for each task ti ∈ W do
vM [i] Load(ti )
The gravitational constant G0 , is initialized in the beginning and 3 Execution_time ETti = Capacity(vM [i] )×(1−degvM [i] )
reduces as the algorithm proceeds, in order to improve the search ⌈ vM [i] ⌉
ETt
accuracy. G(k) is a function of initial value G0 and iteration number 4 Unit_used = i
τ
k, defined as
CPU cycles of vM [i]
)γ ( )
Rate_per_unit = σ × Vcbase × exp
(
k0 5
G(k) = G0 × ,γ < 1 (19) slowest CPU cycle
k 6 Cost = Cost + Unit_used × Rate_per_unit
7 end for
where, γ is a small constant that regulates the reduction in the
8 return Cost
gravitational constant.
20 A. Choudhary et al. / Future Generation Computer Systems 83 (2018) 14–26

Algorithm 3 : Makespan-Calculation
Input: Workflow Application (W ), Mapping (M), Cloud Server Specification (CSS)
Output: Makespan value (Makespan)

1 for each vi ∈ V do
2 VM_time[i] = 0
3 end for
4 for each task ti ∈ W in topological order do
5 if ti .ParentCount ! = 0 then
6 Parent_finishtime = max (Tast_actual_finsh_time[k])
tk ∈pred(ti )
7 end if
8 if ti .ChildCount ! = 0 then
9 Transfer_time = 0
10 for each task tj where tj ∈ succ(ti ) and M [i] ̸ = M [j] do
11 if output data of ti task is not transferred to vM [j] then
t .Outputdatasize
12 Transfertime = Transfertime + i Bandw idth
13 end if
14 end for
15 end if
vM [i]
16 Execution_time ETti = Capacity(v Load(t i)
)×(1−deg vM [i] )
M [i]
17 Actual_start_time = max(Parent_finishtime, VM_time[M[i]])
18 Task_actual_finish_time[i] = Actual_start_time + Execution_time + Transfer_time
19 VM_time[M [i]] = Task_actual_finish_time[i]
20 end for
21 Makespan = VM_boot_time + max ( VM_time[i]) + VM_shutdown_time
vi ∈Cloud
22 return Makespan

Table 3
Parameters used in the illustration.
Parameter Value
Number of VMs 4
Computational power of all VMs 2.0, 3.5, 4.5 and 5.5 MIPS
Network bandwidth 1 MBps
Boot time and shutdown time of VM 0.5 sec
Performance variance of VM 24%
MAX _ITERATION 10
Population size (N) 100
Gravitational constant (G0 ) 5
Weight of makespan and cost (α ) 0.5
Cost time equivalence (β ) 1
Small constant used in gravity (γ ) 0.3
Mass threshold for inferior agents (δ ) 0.1
Small constant used in force (ϵ ) 10

is optimized in terms of makespan and cost. Table 3 shows the

(a) Montage workflow of 16 (b) Cloud environment.
tasks.
Fig. 3. Workflow and cloud environment.
5.8. An illustration
Consider a Montage workflow [29,30] consisting of 16 tasks,
T = {t1 , t2 , . . . , t16 } and a set of 4 VMs, V = {v1 , v2 , v3 , v4 }
as shown in Fig. 3. We have to schedule the workflow on the
given VMs which are fully connected to each other. The output
of this illustration is a mapping of given tasks on the VMs, which
parameters used in this illustration and Table 4 shows the initial
population generated, as described in step 1 of our proposed algo-
rithm.
We now compute a schedule using HEFT for the given workflow.
The resultant schedule is then included into the generated popula-
tion. To compute HEFT schedule, we need to find the priority using
Eq. (8). Now, starting from the highest priority, each task is mapped
to a VM in such a way that it minimizes the EFT . Table 5 shows the
priority as well as the mapping of the tasks to VMs. The makespan
and cost of schedule generated by HEFT are 36.57 sec and $50.31
respectively.
Fig. 4 demonstrates the process of including the mapping gen-
erated by HEFT into the initial population by replacing the shaded
agent. The selection of the agent to be replaced is purely random.
This completes the step 2 of the algorithm.
Now, the current population contains HEFT generated agent as
well as the agents generated randomly. These agents are processed
as described in step 3 through step 16, for a certain number of
iterations. Table 6 shows the details of the best agent identified
 A. Choudhary et al. / Future Generation Computer Systems 83 (2018) 14–26 21

Table 4
Initial population of agents.
Agent 1 4 1 1 2 3 4 4 2 3 1 2 3 2 1 1 4
Agent 2 2 3 1 1 1 2 3 4 4 1 1 2 1 3 3 3
. . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . .
Agent N 1 2 1 1 1 3 3 4 4 1 2 2 2 2 1 1

Fig. 4. Seeding of HEFT solution into population.

Table 5 Table 6
Task priority and task mapping by HEFT. Iteration wise specification of the best agent.
Task Priority Virtual machine Iteration Makespan Cost Fitness
t1 107.71 3 1 36.575226 50.313644 2.250 × 10−2
t2 103.37 2 2 32.338966 49.307339 2.391 × 10−2
t3 107.74 4 3 32.536972 48.014904 2.422 × 10−2
t4 103.24 4 4 32.642235 47.711269 2.428 × 10−2
t5 94.38 2 5 32.642235 47.711269 2.428 × 10−2
t6 94.44 3 6 32.642235 47.711269 2.428 × 10−2
t7 90.26 2 7 32.642235 47.711269 2.428 × 10−2
t8 90.20 4 8 32.642235 47.711269 2.428 × 10−2
t9 90.05 3 9 32.642235 47.711269 2.428 × 10−2
t10 90.08 2 10 32.642235 47.711269 2.428 × 10−2
t11 90.06 4
t12 90.08 4
t13 77.89 4
6.2. Performance metrics
t14 77.57 4
t15 2.63 4
t16 0.28 4 We normalize the makespan and monetary cost similar to that
in [5] and call them the schedule length ratio (SLR) and monetary
cost ratio (MCR) of tasks, as follows.
in each iteration based on calculated fitness value. The resultant Makespan
SLR = m
(20)
schedule is shown in Table 7 having makespan of 32.64 sec and ∑ { vj
}
min ETti
cost as $47.71. ti ∈CP
j=1

Total Cost
6. Experimental results and comparison MCR = (21)
m
min cost(ti , vj )
∑ { }
ti ∈CP
j=1
This section presents the simulation results of the proposed
algorithm and its comparison with three workflow scheduling The denominator is the summation of the minimum execution
algorithms including the standard GSA based approach, HEFT and time and monetary cost of the tasks on the critical path (CP)
HGA as follows. Note that, for the sake of comparison, we convert without communication cost. For a given task graph, an algorithm
the single objective of the HGA (minimization of makespan) into that produces a scheduling plan with lower SLR and lower MCR
bi-objective, keeping all the constraints same as the proposed value is more effective.
algorithm. We also calculate the normalized fitness value for easy compar-
ison and visualization of the overall quality of the results. We use
max-normalization to normalize the absolute fitness value as cal-
6.1. Experimental setup culated using Eq. (10). After applying normalization, the maximum
value is mapped to one and the rest of the values lie in the interval
The simulations were carried out using C++ coding environment (0,1]. Mathematically, max-normalization is defined as
on an Intel(R) Core(TM) i5-2540M CPU with 2.60 GHz and 4GB xi
x̂i = (22)
RAM running on Linux platform. The specifications of the cloud max (xj )
j=1 to N
environment, as well as the parameters used for evaluation of our
proposed algorithm, are given in Table 8. where, x̂i is the normalized value for xi .
22 A. Choudhary et al. / Future Generation Computer Systems 83 (2018) 14–26

Table 7
Resultant schedule of montage with 16 tasks.
Task t1 t2 t3 t4 t5 t6 t7 t8 t9 t10 t11 t12 t13 t14 t15 t16
VM 3 1 4 4 2 3 1 1 3 2 4 4 4 4 4 4

Fig. 5. (a) CyberShake (b) Epigenomics (c) Inspiral (d) Montage (e) SIPHT. (For interpretation of the references to color in this figure legend, the reader is referred to the web
version of this article.)

6.3. Dataset Table 8

Parameters used during experiment.

The proposed algorithm is evaluated on various scientific work- Parameter Value

flows as considered by Bharathi et al. [29] and Juve, et al. [30]. Network bandwidth 1 MBps
These workflows are synthesized using the generator program Boot time and shutdown time of VM 0.5 sec
Performance variance of VM 24%
provided by Pegasus project [31]. It uses information gathered
MAX _ITERATION 200
from actual execution of scientific workflows to generate a syn- Population size (N) 500
thetic workflow. This in turn, is a near approximation of a real Gravitational constant (G0 ) 5
workflow. We used CyberShake (IO and network-intensive), Epige- Weight of makespan (α ) 0.5
nomics (both compute-intensive and network-intensive), Inspiral Cost time equivalence (β ) 50
Small constant used in gravity (γ ) 0.3
(compute-intensive), Montage (network-intensive) and SIPHT (IO
Mass threshold for inferior agents (δ ) 0.1
intensive) in the simulation. For all the types of workflow, we di- Small constant used in force (ϵ ) 10
vide them into three categories based on the number of constituent
tasks as shown in Table 9.
Each workflow has some characteristic features, which play
pivotal roles in process of scheduling. The detailed characterization that it is a single objective scheduling algorithm which focuses on
of each workflow can be found in [29]. Topology of tasks in a makespan only.
given workflow is also an major criterion for scheduling. These To calculate the normalized fitness value, we used two input
workflows have variety of topological features such as pipeline parameters, such as α and β as shown in Table 1. The normalized
(yellow task nodes), data aggregation (red task nodes) and data fitness value shows the overall quality as per the user requirement.
partitioning (green task nodes) as shown in Fig. 5. From Fig. 9(a)–(c), we observe that the proposed HGSA algorithm
performs better than the HEFT, HGA and the GSA. We get the
6.4. Result analysis and performance evaluation better results using the HGSA even for the case where SLR is poor
with respect to the HEFT as the difference in cost is enough to
In this subsection, we evaluate the performance of our proposed compensate for the difference in makespan.
algorithm against the HEFT, standard GSA and the HGA with re-
spect to makespan and monetary cost as follows. The MCR, SLR and
6.5. Analysis of variance (ANOVA)
the normalized fitness are used as the performance metrics for the
comparative analysis which are defined in Section 6.2. Note that
the lower values of the SLR and MCR are desirable as they indicate We also conducted hypothesis testing using ANOVA [9]. It is
lower makespan and cost, respectively. However, the higher value a statistical method which compares the mean of two or more
of the normalized fitness is preferred. groups to determine whether there is a significant difference
We present the results obtained by using the same machine among the groups or not. This test has a null hypothesis (H0 ) and
configuration, same constraints, and the same set of workflow an alternate hypothesis (H1 ), defined as
applications (of various sizes and types). Figs. 6–9 show the bar
charts for the MCR, SLR and the normalized fitness value, so as to H0 : µ1 = µ2 = µ3 = ... = µn (23)
compare between the HGSA, HGA, standard GSA, and the HEFT.
From the figures, we can observe that the MCR of the proposed H1 : Means are not equal (24)
HGSA is better as compared to the HEFT, HGA and the GSA for all
the workflow categories, such as small, medium and large. Thus, During the test, if F statistical < F critical then we fail to reject
it is visible that the performance of the proposed HGSA is better the null hypothesis and all groups have the same mean. But if F
than others in terms of MCR for any aforementioned workflows. statistical > F critical then we reject the null hypothesis and accept
We also observe that the value of the SLR obtained by the proposed the alternate hypothesis. If the alternative hypothesis is accepted,
algorithm is much better than the GSA and the HGA. However, it is we can easily conclude that one of the group is having significant
somewhat lesser to that of given by the HEFT. This is due to the fact differences with the others.
 A. Choudhary et al. / Future Generation Computer Systems 83 (2018) 14–26 23

(a) Monetary cost ratio. (b) Schedule length ratio.

Fig. 6. Results for small sized workflows.

(a) Monetary cost ratio. (b) Schedule length ratio.

Fig. 7. Results for medium sized workflows.

(a) Monetary cost ratio. (b) Schedule length ratio.

Fig. 8. Results for large sized workflows.

24 A. Choudhary et al. / Future Generation Computer Systems 83 (2018) 14–26

(a) Normalized fitness for small sized workflows. (b) Normalized fitness for medium sized workflows.

(c) Normalized fitness value for large sized workflows.

Fig. 9. Comparison of normalized fitness.

Table 9 workflows of various sizes. Tables 10–14 shows the results for each
Category of the workflow based on the number of tasks. workflows of 2000 tasks.
Number of tasks Category As we can see that, for all workflows we have F statistical > F
24 to 60 Small critical. Thus, we can reject the null hypothesis. Therefore, means
100 to 400 Medium of all the groups are significantly different. This implies that the
800 to 2000 Large
performance of HGSA is better and consistent than HGA and GSA.

The test was performed to compare the standard GSA, hybrid
GA and the hybrid GSA. In order to do this experiment, all three
algorithms were executed 10 times for each of the five scientific
7. Conclusion
In this paper, we have presented a hybrid gravitational search
algorithm for scheduling workflows, with the basic objective of re-
ducing the makespan as well as the cost of execution. The efficiency

Table 10
ANOVA using CyberShake workflow of 2000 tasks.
(a) Summary of input
Group Count Sum Average Variance
HGSA 10 3.51E−05 3.51E−06 3.44E−16
GSA 10 2.96E−05 2.96E−06 5.10E−17
HGA 10 3.17E−05 3.17E−06 2.01E−16

(b) ANOVA test result

Source of variation SS df MS F stat P-value F crit
Between groups 1.56E−12 2 7.81E−13 3925.06 5.278E−34 3.35
Within groups 5.37E−15 27 1.99E−16
Total 1.57E−12 29
 A. Choudhary et al. / Future Generation Computer Systems 83 (2018) 14–26 25

Table 11
ANOVA using Epigenomics workflow of 2000 tasks.
(a) Summary of input
Group Count Sum Average Variance
HGSA 10 3.60E−07 3.60E−08 5.90E−20
GSA 10 3.00E−07 3.00E−08 1.30E−20
HGA 10 3.10E−07 3.10E−08 5.90E−20

(b) ANOVA test result

Source of variation SS df MS F stat P-value F crit
Between groups 1.60E−16 2 7.80E−17 1788.54 2.00E−29 3.35
Within groups 1.20E−18 27 4.40E−20
Total 1.60E−16 29

Table 12
ANOVA using Inspiral workflow of 2000 tasks.
(a) Summary of input
Group Count Sum Average Variance
HGSA 10 3.90E−06 3.90E−07 3.70E−18
GSA 10 3.50E−06 3.50E−07 4.50E−19
HGA 10 3.60E−06 3.60E−07 2.30E−18

(b) ANOVA test result

Source of variation SS df MS F stat P-value F crit
Between groups 9.60E−15 2 4.80E−15 2238.84 1.00E−30 3.35
Within groups 5.80E−17 27 2.10E−18
Total 9.70E−15 29

Table 13
ANOVA using Montage workflow of 2000 tasks.
(a) Summary of input
Group Count Sum Average Variance
HGSA 10 8.30E−05 8.30E−06 6.10E−16
GSA 10 5.80E−05 5.80E−06 8.80E−17
HGA 10 7.80E−05 7.80E−06 7.90E−15

(b) ANOVA test result

Source of variation SS df MS F stat P-value F crit
Between groups 3.60E−11 2 1.80E−11 6368.81 7.80E−37 3.35
Within groups 7.70E−14 27 2.90E−15
Total 3.70E−11 29

Table 14
ANOVA using SIPHT workflow of 2000 tasks.
(a) Summary of input
Group Count Sum Average Variance
HGSA 10 5.10E−06 5.10E−07 1.90E−17
GSA 10 4.30E−06 4.30E−07 2.60E−18
HGA 10 4.80E−06 4.80E−07 1.50E−17

(b) ANOVA test result

Source of variation SS df MS F stat P-value F crit
Between groups 3.00E−14 2 1.50E−14 1251.75 2.40E−27 3.35
Within groups 3.23E−16 27 1.20E−17
Total 3.03E−14 29

and usefulness of our algorithm has been demonstrated using var-
ious simulation runs. The results depict that HGSA performs better
than HGA by 14%, GSA by 20% and the HEFT by 35% with respect to
the fitness value. This clearly shows the supremacy of our scheme
over its contemporaries. Moreover, the consistent performance of
HGSA has been validated through the statistical test, ANOVA.
However, in the proposed approach, we have assumed that
the bandwidth between the VMs is fixed and all the tasks of the
workflow consist of simple programming instructions. Therefore,
we would like to extend our approach by considering the scenario,
where the nature of the available bandwidth is variable.Moreover,
we would like to apply the proposed approach for scheduling of
multiple workflows. It should be noted that in the proposed work
the execution time of a task is calculated based on the number of
instructions. However, for the complex tasks, this may not work
accurately. Thus, in the future work, we would like to support
complex tasks as well. For the sake of the simplicity of the work, we
assumed that there is no VM failure during the task execution. This
opens a possible future extension of this work. Our future efforts
will be made towards addressing all these issues accompanied by
a newly proposed scheduling algorithm with it’s simulation on
federated multi-cloud environment.
26 A. Choudhary et al. / Future Generation Computer Systems 83 (2018) 14–26
References
[1] R. Buyya, C.S. Yeo, S. Venugopal, J. Broberg, I. Brandic, Cloud computing and
emerging IT platforms: Vision, hype, and reality for delivering computing as
the 5th utility, Future Gener. Comput. Syst. 25 (6) (2009) 599–616.
[2] F. Tao, Y. Feng, L. Zhang, T. Liao, CLPS-GA: A case library and pareto solution-
based hybrid genetic algorithm for energy-aware cloud service scheduling,
Appl. Soft Comput. 19 (2014) 264–279.
[3] S.G. Ahmad, C.S. Liew, E.U. Munir, T.F. Ang, S.U. Khan, A hybrid genetic algo-
rithm for optimization of scheduling workflow applications in heterogeneous
computing systems, J. Parallel Distrib. Comput. 87 (2016) 80–90.
[4] H. Topcuoglu, S. Hariri, M.Y. Wu, Performance-effective and low-complexity
task scheduling for heterogeneous computing, IEEE Trans. Parallel Distributed
Syst. 13 (3) (2002) 260–274.
[5] S. Su, J. Li, Q. Huang, X. Huang, K. Shuang, J. Wang, Cost-efficient task scheduling
for executing large programs in the cloud, Parallel Comput. 39 (4) (2013) 177–
188.
[6] M.A. Vasile, F. Pop, R.-I. Tutueanu, V. Cristea, J. Kołodziej, Resource-aware
hybrid scheduling algorithm in heterogeneous distributed computing, Future
Gener. Comput. Syst. 51 (2015) 61–71.
[7] A. Verma, S. Kaushal, Cost-time efficient scheduling plan for executing work-
flows in the cloud, J. Grid Comput. 13 (4) (2015) 495–506.
[8] E. Rashedi, H. Nezamabadi-Pour, S. Saryazdi, GSA: A gravitational search algo-
rithm, Inf. Sci. 179 (13) (2009) 2232–2248.
[9] K.E. Muller, B.A. Fetterman, Regression and ANOVA: An Integrated Approach
using SAS Software, SAS Institute, 2002.
[10] M.A. Rodriguez, R. Buyya, Deadline based resource provisioningand scheduling
algorithm for scientific workflows on clouds, IEEE Trans. Cloud Comput. 2 (2)
(2014) 222–235.
[11] F. Fakhfakh, H.H. Kacem, A.H. Kacem, Workflow scheduling in cloud comput-
ing: A survey, in: 2014 IEEE 18th International Enterprise Distributed Object
Computing Conference Workshops and Demonstrations (EDOCW), IEEE, 2014,
pp. 372–378.
[12] J.G. Barbosa, B. Moreira, Dynamic scheduling of a batch of parallel task jobs on
heterogeneous clusters, Parallel Comput. 37 (8) (2011) 428–438.
[13] L.F. Bittencourt, R. Sakellariou, E.R. Madeira, Dag scheduling using a lookahead
variant of the heterogeneous earliest finish time algorithm, in: 2010 18th
Euromicro International Conference on Parallel, Distributed and Network-
Based Processing (PDP), IEEE, 2010, pp. 27–34.
[14] Y. Ding, X. Qin, L. Liu, T. Wang, Energy efficient scheduling of virtual machines
in cloud with deadline constraint, Future Gener. Comput. Syst. 50 (2015) 62–
74.
[15] H.M. Fard, R. Prodan, J.J.D. Barrionuevo, T. Fahringer, A multi-objective ap-
proach for workflow scheduling in heterogeneous environments, in: Procee
dings of the 2012 12th IEEE/ACM International Symposium on Cluster, Cloud
and Grid Computing (Ccgrid 2012), IEEE Computer Society, 2012, pp. 300–309.
[16] S. Abrishami, M. Naghibzadeh, D.H. Epema, Deadline-constrained workflow
scheduling algorithms for infrastructure as a service clouds, Future Gener.
Comput. Syst. 29 (1) (2013) 158–169.
[17] I. Casas, J. Taheri, R. Ranjan, L. Wang, A.Y. Zomaya, A balanced scheduler
with data reuse and replication for scientific workflows in cloud computing
systems, Future Gener. Comput. Syst. (2016).
[18] S.K. Panda, P.K. Jana, Normalization-based task scheduling algorithms for
heterogeneous multi-cloud environment, Inf. Syst. Front. (2016) 1–27.
[19] S.K. Panda, P.K. Jana, Uncertainty-based qos min–min algorithm for heteroge-
neous multi-cloud environment, Arab. J. Sci. Eng. 41 (8) (2016) 3003–3025.
[20] I. Gupta, M.S. Kumar, P.K. Jana, Compute-intensive workflow scheduling in
multi-cloud environment, in: 2016 International Conference on Advances in
Computing, Communications and Informatics (ICACCI), IEEE, 2016, pp. 315–
321.
[21] S. Pandey, L. Wu, S.M. Guru, R. Buyya, A particle swarm optimization-based
heuristic for scheduling workflow applications in cloud computing environ-
ments, in: 2010 24th IEEE International Conference on Advanced Information
Networking and Applications, IEEE, 2010, pp. 400–407.
[22] R. Jena, Multi objective task scheduling in cloud environment using nested pso
framework, Procedia Comput. Sci. 57 (2015) 1219–1227.
[23] S.K. Garg, P. Konugurthi, R. Buyya, A linear programming-driven genetic al-
gorithm for meta-scheduling on utility grids, Int. J. Parallel Emergent Distrib.
Syst. 26 (6) (2011) 493–517.
[24] X. Wang, C.S. Yeo, R. Buyya, J. Su, Optimizing the makespan and reliability for
workflow applications with reputation and a look-ahead genetic algorithm,
Future Gener. Comput. Syst. 27 (8) (2011) 1124–1134.
[25] M. Isard, M. Budiu, Y. Yu, A. Birrell, D. Fetterly, Dryad: Distributed data-parallel
programs from sequential building blocks, in: ACM SIGOPS Operating Systems
Review, Vol.41, ACM, 2007, pp. 59–72.
[26] Y. Xu, K. Li, L. He, T.K. Truong, A dag scheduling scheme on heterogeneous
computing systems using double molecular structure-based chemical reaction
optimization, J. Parallel Distrib. Comput. 73 (9) (2013) 1306–1322.
[27] J. Schad, J. Dittrich, J.A. Quiané Ruiz, Runtime measurements in the cloud:
Observing, analyzing, and reducing variance, Proc. VLDB Endow. 3 (1–2) (2010)
460–471.
[28] J. Walker, D. Halliday, R. Resnick, Fundamentals of Physics, Wiley, Hoboken,
NJ, 2008.
[29] S. Bharathi, A. Chervenak, E. Deelman, G. Mehta, M.H. Su, K. Vahi, Characteriza
tion of scientific workflows, in: 2008 Third Workshop on Workflows in Sup-
port of Large-Scale Science, IEEE, 2008, pp. 1–10.
[30] G. Juve, A. Chervenak, E. Deelman, S. Bharathi, G. Mehta, K. Vahi, Characterizing
and profiling scientific workflows, Future Gener. Comput. Syst. 29 (3) (2013)
682–692.
[31] Pegasus: Workflow generator: (accessed on December. 24th, 2016). https:
//github.com/pegasus-isi/WorkflowGenerator.
Anubhav Choudhary completed his M.Tech in Computer
Science and Engineering from Indian Institute of Technol-
ogy (ISM), Dhanbad, India. He is currently working as a
software engineer in Sandvine Technologies, Bengaluru,
India. His research interests include Cloud Computing and
Machine Learning. He completed his bachelor’s degree
course from Kalinga Institute of Industrial Technology,
Bhubaneswar, India.
Indrajeet Gupta is working as a senior research scholar
in the Department of CSE at IIT (ISM), Dhanbad, India. He
received M. Tech. degree from NIT, Rourkela, India. He
has published more than ten papers in reputed journals
and conferences. His current research interests include
workflow scheduling and load balancing in cloud comput-
ing. He acted as reviewers in many reputed journals and
conferences.
Vishakha Singh received M. Tech. in computer science
and engineering from Indian Institute of Technology (ISM),
Dhanbad, India. Her research interests include energy and
cost-efficient workflow scheduling in cloud computing.
She obtained her bachelor’s degree from Indian institute
of Information Technology, Jabalpur, India. Apart from her
Masters topic, her expertise also covers wireless sensor
networks field.
Prasanta K. Jana received M.Tech. degree in Computer
Science from University of Calcutta, in 1988 and Ph.D. from
Jadavpur University in 2000. Currently he is a Professor
in the department of Computer Science and Engineering,
IIT (ISM), Dhanbad, India. He has contributed 150 research
publications in his credit and co-authored six books and
two book chapters. He has also produced seven Ph.Ds. As
a recognition of his outstanding research contributions, he
has been awarded Senior Member of IEEE (The Institute
of Electrical and Electronics Engineers, USA) in 2010 and
Canara Bank Research Publication Award in 2015. He is
in the editorial board of two international journals and acted as referees in many
reputed international journals. Dr. Jana has acted as the General Chair of the In-
ternational Conference RAIT-2012, Co-chair of National Conference RAIT-2009 and
Convener of the workshop WPDC-2008. He has also served as advisory committee
and programme committee members of several International Conferences. His
current research interest includes wireless sensor networks, cloud computing and
big data clustering. He visited University of Aizu in 2003, Japan, Las Vegas, USA in
2008, Imperial college of London, UK in 2011, University of Macau, Macau in 2012
and Hong Kong in 2015 for academic purpose.