Received October 1, 2019, accepted October 12, 2019, date of publication October 21, 2019, date of current version

November 14, 2019.

Digital Object Identifier 10.1109/ACCESS.2019.2948704

A Task Scheduling Algorithm With Improved

Makespan Based on Prediction of Tasks
Computation Time algorithm for
Cloud Computing
BELAL ALI AL-MAYTAMI 1,2 , PINGZHI FAN1 , ABIR HUSSAIN 3, THAR BAKER 3,

AND PANOS LIATSIS 4

1 Institute
of Mobile communication, Southwest Jiaotong University, Chengdu 611756, China
2 Facultyof Science, Ibb University, Ibb 740005, Yemen
3 Department of Computer Science, Liverpool John Moores University, Liverpool L3 3AF, U.K.
4 Department of Electrical Engineering and Computer Science, Khalifa University of Science and Technology, Abu Dhabi, United Arab Emirates

Corresponding author: Abir Hussain ([email protected])

This work was supported by the 111 project under Grant 111-2-14.

ABSTRACT Cloud computing is extensively used in a variety of applications and domains, however
task and resource scheduling remains an area that requires improvement. Put simply, in a heterogeneous
computing system, task scheduling algorithms, which allow the transfer of incoming tasks to machines,
are needed to satisfy high performance data mapping requirements. The appropriate mapping between
resources and tasks reduces makespan and maximises resource utilisation. In this contribution, we present
a novel scheduling algorithm using Directed Acyclic Graph (DAG) based on the Prediction of Tasks
Computation Time algorithm (PTCT) to estimate the preeminent scheduling algorithm for prominent cloud
data. In addition, the proposed algorithm provides a significant improvement with respect to the makespan
and reduces the computation and complexity via employing Principle Components Analysis (PCA) and
reducing the Expected Time to Compute (ETC) matrix. Simulation results confirm the superior performance
of the algorithm for heterogeneous systems in terms of efficiency, speedup and schedule length ratio, when
compared to the state-of-the-art Min-Min, Max-Min, QoS-Guide and MiM-MaM scheduling algorithms.

INDEX TERMS Scheduling algorithm, task scheduling, resource utilization, cloud computing.

I. INTRODUCTION computing (HC) systems because of the various communi-

A. MOTIVATION AND AIM cation and execution rates amid various processors.
Cloud computing has grown to be a major technological The main aim of cloud computing is to provide a highly
enabler in companies and organizations [1]–[3]. It has been efficient platform for appropriate exploitation of computa-
shown to increase reliability, deliver cost-cutting solutions, tional properties embedded in organizations, and to sup-
and provide 24/7/365 access to hard/soft resources from port the enterprise to capitalize on end-user demands [9].
anywhere based on pay/use pricing policy [4], [5]. The cloud However, the decentralized and heterogeneous nature of
offers services in the structure of Software as a Service cloud networks makes them intricate to deal with. Last but
(SaaS), Infrastructure as a Service (IaaS), and Platform as not least, deciding on suitable assets for tasks has become an
a Service (PaaS) [3]. Task scheduling is a major challenge in acute issue due to the swift rise of users and resources.
widely distributed heterogeneous systems (e.g., cloud com- For heterogeneous clustering systems, task scheduling is
puting), which chooses the preeminent resources for a pro- a computationally demanding problem, even under abridged
vided task. Also, in heterogeneous systems, task scheduling conventions, as it is NP-hard [9], [12].
is more convoluted in comparison to homogeneous The overarching aim of this research is to improve the
performance of task scheduling, while reducing computa-
The associate editor coordinating the review of this manuscript and tional costs. A key objective is to predict the ideal algorithm
approving it for publication was Yaser Jararweh. for incoming/available data as and when needed. In order

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see http://creativecommons.org/licenses/by/4.0/
160916 VOLUME 7, 2019
B. A. Al-Maytami et al.: Task Scheduling Algorithm With Improved Makespan

to achieve this, we perform a systematic analysis of heuris- heuristic algorithms are used in finding an optimal solution,
tic techniques for resource utilization by means of Princi- when meta-heuristic method fail to do so. The associated
pal Components Analysis (PCA) in the cloud environment. solutions are achieved with improved speed, accuracy, opti-
Moreover, we analyze the requirements and consequences of mum transaction and completeness [13]. This section focuses
utilizing Quality of Service (QoS) with the proposed Predic- on popular heuristic algorithms for semi-optimal scheduling.
tion of Tasks Computation Time algorithm (PTCT).
A. MIN-MIN HEURISTIC ALGORITHM
B. CONTRIBUTIONS In the Min-Min scheduling algorithm, tasks with shorter
As described in Section 3, there are many works in the execution time are determined and prioritized, and then,
literature in regards to task scheduling in the cloud comput- resources that generate the minimum accomplished time are
ing environment. This research proposes the following novel assigned to them. This process is executed repeatedly for all
contributions: scheduled tasks. Hence, the Min–Min scheduling algorithm
1. Typically, task scheduling algorithms focus on improv- picks the smaller tasks to be completed first [19], [20].
ing either computation or communication costs in
the cloud data center. The proposed PTCT method B. MAX-MIN HEURISTIC ALGORITHM
decreases both of these costs noticeably, as discussed In contrast to Min-Min, the Max-Min algorithm operates on
and explained in Sections 3 and 5. the concept of completing the largest task first. The time
2. PTCT uses PCA to reduce the size of the required duration of each task is provided in advance and all tasks
matrices (refer to sections 4 and 5). To the best of the are mapped to the appropriate processor. This process is
authors’ knowledge, this is the first attempt of using repeated until all unmapped tasks are processed. It should be
PCA in task scheduling in the context of cloud com- noted that for small-scale distributed systems, Min-Min and
puting. Max-Min are suitably utilized [17], [21]. For all short scale
3. PTCT examines the incoming data (in tasks) to allocate tasks transported in the network, Max-Min simultaneously
the appropriate/capable processor from the data center. schedules the longer tasks, followed by the shorter ones,
In this way, PTCT guarantees to achieve near-optimal while Min-Min performs the opposite, i.e., schedule the
re-/allocation of resources, hence providing superior shorter ones first, followed by the longer tasks, which implies
scheduling performance, compared to benchmark algo- a larger makespan.
rithms, as illustrated in Section 5.
C. QUALITY OF SERVICE (QoS) GUIDED MIN-MIN TASK
C. STRUCTURE SCHEDULING HEURISTIC ALGORITHM
The rest of this research paper is structured as follows. The standard Min-Min heuristic scheme ignores Quality of
Section 2 discusses heuristic scheduling algorithms and Service (QoS) in its implantation. As a result, the effec-
summarizes their essential features. Section 3 provides an tiveness of the algorithm in the cloud is questionable. QoS
overview of related work in task scheduling. Section 4 intro- Guided Min-Min allows the QoS constraint to be utilized
duces the definition of the scheduling problem. Section 5 with standard Min-Min heuristic scheduling [19]. Specifi-
describes the proposed PTCT method and presents strategies cally, some scheduled tasks, particularly those with sensitive
for practical algorithms based on PTCT. Section 6 presents data, may necessitate the use of high bandwidth, while others
simulation results and provides comparisons to the state-of- can be accomplished with low bandwidth. This scheduling
the-art methods. Section 7 provides the conclusions of this algorithm allocates tasks with excessive QoS requests first
research and identifies opportunities for further work. using the Min-Min heuristic. When all tasks entail either high
QoS or low QoS, the scheme requires O(n2 m) computational
II. HEURISTIC SCHEDULING ALGORITHMS complexity.
Scheduling is a decision-making process carried out in real
time, where processors are allocated to an extensive set of D. MIM-MAM ALGORITHM
tasks. Due to resource constraints, task scheduling is a chal- • This scheduling algorithm builds upon the advantages
lenging problem. Various studies investigated task scheduling and limitations of the Max-Min and Min-Min heuristic
algorithms and proposed schemes to improve resource uti- algorithms. In this case, information about the upcoming
lization in widely distributed environments, i.e., cloud com- deadline for each task, arrival rate, cost of execution
puting. Solving the associated NP-hard problem leads to the of using each available resource, and communication
optimal solution. costs is considered [23]. Two types of policies are used
Various heuristics-based schemes have been shown to to classify task scheduling problems, specifically static
provide semi-optimal solutions. Heuristic scheduling algo- and dynamic scheduling. In the former, information
rithms are rule-based, and extensively used in the IaaS about tasks, including communication costs for each
cloud computing environment. They are intended to sort out task, execution and the relationship with other tasks,
challenging problems faster than meta-heuristic algorithms, is determined in advance. In the latter, decisions are
where implantation is time consuming. Furthermore, determined in runtime, since the details of the tasks

VOLUME 7, 2019 160917

 B. A. Al-Maytami et al.: Task Scheduling Algorithm With Improved Makespan

are not obtainable. In addition, dynamic scheduling rion, this is deemed one of the limitations of their study. To
signifies runtime scheduling, while static scheduling overcome this limitation, Sinnen and Sousa [36] use network
is a representation of compile-time scheduling. There contention in their task scheduling method, without consid-
are two major groups of static scheduling algorithms: ering the fee charged to customers for using these resources.
(i) Heuristic-based algorithms, and (ii) Guided Random Two factors must be considered in the cloud computing
Search based algorithms. The former provides good environment, i.e., high performance of data transfer and sat-
approximate solutions with polynomial time complex- isfaction of budget constraints. The authors in [37] and [38]
ity [14]. Dynamic scheduling is faster in execution than introduce a cost-efficient algorithm to select the most appro-
static scheduling, since it’s basically not aware of any priate system in a cloud environment to implement the work-
thread dependencies [7]. flow based on using the deadline and cost saving constraints.
Li et al. [39] illustrate a scheduling algorithm, which can
1) RELATED WORK be applied in large graph processing, where both cost and
As discussed in Section 2, there exists a variety of heuris- schedule length constraints are considered. However, their
tic scheduling algorithms, which can operate in both batch scheme does not consider failed devices.
and online modes. Some of these schemes are appropriate Issues of task scheduling have been widely considered
in heterogeneous scheduling scenarios, however they cannot in the literature. As expected, a multitude of approaches
always attain good makespan, speedup, reduced costs and have been proposed due to its crucial effects on performance
increased efficiency [1], [6]–[8], [13]. Hence, QoS-based [9], [15]. The heuristic algorithm based on list scheduling
techniques are essential in obtaining the maximum objec- strategies [9] is one of the conventional scheduling algorithms
tives so as to retain QoS characteristics for both tasks and for cloud environments. This provides low time complexity,
resources. however the limitations of minimal universality and poor
Wang and Yu [29] propose an enhanced Min-Min algo- convergence have.
rithm to consider the proficiency of task scheduling in cloud In [42], the authors study load balancing in the cloud
computing. As previously indicated, the Min-Min algorithm environment to avoid problems, which may occur due to
first determines the tasks with shorter execution times and increase in power consumption, node failure, and machine
then the resources which result in the shortest times. This can failure. However, the research dealt with a limited number
lead to delays when examining the use of the algorithm in of parameters, e.g., there is no analysis on the effects of
the cloud environment. Zhang et al. [30] propose QoS con- dynamic scheduling, increase in the number of tasks and
straints in the cloud environment as a criteria for scheduling machines, as well the growth of users. In [43], additional
a task in the Min-Min algorithm, named Mul-QoS-Min-Min. parameters are considered. Optimization of task scheduling
The proposed algorithm finds resources with similar tasks to is addressed by introducing the iterative selection operator.
deliver task scheduling, then requests users to carry out their However, this study overlooks the issue of load balancing.
needs. The simulation results indicate that the performance Shimada et al., [44] proposed a novel algorithm, which can
of the Mul-QoS-Min-Min scheme is improved in terms of transfer the task with the shorter path while eliminating
execution times, when benchmarked against the traditional redundant tasks. However, the issue of the increase in the
Min-Min algorithm. number of machines as the number of tasks increases remains
Both Mao et al. [31] endorse the Max-Min algorithm an open challenge. In [45], the authors propose a model to
in order to stabilize the load for the cloud. The algorithm increase the overall system utilization, however, load bal-
conserves a table that holds details about task position and ancing and other performance parameters need to be further
evaluates the real-time workload for virtual machines (VMs) improved. Other works, such as [52], [53] explore the coop-
with the estimated task execution times. The Max-Min algo- eration and collaboration among cloud servers using multi-
rithm boosts the utilization of resources and decreases task agent approaches to best assign resources to incoming tasks.
scheduling response time by using VMs instead of traditional
resources. Li et al. [32] schedules tasks using improved max-
min task scheduling then largest task is too large compared III. PROBLEM FORMULATION
to other tasks in Meta-task in this case overall makespan In this research, we tackle the problem of static scheduling of
is increased because too large task is executed by slowest a single application in a widely distributed and heterogeneous
resource. environment. Let us consider the sets of tasks and processors,
Henning et al. [34] study task scheduling in the parallel T and P, respectively. Let P processors be available for the
method challenge with a fixed number of processors and the set of tasks T, which are not shared during task execution.
best schedule for high performance outcomes. They indicate Let ETC is the Expected Time to compute, the matrix which
that this can be achieved by mapping tasks to machines contains in each row the estimated execution time of a given
according to precedence constraints. In [35], the authors pro- task on each resource, and the estimated execution time of a
pose a task scheduling mechanism for allocating computing resource in each column. The aim is to reduce the makespan
processors to a so-called ‘‘task graph templates’’. Since the time of task execution in the data center. To minimize the
authors do not consider the network connection as a crite- makespan and increase the utilization of resources, the tasks

160918 VOLUME 7, 2019

B. A. Al-Maytami et al.: Task Scheduling Algorithm With Improved Makespan

of parallel applications have to be efficiently scheduled on the TABLE 1. Computation costs of tasks in Fig 1.
available resources using the ETC matrix model [28].

A. TASK GRAPH MODE

Using Directed Acyclic Graph (DAG), it is not possible to
start at one point in the graph and traverse the entire graph,
while evaluating computation costs in the correct order. Let us
consider a DAG, where G = (V, E). In this case, V represents
the set of v nodes. Each node vi ∈ V refers to an application
task, which is comprised of instructions executed on the same
processor. The parameter E represents a set of e connection
boundaries between the tasks. Boundary e(i, j) ∈ E charac-
terizes the task-dependency constraint, where task vi needs
to be completed before task vj is able to start. The DAG
is supplemented by the matrix ETC using v × p tabulated
cost. The parameters v and p are the number of tasks and
the number of processors in the system, respectively. Element
wi,j of W provides the estimated time of completing task vi
on processor pj . The mean execution time of task vi is given
by:
X
wi = wi,j /p (1)
j∈P

Figure 1 represents a simple model commonly used

in the scheduling problem [14]–[17], which allows us to
provide a competitive likeness with state-of-the-art solu-
tions, since these simplifications correspond to real systems.
Table 1 shows the computation costs of the 11 tasks shown
in Figure 1, with randomly generated values to provide a accomplishing an application. Consequently, the task
sophisticated example on extracting the values of ETC. scheduling problem has been widely studied and many
algorithms have been proposed including list scheduling,
clustering, and task duplication scheduling based on Genetic
Algorithm. In summary, list-scheduling algorithms are ideal
in delivering low cost solutions, in comparison to other
approaches. Clustering algorithms perform better in the
case of homogeneous processors. Finally, task duplication
scheduling algorithms are utilized for communication inten-
sive programs. A point to note is that a review of the open
literature on task scheduling revealed a number of enhance-
ments for homogeneous processors [8], [10], [16]–[18],
however there appears to be less progress in the
case of heterogeneous processors [1]–[22]. This pro-
vides further motivation for the development of our
proposed framework in the context of a heterogeneous
environment.
Consider the following two attributes, Earliest Start Time
FIGURE 1. A simple task graph with 11 tasks. (EST) and Earliest Finish Time (EFT), used to outline the
objectives of the task scheduling issue. EST(vi , pj ) repre-
sents the EST for task vi on processor pj , and similarly,
IV. PROPOSED METHOD EFT(vi , pj ) represents EFT for task vi on processor pj . EST(vi )
This section introduces the general framework of the pro- and EFT(vi ) represent the values of these attributes over the
posed PTCT algorithm, including algorithmic details. set of processors, respectively. For any initial entry task,
ventry , EST(ventry ) = 0, the values of EST and EFT are
A. OVERVIEW calculated from the entry to the exit tasks, traveling the task
In heterogeneous computing, effective task scheduling is graph from top to bottom. All immediate predecessor tasks of
of the utmost importance to increase the advantages of vi should be scheduled to allow the calculation of EST. The

VOLUME 7, 2019 160919

 B. A. Al-Maytami et al.: Task Scheduling Algorithm With Improved Makespan

task scheduling problem is defined as follows:

n
EST vi , pj = max pavail [vi ,pj ] , max EFT vp , vi

+ C vp , vi



(2)
where vp pred (vi ) , EFT vp , pk = EFT vp

C vp , vi = 0 when k = j


(3)
EFT vi , pj = T vi , pj + EST vi , pj




(4)

Pavail[vi,pj] is defined as the earliest time for processor pj

to execute task vi , pred(vi) while EST(vi , pj ) represents the
maximum time when processor pj will be available. k is a
counter, meantime this represents the time where last message
arrives from any of task vi predecessors.

makespan = EFT (vExit , pj ), (5)

where vExit is the exit task and ETC(i,j) and PCA are defined FIGURE 2. The architecture of the proposed PTCT scheme.
as:
The proposed PTCT algorithm is shown in Algorithm 1,
ETC(i, j) while Algorithm 2 illustrates how to calculate the execution
ETC1,1 ETC1,2 ETC1,3 . . .
 
ETC1,j time of task vi on host pj and Algorithm 3 provides details
 ETC2,1 ETC2,2 ETC2,3 . . . ETC2,j  about calculating the expected execution time of task vi on
 .. .. .. .. .. 
 
host pj .
=  . . . . . 
 .. .. .. .. .. 

 . . . . .  V. SIMULATION RESULTS
ETCi,1 ETCi,2 ETCi,3 . . . ETCi,j Simulation experiments were run in MATLAB R2013a on
PCA(ETC) a PC with Intel Core i5 processor, using 2.40 GHz CPU
and 8 GB RAM. Windows 7 was utilised as the OS for
λ(ETC1 ) ... ... ... ...
 
.. .. .. .. the platform. The proposed PTCT algorithm was bench-
. λ(ETC2 ) . . .
 
  marked with Min-Min [20], Max-Min [21], QoS guided [33]
. . .. ..
 
=
 .. .. λ(ETC3 ) . .

 and MiM-MaM [23], developed for heterogeneous system.
.. .. .. .. Simulation experiments were repeated 20 times, using the
..
 
.
 
 . . . .  parameters in Table 3.
... ... ... . . . λ(ETCi )
TABLE 3. Simulation parameters.
Let us consider the notation of Table 2.

TABLE 2. Notation.

The following assumptions were made:

1. Expected Time to Compute (ETC) of size v × p is
used, where v and p represent the number of tasks and
resources, respectively.
2. Tasks have no priorities associated with them.
3. Independent tasks are assigned to available resources.
4. The availability of resources and the number of tasks to
be executed are known in advance [24].
5. We consider three types of QoS data: high, medium,
and low.

B. SYSTEM MODEL OF THE PROPOSED PTCT A. PERFORMANCE ANALYSIS

The aim of the proposed PTCT scheme is to minimize the Performance analysis was carried out based on three quality
scheduling rate (makespan), as well as estimate the optimal measures, i.e., makespan, speedup and efficiency. These were
algorithm for scheduling as shown in Figure 2. defined as follows:

160920 VOLUME 7, 2019

B. A. Al-Maytami et al.: Task Scheduling Algorithm With Improved Makespan

TABLE 4. Makespan of PTCT vs four state-of-the-art algorithms as a function of the number of tasks.

FIGURE 3. Makespan of state-of-the-art algorithms vs the proposed PTCT scheme as a function of the number of tasks.

a) Makespan: this represents the main quality measure as time by estimating the finishing time of the last
it provides the completion time for all tasks in a graph. task [27] and is calculated according to:
Makespan is utilized to find the maximum completion Makespan = EFT (vi , pj ) (6)

VOLUME 7, 2019 160921

 B. A. Al-Maytami et al.: Task Scheduling Algorithm With Improved Makespan

FIGURE 4. Speedup of different algorithm in comparison to our proposed PTCT scheme.

b) Speedup: this is defined as the ratio of the sequential schedule is attained for the task ready to run at its next
schedule length calculated by allocating all tasks to the scheduled time, thus decreasing the time complexity
fastest processor over the execution time of the task of each schedule. The computational complexity of
schedule (makespan). as shown in Equation (7). The MiN-MiN, MaX-MiN, QoS-Guide and MiM-MaM is
sequential execution time is the cumulative computa- O (n2 m), where n refers to the number of nodes, and m
tion cost when assigning all the tasks vi sequentially to is the number of edges. The computational complexity
a single computing host p, is the host, H the set of hosts, of PTCT is O(m).because of the matrix of ETC will
V is the set of tasks contains only one dimension instead of two.
P
minp∈H ( vi ∈V wi,j ) Makespan is the maximum finish time of the exit task in
Speedup = (7) the scheduled DAG. From Figure 3, it can be noted that the
makespan makespan of PTCT decreases following the application of
c) Efficiency: this is the ratio of speedup to the total num- PCA. Since the PTCT algorithm minimizes the communica-
ber of processors, p, utilized to schedule the entire DAG tion overheads, the time required for completing application
application: execution by the PTCT algorithm is lower than the bench-
marked algorithms.
Speedup
Efficincy = (8) In Equation 7, wi,j represents the weight of task ti on
p processor pj . Speedup is a good quality measure for execut-
d) Complexity: Using the PCA algorithm for all tasks and ing the application program using a parallel system. From
machines is feasible for low-complexity scheduling Figure 4, it can be seen that the speedup of PTCT algorithm
algorithms. The information requested from all tasks is is higher than the other algorithms, since ETC is reduced
utilized to calculate the critical path and the scheduling to one instead of two dimensions. Figure 5 illustrates that
algorithm is executed in such a way that it stops after a the efficiency of the proposed algorithm when benchmarked

160922 VOLUME 7, 2019

B. A. Al-Maytami et al.: Task Scheduling Algorithm With Improved Makespan

FIGURE 5. Efficiency of state-of-the-art algorithms vs the proposed PTCT scheme as a function of the number of tasks.

with the state of the art algorithms. It shows improved matrix, that will leads to reduce the time used to complete
results, in particular, when the number of tasks is increased to tasks. We used 8 processors and a range of [10-90] tasks.
above 40. The results in Table 4 and Figure 3 indicate that the algorithm
provided positive results, i.e., reduced times to complete tasks
on resources, due to reducing the size of the ETC matrix by
B. DISCUSSION using PCA.
For comparison purposes, the makespan, speedup and effi- Figure 4 also illustrates positive results in terms of speedup,
ciency are used to illustrate the performance of makespan. which is one of the essential criteria for measuring the per-
Figure 3 shows comparative results of four state-of-the- formance of algorithms for the scheduling task. In addition
art algorithms with the PTCT scheme. In the experiments, to the efficiency, which refers to the ratio of speedup with the
the same data was used to compare the performance of the number of processors used, Figure 5 depicts better results for
algorithms. Compare the PCTC algorithm with each algo- PTCT in comparison with the other algorithms. The simula-
rithm separately. In other words, compare each algorithm tion results indicate that the efficiency of our proposed tech-
before and after using PTCT. The results show an improve- nique is significantly improved, when increasing the number
ment in the performance of each algorithm with PTCT. of tasks apart from the Min-Min algorithm, which shows sim-
By repeating the experiment 20 times, the best algorithm will ilar performance (when the number of tasks is less or equal
be selected. MiN-MiN shows less effect with using PTCT, to 90). However, Min-Min does not take into consideration
and this usually depends on the quality of the used data. But QoS as is the case with the proposed algorithm. Further-
as long as the PTCT algorithm decreases the size of the ETC more, with larger number of tasks (>90), additional simu-

VOLUME 7, 2019 160923

 B. A. Al-Maytami et al.: Task Scheduling Algorithm With Improved Makespan

Algorithm 1 Prediction of Tasks Computation Time Algorithm 3 Main Algorithm to Calculate the Execu-
(PTCT) tion Time of Task vi on Host pj
Input: Input:
Output: p = processors;
1: Generate data (v, p) v = tasks;
2: Group the data in S groups, where each group can ETC(vi , pj )
include high and low QoS Output: PTCTm (ETC(vi , pj )), m= number of
3: For each group of data, run six algorithms and output algorithms
the makespan time for each algorithm 1: Group the data (v, p)in S groups randomly, where
4: Hi ={h1 , h2 , . . . . . . ., hm ), m is number of machines, each group may include high and low QoS
H (hosts), i=1,. . . , 4 (for all algorithms) 2: While there is a group of QoS
5: Select the best pair (Hk , Fk ), i.e., minimum 3: For all Gn (ETC(vi , pj )) (generate the matrix
makespan denoting the best algorithm k. ETC
6: Use PCA 4: For all algorithms = m
m
7: ψ =
1 P
hi mean value for every dimension of the 5: Compute makespan for each algorithm
m
i=1 6: Fm ={f1 , f2 , .., fm };
matrix ETC. 7: makespan PCA(Fm ={f1 , f2 , . . . fm });
8: φi = |hi − ψ| 8: Select best pair (minimum makespan)
9: Set A = [φ1 , φ2 , ..........φm ] Eq.5
10: Covariance C = A AT (Fk , Gk ), which denotes the best
11: Find the eigenvector W=A AT algorithm k
12: Save <U, k> in the database 9: End for
13: Find min||W – Ui ||, i.e., the closest eigenvector in 10: End for
the database and choose its algorithm Fi 11: End while.
14: Predicated algorithm.

VI. CONCLUSION
Algorithm 2 Calculation of Execution Time of Task
In this work, a novel algorithm, Prediction of Tasks Com-
vi on Host pj
putation Time, was presented. This results in a performance
Input: improvement in cloud-based task scheduling by using Prin-
p = processors; cipal Component Analysis. This permits the reduction of the
v = tasks; size of the Expected Time to Compute (ETC) matrix.
Output: ETC(vi , pj ) The proposed algorithm was applied to simulated task
1: While there are more tasks to be scheduled graphs, and its performance was assessed in terms of
2: For all vi to schedule speed-up, makespan, schedule length ratio and efficiency.
3: For all pj The simulation results showed improved performance,
4: Compute CTi,j =CT(vi , pj ) when benchmarked with four state-of-the-art scheduling
5: End for loop algorithms, namely Min-Min, Max-Min, QoS-guided and
6: Compute i = F(CTi,1 , CTi,2 , . . . . . . . . . ) MiM-MaM. In the cloud computing context, the simulation
7: End for loop results indicated that the proposed PTCT can reduce the
8: Determine the best F match m overall makespan and task execution time.
9: Find minimum CTm,n The simulation setup was based on static scheduling, where
10: task m will be scheduled on n task arrival at the processors and speed are assumed to
11: End while loop. be known. Future work will consider dynamic scheduling
for real-world application graphs and benchmarking in real-
world problems. The focus will be on improving the total
energy utilization and consumption of task scheduling using
lation results indicated that the proposed PTCT generated
the PTCT algorithm and comparing the findings with relevant
significantly improved results in comparison to MiN-MiN
state-of-the-art algorithms for cloud energy consumption,
algorithm. This in agreement with existing literature research
such as GreeDi and GreeAODV [47]–[51].
[40], [41].
One limitation of the proposed algorithm is that minimiz- REFERENCES
ing the ETC matrix may lead to reduced accuracy values. [1] A. I. Avetisyan, R. Campbell, I. Gupta, M. T. Heath, S. Y. Ko,
Nevertheless, the simulation results demonstrated that this did G. R. Ganger, M. A. Kozuch, D. O’Hallaron, M. Kunze, T. T. Kwan,
K. Lai, M. Lyons, D. S. Milojicic, H. Y. Lee, Y. C. Soh, N. K. Ming,
not affect the quality measures benchmarked against the state- J.-Y. Luke, and H. Namgoong, ‘‘Open cirrus: A global cloud computing
of-the-art techniques. testbed,’’ Computer, vol. 43, no. 4, pp. 35–43, Apr. 2010.

160924 VOLUME 7, 2019

B. A. Al-Maytami et al.: Task Scheduling Algorithm With Improved Makespan

[2] S. K. Panda and P. K. Jana, ‘‘Efficient task scheduling algorithms for [23] S. V. Kfatheen and M. N. Banu, ‘‘MiM-MaM: A new task scheduling
heterogeneous multi-cloud environment,’’ J. Supercomput., vol. 71, no. 4, algorithm for grid environment,’’ in Proc. Int. Conf. Adv. Comput. Eng.
pp. 1505–1533, Apr. 2015. Appl., Mar. 2015, pp. 695–699.
[3] R. Buyya, C. S. Yeo, S. Venugopal, J. Broberg, and I. Brandic, ‘‘Cloud [24] S. V. Kfatheen and A. Marimuthu, ‘‘ETS: An efficient task scheduling
computing and emerging IT platforms: Vision, hype, and reality for deliv- algorithm for grid computing,’’ Adv. Comput. Sci. Technol. vol. 10, no. 10,
ering computing as the 5th utility,’’ Future Gener. Comput. Syst. vol. 25, pp. 2911–2925, 2017.
no. 6, pp. 599–616, Jun. 2009. [25] S. H. H. Madni, M. S. A. Latiff, Y. Coulibaly, and S. M. Abdulhamid,
[4] K. A. Beaty, V. K. Naik, and C.-S. Perng, ‘‘Economics of cloud com- ‘‘Resource scheduling for infrastructure as a service (IaaS) in cloud com-
puting for enterprise IT,’’ IBM J. Res. Develop. vol. 55, no. 6, pp. 1–12, puting: Challenges and opportunities,’’ J. Netw. Comput. Appl., vol. 68,
Nov./Dec. 2011. pp. 173–200, Jun. 2016.
[5] I. Foster, Y. Zhao, I. Raicu, and S. Lu, ‘‘Cloud computing and grid com- [26] S. M. Abdulhamid, M. S. A. Latiff, G. Abdul-Salaam, and S. H. H. Madni,
puting 360-degree compared,’’ in Proc. Grid Comput. Environ. Workshop, ‘‘Secure scientific applications scheduling technique for cloud computing
Nov. 2008, pp. 1–10. environment using global league championship algorithm,’’ PLoS ONE,
[6] A. Masood, ‘‘HETS: Heterogeneous edge and task scheduling algorithm vol. 7, Jul. 2016, Art. no. e0158102.
for heterogeneous computing systems,’’ in Proc. IEEE 17th Int. Conf. [27] S. H. H. Madni, M. S. A. Latiff, M. Abdullahi, S. M. Abdulhamid, and
High-Perform. Comput. Commun., Aug. 2015, pp. 1865–1870. M. J. Usman, ‘‘Performance comparison of heuristic algorithms for task
[7] R. Hoffmann, A. Prell, and T. Rauber, ‘‘Dynamic task scheduling scheduling in IaaS cloud computing environment,’’ PLoS ONE, vol. 12,
and load balancing on cell processors,’’ in Proc. 18th Euromicro no. 5, May 2017.
Int. Conf. Parallel, Distrib. Netw.-Based Process. (PDP), Feb. 2010, [28] B. A. Al-Maytami, P. Fan, and A. Hussain, ‘‘I-MMST: A new task schedul-
pp. 205–212. ing algorithm in cloud computing,’’ in Proc. Int. Conf. Intell. Comput.
[8] E. U. Munir, J. Li, and S. Shi, ‘‘QoS sufferage heuristic for independent Cham, Switzerland: Springer, 2018.
task scheduling in grid,’’ Inf. Technol. J., vol. 6, no. 8, pp. 1166–1170, [29] G. Wang and H. C. Yu, ‘‘Task scheduling algorithm based on improved
Aug. 2007. min-min algorithm in cloud computing environment,’’ Appl. Mech. Mater.,
[9] R. K. Bawa and G. Sharma, ‘‘Modified min-min heuristic for job schedul- vol. 303, pp. 2429–2432, Feb. 2013.
ing based on QoS in grid environment,’’ in Proc. 2nd Int. Conf. Inf. Manage. [30] Y. Zhang and B. Xu, ‘‘Task scheduling algorithm based-on QoS con-
Knowl. Economy (IMKE), Dec. 2013, pp. 166–171. strains in cloud computing,’’ Int. J. Grid Distrib. Comput., vol. 8, no. 6,
[10] J. Napper and P. Bientinesi, ‘‘Can cloud computing reach the top500?’’ pp. 269–280, 2015.
in Proc. Combined Workshops Unconventional High Perform. Com- [31] Y. Mao, X. Chen, and X. Li, ‘‘Max–min task scheduling algorithm for load
put. Workshop Plus Memory Access Workshop, Ischia, Italy, May 2009, balance in cloud computing,’’ in Proc. Int. Conf. Comput. Sci. Inf. Technol.
pp. 17–20. New Delhi, India: Springer, 2014, pp. 457–465.
[11] W. E. Dong, W. Nan, and L. Xu, ‘‘QoS-oriented monitoring model of cloud [32] X. Li, Y. Mao, X. Xiao, and Y. Zhuang, ‘‘An improved max-min task-
computing resources availability,’’ in Proc. Int. Conf. Comput. Inf. Sci., scheduling algorithm for elastic cloud,’’ in Proc. Int. Symp. Comput.,
Jun. 2013, pp. 1537–1540. Consum. Control, Jun. 2014, pp. 340–343.
[12] C. Zhang, R. Huang, and J. Zhang, ‘‘Distributed adaptive consen- [33] H. Han, Q. Deyui, W. Zheng, and F. Bin, ‘‘A Qos guided task
sus tracking of unknown heterogenous linear systems via output feed- scheduling model in cloud computing environment,’’ in Proc.
back,’’ in Proc. 35th Chin. Control Conf., Chengdu, China, Jul. 2016, 4th Int. Conf. Emerg. Intell. Data Web Technol., Sep. 2013,
pp. 8008–8013. pp. 72–76.
[13] Z. Beheshti and S. M. Shamsuddin, ‘‘A review of population-based meta- [34] S. Henning, K. Jansen, M. Rau, and L. Schmarje, ‘‘Complexity and inap-
heuristic algorithms,’’ Int. J. Adv. Soft Comput. Appl., vol. 5, no. 5, pp. 1–5, proximability results for parallel task scheduling and strip packing,’’ in
2013. Proc. Int. Comput. Sci. Symp. Russia. Cham, Switzerland: Springer, 2018,
[14] C. Feng, H. Xu, and B. Li, ‘‘An alternating direction method approach to pp. 169–180
cloud traffic management,’’ Jul. 2014, arXiv:1407.8309. [Online]. Avail- [35] J. Wolf, N. Bansal, K. Hildrum, S. Parekh, D. Rajan, R. Wagle,
able: https://arxiv.org/abs/1407.8309 K.-L. Wu, and L. Fleischer, ‘‘SODA: An optimizing scheduler for
[15] S. Begum and P. C S R, ‘‘Stochastic based load balancing mecha- large-scale stream-based distributed computer systems,’’ in Proc. 9th
nism for non-iterative optimization of traffic in cloud,’’ in Proc. Int. ACM/IFIP/USENIX Int. Conf. Middleware, Leuven, Belgium, Dec. 2008,
Conf. Wireless Commun., Signal Process. Netw. (WiSPNET), Mar. 2016, pp. 306–325.
pp. 1249–1254. [36] O. Sinnen and L. A. Sousa, ‘‘Communication contention in task schedul-
[16] A. V. Smirnov, K. A. Borisenko, A. V. Shorov, and E. S. Novikova, ing,’’ IEEE Trans. Parallel Distrib. Syst., vol. 16, no. 6, pp. 503–515,
‘‘Network traffic processing module for infrastructure attacks detection in Jun. 2005.
cloud computing platforms,’’ in Proc. 16th IEEE Int. Conf. Soft Comput. [37] V. Ruben den Bossche, K. Vanmechelen, and J. Broeckhove, ‘‘Cost-
Meas. (SCM), May 2016, pp. 199–202. optimal scheduling in hybrid IaaS clouds for deadline constrained work-
[17] L. Kang and X. Ting, ‘‘Application of adaptive load balancing algorithm loads,’’ in Proc. IEEE 3rd Int. Conf. Cloud Comput., Miami, FL, USA,
based on minimum traffic in cloud computing architecture,’’ in Proc. Int. Jul. 2010, pp. 228–235.
Conf. Logistics, Inform. Service Sci. (LISS), Jul. 2015, pp. 1–5. [38] V. Ruben den Bossche, K. Vanmechelen, and J. Broeckhove, ‘‘Cost-
[18] R. Sahu and A. K. Chaturvedi, ‘‘Many-objective comparison of twelve efficient scheduling heuristics for deadline constrained workloads on
grid scheduling heuristics,’’ Int. J. Comput. Appl., vol. 13, no. 6, pp. 9– hybrid clouds,’’ in Proc. 3rd IEEE Int. Conf. Cloud Computing Technol.
17, Jan. 2011. Sci., Athens, Greece, Nov./Dec. 2011, pp. 320–327.
[19] T. Amudha and T. T. Dhivyaprabha, ‘‘QoS priority based scheduling algo- [39] J. Li, S. Su, X. Cheng, Q. Huang, and Z. Zhang, ‘‘Cost-conscious
rithm and proposed framework for task scheduling in a grid environment,’’ scheduling for large graph processing in the cloud,’’ in Proc. IEEE Int.
in Proc. Int. Conf. Recent Trends Inf. Technol. (ICRTIT), Tamil Nadu, India, Conf. High Perform. Comput. Commun., Banff, AB, Canada, Sep. 2011,
Jun. 2011, pp. 650–655. pp. 808–813.
[20] G. Patel, R. Mehta, and U. Bhoi, ‘‘Enhanced load balanced min-min [40] S. Anousha, S. Anousha, and M. Ahmadi, ‘‘A new heuristic algorithm for
algorithm for static meta task scheduling in cloud computing,’’ Procedia improving total completion time in grid computing,’’ in Multimedia and
Comput. Sci., vol. 57, pp. 545–553, Jan. 2015. Ubiquitous Engineering. Berlin, Germany: Springer, 2014, pp. 17–26.
[21] J. K. Konjaang, J. Y. Maipan-uku, and K. K. Kubuga, ‘‘An efficient [41] J. Y. Maipan-uku, A. Muhammed, A. Abdullah, and M. Hussin, ‘‘Max-
max-min resource allocator and task scheduling algorithm in cloud com- average: An extended max-min scheduling algorithm for grid computing
puting environment,’’ Nov. 2016, arXiv:1611.08864. [Online]. Available: environtment,’’ J. Telecommun., Electron. Comput. Eng., vol. 8, no. 6,
https://arxiv.org/abs/1611.08864 pp. 43–47, Sep. 2016.
[22] X. He, X. Sun, and G. Von Laszewski, ‘‘QoS guided min-min heuristic for [42] R. Pratap and T. Zaidi, ‘‘Comparative study of task scheduling algorithms
grid task scheduling,’’ J. Comput. Sci. Technol., vol. 18, no. 4, pp. 442–451, through Cloudsim,’’ in Proc. 7th Int. Conf. Rel., Infocom Technol. Optim.
Jul. 2003. (ICRITO), Aug. 2018, pp. 397–400

VOLUME 7, 2019 160925

 B. A. Al-Maytami et al.: Task Scheduling Algorithm With Improved Makespan

[43] D. Wu, ‘‘Cloud computing task scheduling policy based on improved [49] T. Baker, J. M. García-Campos, D. G. Reina, S. Toral, H. Tawfik,
particle swarm optimization,’’ in Proc. Int. Conf. Virtual Reality Intell. Syst. D. Al-Jumeily, and A. Hussain, ‘‘GreeAODV: An energy efficient
(ICVRIS), Aug. 2018, pp. 99–101. routing protocol for vehicular ad hoc networks,’’ in Proc. Int.
[44] K. Shimada, I. Taniguchi, and H. Tomiyama, ‘‘Work-in-progress: Conf. Intell. Comput. Cham, Switzerland: Springer, Aug. 2018,
Communication-aware scheduling of data-parallel tasks,’’ in Proc. Int. pp. 670–681.
Conf. Compil., Archit. Synth. Embedded Syst. (CASES), Sep./Oct. 2018, [50] I. A. Ridhawi, M. Aloqaily, Y. Kotb, Y. Jararweh, and T. Baker,
pp. 1–2. ‘‘A profitable and energy- efficient cooperative fog solution for IoT ser-
[45] S. Yang and Q. Deyu, ‘‘Study on static task scheduling based on heteroge- vices,’’ IEEE Trans. Ind. Informat., to be published.
neous multi-core processor,’’ in Proc. Int. Conf. Comput. Netw., Electron. [51] S. Oueida, Y. Kotb, M. Aloqaily, Y. Jararweh, and T. Baker, ‘‘An edge
Automat. (ICCNEA), Sep. 2017, pp. 180–182. computing based smart healthcare framework for resource management,’’
[46] K. Baital and A. Chakrabarti, ‘‘Dynamic scheduling of real-time tasks in Sensors, vol. 18, no. 12, p. 4307, Dec. 2018.
heterogeneous multicore systems,’’ IEEE Embedded Syst. Lett., vol. 11.1, [52] M. Al-Khafajiy, T. Baker, H. Al-Libawy, Z. Maamar, M. Aloqaily, and
pp. 29–32, Mar. 2018. Y. Jararweh, ‘‘Improving fog computing performance via fog-2-fog
[47] T. Baker, B. Al-Dawsari, H. Tawfik, D. Reid, and Y. Ngoko, ‘‘GreeDi: An collaboration,’’ Future Gener. Comput. Syst., vol. 100, pp. 266–280,
energy efficient routing algorithm for big data on cloud,’’ Ad Hoc Netw., Nov. 2019.
vol. 35, pp. 83–96, Dec. 2015. [53] Y. Kotb, I. Al Ridhawi, M. Aloqaily, T. Baker, Y. Jararweh, and H. Tawfik,
[48] T. Baker, M. Asim, H. Tawfik, B. Aldawsari, and R. Buyya, ‘‘An energy- ‘‘Cloud-based multi-agent cooperation for IoT devices using workflow-
aware service composition algorithm for multiple cloud-based IoT appli- nets,’’ J. Grid Comput., pp. 1–26, 2019.
cations,’’ J. Netw. Comput. Appl., vol. 89, pp. 96–108, Jul. 2017.

160926 VOLUME 7, 2019