You are on page 1of 6

2020 2nd International Conference on Sustainable Technologies for

Industry 4.0 (STI), 19-20 December, Dhaka

Performance Analysis and Comparison Among


Different Task Scheduling Algorithms in Cloud
Computing
2020 2nd International Conference on Sustainable Technologies for Industry 4.0 (STI) |978-1-6654-0489-1 /20/$31.00 ©2020 IEEE | DOI: 10.1109/STI50764.2020.9350466

Md. Tanvir Alam Siddique, Selina Sharmin, Tanvir Ahammad


Department of Computer Science and Engineering, Jagannath University, Dhaka, Bangladesh
tanviralamsiddique0@gmail.com, {selina, tanvir}@cse.jnu.ac.bd

major research area in cloud computing. Task scheduling in


Abstract—In the field of cloud computing, efficient task cloud computing defines how to map a set of tasks to a set
scheduling plays a vital role. Cloud computing is an on-demand of virtual machines. An effective task scheduling algorithm
service where resources are shared virtually by the users to
complete their tasks. The main concept of task scheduling is helps us to increase overall system performances. Moreover,
to execute the tasks in such an order that can enhance the due to poor task scheduling, most of the time cloud resources
performance with full resource utilization and minimum loss. are underutilized. In the literature, various task scheduling
Already a lot of scheduling algorithm exists targeting to various again identifies the most suitable scheduling algorithm under
goals like reducing execution time, cost, makespan and increase different user requirements or criteria and performance
resource utilization, load balancing, etc. In this work, we have
mainly carried out the performance analysis and comparison parameters throws a huge research challenge. Therefore,
among various well-known task scheduling algorithms. The a proper comparative analysis among task scheduling can
algorithms are examined and implemented in a well-known assist users and providers to enhance the system performances.
cloud simulator (CloudSim) to observe their performances by
varying different parameters. Finally, the performance metrics In the literature, some works have been done to observe the
like makespan, waiting time, turnaround time, throughput, and
load balancing are observed by plotting the results graphically. performances among the different task scheduling algorithms.
Along with the comparative analysis, this work also assists one In [4] [5] [6] the authors make analysis between the max-min
to identify a better scheduling algorithm based on different and min-min algorithms. Analysis and comparison among the
scheduling constraints and criteria. SJF, FCFS, max-min, min-min task scheduling algorithms
Index Terms—Cloud computing, Comparative analysis, are done in [7] [8]. In [17] load balancing of Round-robin
CloudSim, Load balancing, Makespan, Task Scheduling, Virtu-
alization, Datacenter.
(RR) is compared with max-min and min-min algorithms.
In [18] details about particle swarm optimization based task
scheduling (PSO) and various aspects of PSO are described.
I. I NTRODUCTION
However, most of the works have done the theoretical
Cloud computing is an on-demand service that offers analysis. Without a proper graphical analysis sometimes it
shared services, applications, infrastructure, and other devices is hard for the users to get a proper analysis of the algorithms.
on a pay-as-you-go basis depending on the requirements of
the customer [1]. In this smart era, sustainable development In this work, we have studied six well-known task
is not possible without cloud computing. Cloud computing scheduling algorithms and implement them on cloud
is crucial for industry revolution 4.0 cause it helps to collect simulation toolkit (CloudSim). Although the algorithms are
and consolidate the industry’s data while also providing quite not up to date, still they are widely used. Varying the
an open-source collaboration platform to speed up and number of tasks and virtual machines some performance
optimize research for whole industry profits. Normally, cloud metrics (makespan, waiting time, throughput, turnaround
computing provides us three main services SaaS (Software time and load balancing) are observed. Finally, the results
as a Service), IaaS (Infrastructure as a Service), and PaaS are plotted graphically and performance analysis, as well
(Platform as a Service) [1]. as comparison among those algorithms, have done so that
one can easily able to identify which algorithm is better for
Millions of physical machines operate digitally in cloud specific requirements or criteria.
technology, as well as the cloud provides the illusion of
unlimited resources that can be continuously scaled up or The remainder of this paper is arranged as follows, Section
scaled down by using virtualization. Since cloud services II outlines the relevant works of comparison-based scheduling
are shared by a large number of clients and their tasks are algorithms in the cloud computing setting and Section III
submitted to the cloud, planning and carrying out these tasks provides an overview of our task scheduling algorithms that
has just become a difficulty while maintaining the various have been implemented. Section IV discusses the experimental
user requirements and QoS. Therefore, task scheduling is a findings and the comparative study and ultimately, section V

978-1-6654-0489-1/20/$31.00©20XX IEEE

Authorized licensed use limited to: BOURNEMOUTH UNIVERSITY. Downloaded on June 27,2021 at 06:27:46 UTC from IEEE Xplore. Restrictions apply.
discusses the conclusion and future work respectively. parameters. Also plotting the results graphically would aid
the users to make a quick and efficient decision to fulfill their
II. R ELATED W ORKS demands.
Most of the existing works in literature deal with generating
III. OVERVIEW OF S ELECTED TASK S CHEDULING
new scheduling algorithms. However, very few works have
A LGORITHMS
been done to provide analysis and comparison among the
task scheduling algorithms. In [4], a comparative study of We have described the scheduling processes, parameters,
Min-Min and Max-Min algorithms is based on the makespan and other features of our task scheduling algorithms imple-
parameter that was conducted by the authors. A static method mented and compared in the cloud computing environment in
and task-independent and batch mode scheduling are used by this section. There are three main strategies in task scheduling.
both methods. Sanjay et al. in [5], the Min-Min and Max-Min Firstly, in-network properties and their associated state data
algorithms have been used for job scheduling in the cloud will be collected. Then, based on various criteria, the target
environment. They have used time-shared and space-shared resource is chosen. Finally, it submits the task to the chosen
virtual machines models. Tested in CloudSim, they would resource [2].
like to demonstrate that the Max-min approach is more A. First Come First Serve(FCFS)
effective than other strategies in space-shared mode, and the
Min-Min approach provides minimal task processing delays First Come First Serve means that the jobs are done
and Max-Min gives appropriate load balancing. according to the time of arrival of the task order. The FCFS
scheduling strategy for internal job scheduling is supported by
M. Joundy et al. [7] have shown a comparative study the CloudSim toolkit. A clear policy that assigns a VM to the
among two hybrids, SJF, RR, Max-Min, and Min-Min host on an FCFS basis is the default policy enforced by VM
scheduling algorithms in graphically. To determine their provisioning [15]. The queue is handled by the (first in first to
efficiency, they checked those methods using 3 distinct data go) FIFO process. Quick jobs at the end of a queue can wait
sets. Their experimental findings ensure that the Min-Min until a large task is done at the front of the queues [20] . It is
algorithm’s dominance over state-of-the-art algorithms and straightforward and simple to understand and non-preemptive.
SJF demonstrates a reasonable waiting period, but there is B. Shortest Job First(SJF)
more response time. In [8] a comparative analysis among SJF,
Shortest Job First Scheduling is used by putting the shorter
FCFS, Max-Min, and Min-Min scheduling algorithms has
task in front of the queue and the longer tasks at the end of
done based on mean, max, and variance execution time and
a queue to order a series of tasks. This minimizes the overall
load balancing graphically. They further show that Min-Min
waiting period. It also suffers from starvation for longer tasks
achieves better makespan compared to other methods.
when a large number of small jobs are open. It will not execute
if a job with a greater burst time has the highest priority as it
In papers [9] [10] [11] [12] [13] [14] [15] [16] a number
is a non-preemptive scheduling algorithm. In severe situations,
of existing algorithms (FCFS, SJF, RR, PSO etc.) for job
the tasks with big service time can never be served if the
scheduling are compared and tabulated with each other
system has little idle time [7].
through their effects or evaluated theoretically [17]. It helps
to consider the wide variety of choices for scheduling for C. Round Robin Scheduling Algorithm(RR)
selecting one for a particular environment. However, graphical One of the easiest, most traditional, and most used task
observation is absent here. scheduling is known to be the Round Robin algorithm. It
transfers the load to all the resources evenly. The big advantage
Valarmathi et al in [18] detailed theme was given about the is that tasks are performed wisely in a series of turns and there
Particle Swarm Optimization (PSO) based algorithm and its is a necessity to late until the past tasks is executed. Therefore,
several task allocation variants in a cloud environment. PSO it has no starvation problem. But the poor hand is, it consumes
produces better results for task Scheduling Problems. This a lot of effort to achieve all the jobs if the workload is high
paper offers the researchers insight into different dimensions and the queue is filled, and a perfectly appropriate quantum
of PSO and Task Scheduling. In [19], centered on the of time is difficult to determine [9]. The use of too short
PSO methodology, they proposed a multi-objective process quantum time induces too many context switches and reduces
scheduling scheme. An effective encoding method with the performance of the CPU. Using too long time quantum,
a weighted linear transforming fitness value that contains poor response time like FCFS in some situations [7].
two metrics of scheduling, i.e. make-pan and utilization of
resources. They showed that the scheduling of workflows D. Min-Min scheduling Algorithm
based on PSO outshines GA over multiple parameters. This algorithm is split into 2 stages. The estimated com-
pletion time of each assignment in a metatask is determined
While these works provide comparison and analysis of for each resource in the first stage. In the second stage, the
different task scheduling algorithms, in this work we aim task is chosen and assigned to the corresponding element by
to analyze them by implementing them based on different shortest estimated completion time. Then the chosen task is

Authorized licensed use limited to: BOURNEMOUTH UNIVERSITY. Downloaded on June 27,2021 at 06:27:46 UTC from IEEE Xplore. Restrictions apply.
erased from the meta-list. This process is repeated before cloudlet length is generated randomly so the value of all
you map all the tasks in the metatask [4]. This raises the performance metrics is changed randomly in a certain range
cumulative completion time of all the tasks and thereby raises based on specific algorithm’s processes. For this cause, we
the makespan. To finish their operation, the long tasks have to determine the average value for all performance metrics for
pause for smaller tasks. It causes for unbalancing load in Min- doing the comparative analysis. For getting the average value,
Min [20]. The min-min method is easy, but at the time of the we run each algorithm 10 to 3000 times. For the simulation
number of tasks in the metatask is small relative to large tasks, analysis, we observe the throughput, waiting time, makespan,
it gives instant results. On the other hand, if the number of turnaround time and load balancing by varying the number
smaller tasks is overlaid by large-size, it offers bad utilization of tasks, and virtual machines.
of resources and bot good makespan since large tasks have to
pause until all low size tasks to be done [4]. Throughput is calculated as the total number of tasks per
second that are completed. The average throughput of all
E. Max-Min Scheduling Algorithm compared task scheduling is shown in Figure 1. In comparison
Max-Min resolves some problem of the Min-Min algorithm to the remaining methods, higher throughput is found for PSO
(when the number of large tasks is higher than the small task and Min-Min. Max-Min also provides medium-throughput as
then larger makespan). The Max-Min algorithm works the compared to the other five. Here, the PSO executes 19 tasks
same procedures except in the second level, the key difference per 10 seconds and Min-Min executes 17 tasks per 10 seconds
is that a task is chosen that has the largest time of completion considering 5 VM and 15 tasks for each algorithm. PSO can
rather than the minimum completion time as in min-min and find near-optimal solutions for mapping all tasks to the set of
allocated to a element that offers the less time of completion. available resources so it achieves high throughput. For Min-
This procedure is iterated until all the tasks are mapped or the Min, there are only smaller tasks on quicker machines and the
meta task becomes null [4]. Therefore, the larger tasks can be throughput is improved here. Among those methods, FCFS,
performed first on quicker machines in max-min, and smaller SJF and RR achieve less throughput.
tasks can be performed on other possible machines in parallel,
resulting in a better make-pan and balanced workload than the
previous process [11].

F. Particle Swarm Optimization Algorithm(PSO)


The technique is based on observations of swarms such
as schools of fish and bird flocking. The output is extracted
by the best fitness according the function of fitness. Finally,
the best solution is extracted from the fitness values of each
particle and the strongest optimal solution is assumed to be the
maximum value. For the practice population, it starts with free
parameters of velocity and position and tasks are regarded as
particles here. Like GA, the fitness function tests each particle Fig. 1. Average throughput Comparison among six algorithms
[20]. In the late iterative procedure, it can drop into local
optima and have a poor rate of convergence. PSO is the easiest Load balancing is an important factor in cloud computing.
implementation and a polynomial connected to the size of the Load balancing means, how to distribute the workloads
problem is its computational complexity. The traffic workload among different cloud resources. Figure 2 illustrates the
using PSO is balanced and can be used for any number of tasks average load distribution of all the compared algorithms. It
and resources, and almost optimal solutions can be identified shows that RR, Max-Min, then PSO perform proper load
for mapping all tasks to the collection of resources available balance compared to the rest of the methods. Max-Min
[20]. gives better load balancing since smaller jobs are executed
concurrently while other longer jobs are executed. FCFS,
IV. E XPERIMENTAL R ESULTS AND C OMPARATIVE SJF and Min-Min achieve worst load balancing. In Min-Min
P ERFORMANCE A NALYSIS and SJF, the long tasks have to wait for smaller tasks to end
We are using the CloudSim simulation toolkit for the their execution so the workload can’t be distributed properly
comparative performance study of the selected approaches. means an Unbalanced load.
Our cloudlet (task) properties are: number of CPU is 1,
file-size is 300, the output size is 300 and the cloudlet length Makespan is the time differential between a series of tasks
is generated randomly in CloudSim. And VM’s properties beginning and finishing. Figure 3 and Figure 4 illustrate
are: 512 MB RAM, 250 MIPS (million instruction per the average makespan of all the compared algorithms. The
second), bandwidth = 1000, 1 CPU and VMM name is value of the average makespan for each algorithm has been
”Xen”. Those all properties are the same for our selected calculated after running the code 3000 times in a loop, for
algorithms at the time of simulation in CloudSim. Since better accuracy. Figure 3 shows the average makespan by

Authorized licensed use limited to: BOURNEMOUTH UNIVERSITY. Downloaded on June 27,2021 at 06:27:46 UTC from IEEE Xplore. Restrictions apply.
TABLE II
AVERAGE M AKESPAN T IME ( VARYING THE NUMBER OF VM’ S )

Number Average Makespan Time (in millisecond)


of VM’s SJF RR Min-Min Max-Min FCFS PSO
5 60259 65892 54116 50236 70498 45896
10 35986 38952 31871 28659 41875 27123
15 24589 29865 24040 22569 31880 16985
20 23594 24563 20031 16359 26572 15264
25 17865 18659 17524 14569 23261 11456
30 16589 20356 15854 12594 21040 11256

Fig. 2. Average load distribution Comparison among six algorithms

changing the number of tasks and generated from the value of


Table I. In this figure, PSO had proved superiority among those
algorithms by considering makespan, because PSO can find
near-optimal solutions. Max-Min has also reduced makespan
since smaller jobs are executed concurrently while other longer
jobs are executed. FCFS and RR (for a large number of
tasks, more context switch, and increase makespan) get the
highest makespan compared to the rest of the methods. Figure
4 shows the average makespan by changing the number of
VMs and generated from the value of Table II. Here we see Fig. 4. Average makespan Comparison by changing the number of VMs
that, if the number of VM’s is increased then the makespan
is decreased noticeably. In this figure, in the same way, PSO Figure 5 and Figure 6 illustrate the average waiting time
gives minimum makespan and then, the Max-Min gives better of all the compared algorithms. We run the code 3000 times
makespan. in a loop for getting this. Figure 5 shows by changing the
number of tasks and Figure 6 shows by changing the number
TABLE I of VMs. It is observed from the results in Figure 5 and Figure
AVERAGE M AKESPAN T IME ( VARYING THE NUMBER OF TASKS ) 6, PSO outperforms compared to the other five algorithms
because PSO can find near-optimal solutions by evaluating the
n. of Average Makespan Time (in millisecond) fitness value of each task effectively. Min-Min also provides
Tasks SJF RR MinMin MaxMin FCFS PSO acceptable waiting time because the shortest tasks are assigned
15 10256 11420 9429 8856 12278 8235 in the fastest machine. SJF also provides less waiting time
30 17256 19853 16272 14895 21306 14025
60 32598 34895 29240 27896 38360 23985 because it reduces the waiting time of the short jobs though it
120 60259 65986 53969 50289 70261 45896 increases the waiting time of the long jobs. On the other hand,
240 86957 120569 102029 90258 131665 81256 FCFS and RR give the highest average waiting time among
480 205694 235589 193847 165982 250000 155892
those six algorithms.

Fig. 5. Average waiting time Comparison by changing the number of tasks


Fig. 3. Average makespan Comparison by changing the number of tasks

Authorized licensed use limited to: BOURNEMOUTH UNIVERSITY. Downloaded on June 27,2021 at 06:27:46 UTC from IEEE Xplore. Restrictions apply.
V. C ONCLUSION
In this research, we compared six traditional but widely
used task scheduling algorithms in the cloud computing en-
vironment. Finally, from the experimental assessment, we
can infer that the PSO performed successfully and won the
competition against the six existing methods implemented,
taking into account the average waiting time, turnaround
time, makespan, load balancing, and throughput. When we
consider load balancing then the Max-min, PSO and RR can be
better choices. Since the execution of the task is performed a
cyclic way, so RR achieves acceptable load balancing and less
Fig. 6. Average waiting time Comparison by changing the number of VMs response time but it is not considering other criteria. Min-Min
and SJF showing the good waiting time but its bottleneck is
a starvation problem. Max-Min also provides better makespan
We have also observed the turnaround time that indicates
and turnaround time. In the future, in the cloud computing
the summation of waiting time and the execution time for
world, we want to conduct performance analysis and com-
a specific task. Figure 7 and Figure 8 shows the average
parative study of newly developed and successful scheduling
turnaround time of all the compared algorithms. For getting
algorithms.
this we run our code 3000 times in a loop. Figure 7 has shown
by changing the number of tasks and Figure 8 has shown by R EFERENCES
changing the number of VMs. It is observed from the results
in Figure 7 and Figure 8, PSO outperforms compared to the [1] Peter Mell and Tim Grance, “The nist definition of cloud computing”,
2011.
other five algorithms. Max-Min also provides less turnaround [2] A. Gupta, H. S. Bhadauria, A. Singh and J. C. Patni, ”A theo-
time since smaller jobs are executed concurrently while other retical comparison of job scheduling algorithms in cloud computing
longer jobs are executed. On contrary, the FCFS, SJF, RR and environment”, 2015 1st International Conference on Next Generation
Computing Technologies (NGCT), Dehradun, 2015, pp. 16-20, doi:
Min-Min give the highest turnaround time. 10.1109/NGCT.2015.7375074.
[3] Malhotra, Agarwal, and Jaiswal, “Virtualization in cloud com-
puting”, Journal of Information Technology Software Engineering,
doi:10.4172/2165-7866.1000136, vol. 04, 2014.
[4] Neha Sharma and D. Sanjay Tyagi, “A comparative analysis of min-min
and max-min algorithms based on the makespan parameter”, Interna-
tional Journal of Advanced Research in Computer Science, vol. 8, no.
03, April 2017.
[5] Sanjay Kumar and Atul Mishra, “Application of min-min and max-
min algorithm for task scheduling in cloud environment under time
shared and space shared vm models”, International Journal of Computing
Academic Research (IJCAR), vol. 04, no. 06, pp. 182–190, December
2015.
[6] R. Jemina Priyadarsini and L. Arockiam, “Performance evaluation of
min-min and max-min algorithms for job scheduling in federated cloud”,
International Journal of Computer Applications (0975 â 8887), vol. 99,
no. 18, August 2014.
Fig. 7. Average turnaround time Comparison by changing the number of [7] M. Joundy, S. Sarhan, and S. Elmougy, “Task scheduling algorithms
tasks in cloud computing: A comparative study”, International Journal of
Intelligent Computing And Information Science.
[8] Essaies Meriam and Nabil Tabbane, “A survey on cloud computing
scheduling algorithms”, 2016 Global Summit on Computer Information
Technology, DOI: 10.1109/GSCIT.2016.6, 2017.
[9] Abhishek Gupta, H. S. Bhadauriay, Annapurna Singhz, and Jagdish
Chandra Patnix, “A theoretical comparison of job scheduling algorithms
in cloud computing environment”, 2015 1st International Conference on
Next Generation Computing Technologies (NGCT-2015), 4-5 September
2015.
[10] Arnav Wadhonkar and Deepti Theng, “A survey on different schedul-
ing algorithms in cloud computing”, International Conference on Ad-
vances in Electrical, Electronics, Information, Communication and Bio-
Informatics (AEEICB16), 2016.
[11] Teena Mathew, K. Chandra Sekaran, and John Jose, “Study and
analysis of various task scheduling algorithms in the cloud comput-
ing environment”, International Conference on Advances in Comput-
ing,Communications and Informatics (ICACCI), IEEE, 2014.
[12] Aswathy B Namboothiri and D. R. Joshua Samuel Raj, “A comparative
Fig. 8. Average turnaround time Comparison by changing the number of study on job scheduling algorithm augmenting load balancing in cloud”,
VMs Second International Conference on Science Technology Engineering
And Management(ICONSTEM), IEEE, 2016.

Authorized licensed use limited to: BOURNEMOUTH UNIVERSITY. Downloaded on June 27,2021 at 06:27:46 UTC from IEEE Xplore. Restrictions apply.
[13] V.Seethalakshmi, D. V. Govindasamy, D. V. Akila, G. Sivaranjini, K.
Sindhuja, and K. Prasanth, “A survey of different workflow scheduling
algorithms in cloud computing”, International Conference on Compu-
tation of Power,Energy, Information and Communication (ICCPEIC),
IEEE, 2019.
[14] Arghavan Keivani and Jules-Raymond Tapamo, “Task scheduling in
cloud computing: A review”, IEEE, 2019.
[15] Kavyasri M N and D. B. Ramesh, “Comparative study of scheduling
algorithms to enhance the performance of virtual machines in cloud
computing”, 2016.
[16] Yash P. Dave, Avani S. Shelat, Dhara S. Patel, and Rutvij H. Jhaveri,
“Various job scheduling algorithms in cloud computing: A survey”, in
ICICES2014 - S.A.Engineering College, Chennai, Tamil Nadu, India,
IEEE, 2014.
[17] Graves and Alex, “Load balancing strategy using round robin algo-
rithm”, Asia-pacific Journal of Convergent Research Interchange, vol.
02, no. 03, pp. 1–10, September 30, 2016.
[18] R. Valarmathi and D. T. Sheela, “A comprehensive survey on task
scheduling for parallel workloads based on particle swarm optimisation
under cloud environment”, Second International Conference On Com-
puting and Communications Technologies(ICCCT-17), 2017.
[19] Shubham, R. Gupta, V. Gajera, and P. K. Jana, “An effectivemulti-
objective workflow scheduling in cloud computing: A pso based ap-
proach”, Ninth International Conference on Contemporary Computing
(IC3), 2016, pp. 1–6.
[20] Ruba Abu Khurma, Heba Al Harahsheh, and Ahmad Sharieh, “Task
scheduling algorithm in cloud computing based on modified round robin
algorithm”, Journal of Theoretical and Applied Information Technology,
vol. 96, no. 17, pp. 5869–5888, 15th September 2018.

Authorized licensed use limited to: BOURNEMOUTH UNIVERSITY. Downloaded on June 27,2021 at 06:27:46 UTC from IEEE Xplore. Restrictions apply.

You might also like