Heuristic Initialization of PSO Task Scheduling Al
Heuristic Initialization of PSO Task Scheduling Al
a r t i c l e i n f o a b s t r a c t
Article history: Task scheduling is one of the major issues in cloud computing environment. Efficient task scheduling is
Received 21 July 2020 substantial to attain cost-effective execution and improve resource utilization. The task scheduling prob-
Revised 3 November 2020 lem is classified to be a nondeterministic polynomial time (NP)-hard problem. This feature attracts
Accepted 4 November 2020
researchers to utilize nature inspired metaheuristic algorithms. Initializing searching solutions randomly
Available online xxxx
is one of the key features in such optimization algorithms. However, assisting metaheuristic algorithms
with effective initialized solutions can significantly improve its performance. In this paper, an improved
Keywords:
initialization of particle swarm optimization (PSO) using heuristic algorithms is proposed. Longest job to
Particle swarm optimization
Task scheduling
fastest processor (LJFP) and minimum completion time (MCT) algorithms are used to initialize the PSO.
Cloud computing The performance of the proposed LJFP-PSO and MCT-PSO algorithms are evaluated in minimizing the
Metaheuristic algorithms makespan, total execution time, degree of imbalance, and total energy consumption metrices.
Load balancing Moreover, the performance of the proposed algorithms is compared with recent task scheduling meth-
Virtual machines ods. Simulation results revealed the effectiveness and superiority of the proposed LJFP-PSO and MCT-
PSO compared to the conventional PSO and comparative algorithms.
Ó 2020 The Authors. Production and hosting by Elsevier B.V. on behalf of King Saud University. This is an
open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
1. Introduction software as a service (SaaS), in which users can use the software
directly from the cloud without the need of installing it locally
The tremendous increase of internet information processing (Devaraj et al., 2020). Cloud providers exploit virtualization tech-
lead to the development of cloud computing system. Cloud com- nology and supply users with computing resources in the form of
puting plays an important role in offering technology services virtual machines (VMs). When multiple users request services
through the internet. It provides users with computer system from the cloud in the form of tasks, an efficient task scheduling
resources, such as data storage and computing power, without is required to improve the overall performance of the cloud com-
direct active management. A cloud can provide three types of ser- puting. Efficient task scheduling allows optimal allocation of
vices related to the infrastructure, platform, and software. The first resources among the requested tasks in a finite time to achieve
service is infrastructure as a service (IaaS) which provides infras- desired quality of service (QoS) (Adhikari et al., 2019). The problem
tructure services such as storage system and computation of optimal scheduling is to build a schedule of tasks to VMs alloca-
resources. Second service is platform as a service (PaaS), which tion subject to some constraints to optimize a given objective.
enables clients to build their applications on the provided plat- Therefore, the task scheduling algorithm is one of the core ele-
form. The third service provides the software to the users termed ments of each cloud infrastructure.
In cloud computing, the main types of scheduling mechanisms
are cloud service scheduling, heuristics scheduling, workflow
⇑ Corresponding author.
scheduling, static scheduling and dynamic scheduling (Arunarani
E-mail addresses: [email protected] (S.A. Alsaidy), amenah.abbood@sc.
et al., 2019). However, scheduling techniques can be classified
uobaghdad.edu.iq (A.D. Abbood), [email protected] (M.A. Sahib).
based on the complexity of the algorithm as heuristic, meta-
Peer review under responsibility of King Saud University.
heuristic, and hybrid task scheduling approaches (Elaziz et al.,
2019). Heuristic algorithms are used for static scheduling such
algorithms include minimum execution time (MET), minimum
Production and hosting by Elsevier completion time (MCT), shortest job to fastest processor (SJFP),
https://doi.org/10.1016/j.jksuci.2020.11.002
1319-1578/Ó 2020 The Authors. Production and hosting by Elsevier B.V. on behalf of King Saud University.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
Please cite this article as: S.A. Alsaidy, A.D. Abbood and M.A. Sahib, Heuristic initialization of PSO task scheduling algorithm in cloud computing, Journal of
King Saud University – Computer and Information Sciences, https://doi.org/10.1016/j.jksuci.2020.11.002
S.A. Alsaidy, A.D. Abbood and M.A. Sahib Journal of King Saud University – Computer and Information Sciences xxx (xxxx) xxx
longest job to fastest processor (LJFP), Min-Min, Max-Min (Elaziz issues are significantly affected by exploring an optimal task
et al., 2019; Mishra et al., 2020). The task scheduling problem is scheduling. A scheduling model in the cloud environment is shown
defined to be a non-deterministic polynomial (NP)-hard optimiza- in Fig. 1.
tion problem (Golchi et al., 2019). Therefore, the need for meta- The datacenter broker (DB) is responsible for detecting and
heuristic algorithms is necessary to deal with such multimodal gathering all information related to the available resources (VMs)
optimization problems. The main metaheuristic algorithms used and all remaining resources that may be available in future. The
are particle swarm optimization (PSO) (Mansouri et al., 2019; DB collect this information with the help of cloud information ser-
Kumar and Sharma, 2018), ant colony optimization (ACO) vices (CIS). The hypervisor or virtual machine monitor (VMM) acts
(Mahato et al., 2017), genetic algorithm (Keshanchi et al., 2017), as an interface between the host operating system and VMs. Tasks
discrete symbiotic organism search (DSOS) (Abdullahi et al., are submitted to a task waiting queue to be scheduled on VMs
2016), gravitational search algorithm (GSA) (Chaudhary and based on a defined scheduling algorithm in DB.
Kumar, 2018). The problem of scheduling involves scheduling a given set of
In order to obtain improved performance, researchers used tasks on a given set of resources in the form of virtual machines
hybrid metaheuristic algorithms. Examples of such hybridization subject to some constraints to optimize some objective function.
include; firefly and PSO (Golchi et al., 2019; Devaraj et al., 2020), In cloud computing, the effectiveness of task scheduling is mea-
Q-learning and PSO (Jena et al., 2020), firefly and simulated anneal- sured by various system performance parameters (metrics). Gener-
ing (Fanian et al., 2018), hybridized whale optimization algorithm ally, these optimization criteria are categorized into two demands
(Strumberger et al., 2019). based on cloud service: cloud users demand and cloud service pro-
In using PSO for task scheduling, the particles are initialized viders demand. On one hand, makespan (MS) and total execution
randomly by allocating each task to one of the available machines time (TET) are user demand criteria. On the other hand, degree of
without constraints. Random initialization is one of the key fea- imbalance (DI) and energy consumption are provider demand cri-
tures of PSO in spreading the particles over the search space. How- teria. In single objective task scheduling, only one criterion is con-
ever, if these random solutions are chosen wisely, the searching sidered, however, in multi-objective criteria, two or more
process of tasks to VMs allocation can be boosted to the optimal parameters are considered as an objective (Belgacem et al., 2020;
solution. Moreover, heuristic algorithms such as MCT, SJFP, LJFP, Devaraj et al., 2020).
Min-Min, and Max-Min are non-optimized basic algorithms. How- In recent years, a considerable amount of literature has been
ever, compared to metaheuristic algorithms, they are simple to published on task scheduling in cloud computing environment.
implement. If the PSO can be assisted by suggesting a good starting These studies can be categorized into two main groups, heuristic
search point, the performance of the algorithm can be improved. or metaheuristic algorithms, and hybrid of heuristic and meta-
In this paper, the proposed method is to utilize heuristic heuristic algorithms. (Kaur and Kaur, 2019) proposed two heuristic
scheduling algorithms to initialize the PSO search process. The metaheuristic hybridized approaches to optimize the load balanc-
two standard algorithms are LJFP and MCT. The proposed com- ing in cloud environment. The two heuristic algorithms are pre-
bined algorithms termed LJFP-PSO and MCT-PSO are expected to dicted earliest finish time (PEFT) and heterogeneous earliest
improve the classical PSO. finish time (HEFT). The output of these two heuristics are
The major contributions of this paper can be summarized as employed to initialize the ACO algorithm parameters to speed up
follows: the convergence rate and reduce randomness. (Golchi et al.,
2019) proposed a hybrid of firefly and improved particle swarm
1. Formulating the problem of task scheduling and defining the optimization (IPSO) algorithm for load balancing and tasks
mathematical model and the objective functions for which the scheduling in cloud computing environment. The random nature
allocation of tasks to VMs is optimized. of the operators of the IPSO algorithm make it sensitive to the ini-
2. Proposing a hybridized task scheduling algorithm and designing tial population. Improper selection of initial population signifi-
it by replacing the random initialization of the metaheuristic cantly effects the algorithm’s convergence. Therefore, firefly
algorithm (selected to be PSO algorithm) using heuristic algo- algorithm is utilized to initialize the IPSO. (Belgacem et al., 2020)
rithms (selected to be MCT and LJFP algorithms) termed as proposed a multi-objective search algorithm called spacing
MCT-PSO and LJFP-PSO. multi-objective antlion algorithm (S-MOAL) then (Abualigah and
3. Evaluating and comparing the performance of the proposed Diabat, 2020) proposed a hybrid multi-objective antlion optimiza-
MCT-PSO and LJFP-PSO with four recent and well-known task tion algorithm (MALO) with elite-based differential evolution (DE)
scheduling algorithms. The evaluation is performed using four for solving task scheduling problems in cloud computing environ-
objective metrics which are makespan (MS), total execution ments. The DE is used to improve the exploitation search-ability of
time (TET), degree of imbalance (DI), and total energy consump- the ALO since the DE algorithm demonstrates the powerful fea-
tion (TEC). tures of the genetic algorithm (GA) and the evolution strategy
(ES). (Elaziz et al., 2019) also used DE algorithm as a local search
The rest of the paper is organized as follows. In Section 2, signif- technique to enhance the exploitation ability of standard moth
icant related works are presented. The problem of task scheduling search algorithm (MSA) and termed it as (MSDE). The MSDE is used
in cloud computing environment is described and modelled in sec- to balance the load of tasks among VMs. (Strumberger et al., 2019)
tion 3. The proposed algorithms are explained in Section 4. Simu- introduced a hybridized whale optimization algorithm (WOA) for
lation results are discussed in Section 5. Finally, Section 6 tackling the resource scheduling problem. The enhanced WOA is
concludes the paper. improved by performing a hybridization with the artificial bee col-
ony (ABC) and firefly algorithms (FA) to balance between exploita-
2. Related work tion (intensification) and exploration (diversification) mechanisms.
The main research works related to this paper are listed and
Researchers attention and concern has been attracted by compared in Table 1.
many issues facing cloud computing. Such major issues include As can be seen from Table 1, the works are based on different
resource management, load balancing, migration to clouds, privacy types of metaheuristic algorithms. However, the common relation-
and security, energy-efficiency, availability and scalability, ship between these related approaches are the use of random pop-
interoperability and compatibility (Al-Arasi and Saif, 2020). These ulation in initializing the metaheuristic and hybrid metaheuristic
2
S.A. Alsaidy, A.D. Abbood and M.A. Sahib Journal of King Saud University – Computer and Information Sciences xxx (xxxx) xxx
algorithms. Generally, the convergence of metaheuristic algo- The ultimate output of the scheduling algorithm is an N t N v
rithms is significantly affected by the initial population. The ran- allocation matrix stating which task should be executed by which
dom nature is a fundamental need to avoid local minimum traps. VM. The allocation matrix is defined to be:
However, if some particles are assisted heuristically in selecting 2 3
x11 x1Nv
an effective or near optimum starting points, then the algorithm 6 . .
convergence can be improved. In this paper, the idea of the pro- X¼6 . . .. 7 7;wherexij ¼
1 if T i is allocated to V j
4 . . . 5 0 if T i is not allocated to V j
posed algorithm is to utilize heuristic algorithms for initializing
xNt 1 xNt Nv
the papulation and hence improve the algorithm’s performance
significantly. ð2Þ
With the condition:
3. Task scheduling problem formulation XNv
x
j¼0 ij
¼ 1For1 i Nt ð3Þ
In cloud computing, tasks are scheduled based on the objective
of improving the performance of various quality of service param- Thus, the j-th virtual machine execution time (VET j ) is given by:
eters. The problem of task scheduling is how to find the optimal XNt
allocation of a number of tasks to a given number of virtual pro- VET j ¼ x
i¼0 ij
ECT ij For1 j Nv ð4Þ
cessing machines. To model this problem, the following assump-
To evaluate the performance of a given scheduling algorithm,
tions are considered:
four metrics will be considered; MS, TET, DI, and TEC which are
given in the following equations respectively:
1- All submitted tasks are independent.
2- Tasks cannot migrate between virtual machines, in other MS ¼ MAXfVET j gFor1 j Nv ð5Þ
words, one task is not allowed to be assigned to more than
one virtual machine. XNv
TET ¼ j¼1
VET j For1 j Nv ð6Þ
3- The virtual machines are heterogeneous having different
processing abilities and different power consumptions.
4- The VMM allows the optimization of the resources utiliza- MS VET min
DI ¼ Nv ð7Þ
tion with a minimum energy consumption. TET
Where VET min is the minimum total execution time between all
Consider N t independent tasks defined by a task length set, machines. During the execution of tasks, a VM can have one of two
T ¼ fT 1 ; T 2 ; ; T Nt g, where T i is the i-th task length in million states; active and idle states. The idle state of a VM consumes 60–
instructions (MI). Each task in T, is bounded by T max and T min , i. 70% of energy consumption of active state of that VM (Jena et al.,
e., T min T i T max .
2020; Mishra et al., 2020). Accordingly, assume virtual machine j
Also, consider N v virtual machines defined by execution rate consumes energy equal to aj (joule/MI) in active state and bj
set, V ¼ fV 1 ; V 2 ; ; V Nv g, where V j is the j-th virtual machine exe-
(joule/MI) in idle state. A VM j reside (VET j ) sec. in active state
cution rate in million instructions per second (MIPS). Each virtual
and (MS VET j ) sec. in idle state, thus, the total energy consump-
machine in V, is bounded by V max and V min , i. e., V min V j V max .
tion can be given as:
Similarly, we define the power set P ¼ fP1 ; P2 ; ; P Nv g, where Pj
XNv
is the j-th virtual machine power consumption in watt bounded TEC ¼ ½½VET j aj þ MS VET j bj V j ð8Þ
j¼1
by Pmax and Pmin . The estimated computation time for executing
task i on virtual machine j is given by: The allocation of tasks stated by the scheduling algorithms are
optimized based on the aforementioned metrics. The following
Ti
ECT ij ¼ ð1Þ section present the method of the proposed scheduling algorithm
Vj along with various standard ones.
3
S.A. Alsaidy, A.D. Abbood and M.A. Sahib Journal of King Saud University – Computer and Information Sciences xxx (xxxx) xxx
Table 1
Comparison of related work.
4. Proposed scheduling algorithm an optimal solution (best particle position) within a predefined
search space. The particle cooperates between each other sharing
Particle swarm optimization is one of the main meta-heuristic information about their current situation in the search process to
techniques utilized in solving optimal task scheduling problem. update their new position. The update equation of the new particle
PSO is an evolutionary computational technique motivated by position is given by:
social behavior of animal species living in large colonies like birds,
ants or fish (Sardaraz and Tahir, 2019). The mathematical mod- pltþ1 ¼ ptl þ v ltþ1 ð9Þ
elling of the PSO algorithm is based on a number of agents grouped
in a swarm called particles. Each particle is considered as a solution where ptl is the l-th particle at iteration t, and v tþ1
l is the velocity of
to the optimization problem. The particles are utilized to search for the l-th particle at iteration t þ 1 defined by:
4
S.A. Alsaidy, A.D. Abbood and M.A. Sahib Journal of King Saud University – Computer and Information Sciences xxx (xxxx) xxx
v tþ1 ¼ xv tl þ c1 r1 pbest
t t
pl þ c2 r 2 pgbest pl ð10Þ expected to improve the classical PSO. The pseudo code of the LJFP,
l l
MCT, and the Proposed LJFP-PSO and MCT-PSO are presented in
Where pbestl is the best position found so far associated with par- algorithms 1, 2, and 3 respectively.
ticle l. pgbest is the best global position found among all pbest l for
(1 l L), where L is the population size (number of particles). Algorithm 1: Pseudo-code of MCT algorithm
c1 and c2 are the acceleration constants, r1 and r 2 are random num- 1 Input:
bers between 0 and 1, x is the inertia factor. The velocity equation 2 No of tasks (Nt),
is influenced by three components; inertia, personal influence, and 3 Task lengths (T),
social influence represented by v tl , pbestl ; and pgbest respectively. 4 No. of VMs (Nv),
For the task scheduling problem, each particle is structured by 5 VMs execution rates (V)
an allocation matrix as given by equation (up) and shown in 6 Output:
Fig. 2. Therefore, the swarm is a three-dimensional matrix of size 7 Allocation matrix (task to VM) of size Nt Nv
ðN t N v LÞ. 8 Start:
The allocation matrix has two main features; First, all elements 9 Calculate the estimated execution time matrix of size
have a binary value of 0 or 1 (xij 2 f0; 1g), second, for a given par- Nt Nv
ticle, each task allocation row contains one element equal to one 10 Initialize the allocation matrix (task to VM) of size Nt Nv
and the rest are 0. For example, as shown in Fig. 2, particle 1 (p1 ) to zeros
assigns task 1 (T 1 ) to machine 1 (V 1 ), task 2 (T 2 ) to machine Nv 11 For i = 1 to no. of tasks Nt
(V Nv ), and task Nt (T Nt ) to machine 2 (V 2 ),. 12 Find the minimum execution time of task i among all VMs
Similarly, the velocity 3D matrix will have the same size of 13 Set Mn = minimum execution time,
swarm such that each velocity window of size ðN t N v Þ is associ- 14 Set jmin = the index of the VM at which the minimum found
ated with one particle. To ensure convergence of algorithm, the 15 Allocate task i to VM at index jmin in the allocation matrix
velocity matrix elements are bounded by the interval [v min , v max ]. 16 Add Mn to all values of column jmin in the execution time
In the conventional PSO, the particles are updated according to matrix
Eq. (9). However, in the problem of task scheduling where particles 17 End
are defined with binary values, the particles update equation is
derived based on the binary PSO as follows; for each velocity of
particle l, get the index jmax of the machine at which the velocity
In algorithm 1, the main loop of the MCT starts at step 11 and
is maximum compared to all values in the i-th row. Set the corre-
ends at step 17. This loop will run for Nt iterations equal to the
sponding element of particle l at location (jmax ; i) to ‘10 and all other
number of tasks. At each iteration i, the minimum execution time
locations at ðj; iÞgj¼jmax to ‘00 . As shown in Fig. 3, for example, for task
of task i at VM j is found. Then in step 15, task i, is assigned to VM j,
1, the maximum velocity appears at location (i ¼ 1, jmax ¼ 2) with a then, in step 16, the minimum execution time found is added to all
value equal to v 1;2 , thus task 1 will be allocated to VM 2 by setting tasks with index (i + 1) to Nt and the process repeats until the loop
x12 ¼ 1 and the rest values of the row are set to 0. ends.
In PSO, the particles are initialized randomly by allocating each
task randomly to one of the available machines without con- Algorithm 2: Pseudo-code of LJFP algorithm
straints. Random initialization is one of the key features of PSO
in spreading the particles over the search space. However, if we 1 Input:
could select a good starting search point the performance of the 2 No. of tasks (Nt),
algorithm can be improved. Therefore, the proposed method is to 3 Task lengths (T),
utilize heuristic scheduling algorithms to initialize the PSO search 4 No. of VMs (Nv),
process. The two standard algorithms are LJFP and MCT. The pro- 5 VMs execution rates (V)
posed combined algorithms termed LJFP-PSO and MCT-PSO are 6 Output:
7 Allocation matrix (task to VM) of size Nt Nv
8 Start:
9 Calculate the estimated execution time matrix of size
Nt Nv
10 Initialize the allocation matrix (task to VM) of size Nt Nv
to zeros
11 Sort the task according to their lengths in descending order
12 Sort the VMs according to their execution rate in descending
order
13 For i = 1 to no. of tasks Nt
14 If mod(i,Nv) = 0
15 VM index j = Nv
16 else
17 VM index j = mod(i, Nv)
18 end
19 Allocate task i to VM at index j in the allocation matrix
20 End
loop of the LJFP run for Nt iterations equal to the number of tasks (MS, TET, DI, and TEC). The proposed algorithms will consider only
from step 13–20. In each iteration, the sorted task is assigned to one of these four metrics as an objective to find the optimum task
its corresponding sorted VM. The assignment process repeats in scheduling. The local and global solutions along with their best fit-
cyclic form such that when the tasks exceeds the number of VMs ness values are found in step 17. In steps 18 and 19, the velocity and
the assignment starts over to the first sorted VM. position matrixes are updated according to Eqs. (10) and (9) respec-
tively. At the end of the last main loop iteration, the optimal
Algorithm 3: Pseudo-code of LJFP-PSO and MCT-PSO scheduling solution will be found in the global best particle alloca-
algorithms tion matrix.
1 Input:
2 No. of tasks (Nt), 5. Simulation results and discussion
3 Task lengths (T),
4 No. of VMs (Nv), In this section, the performance of the proposed algorithm is
5 VMs execution rates (V) evaluated and compared with standard and swarm-based algo-
6 Output: rithms. The scheduling algorithms are simulated on MATLAB soft-
7 Allocation matrix (task to VM) of size Nt Nv ware installed on a PC with IntelÒ CoreTM i7-55000U CPU with
8 Start: 2.40 GHz, and RAM of 8 GB running on 64-bit Windows 10 Pro
9 Calculate the estimated execution time matrix of size operating system platform. The parameters of the virtual cloud
Nt Nv computing environment and the PSO algorithm are defined in
10 Initialize the allocation matrix (task to VM) of size Nt Nv Tables 2 and 3 respectively.
to zeros The scheduling algorithms selected for comparison are; MCT,
11 Initialize PSO parameters SJFP, LJFP, MinMin, MaxMin, PSO, and SJFP-PSO. Fig. 4, shows the
12 Initialize particles according to LJFP algorithm for LJFP-PSO, simulation method structure. In addition, four recent metaheuris-
or tic algorithms (MALO, MSDE, WOA, and DSOS) are compared the
13 Initialize particles according to MCT algorithm for MCT-PSO. proposed algorithms.
14 For it = 1 to no. of iterations The performance is compared based on four main metrics; MS,
15 For i = 1 to no. of particles TET, DI, and TEC. At first, for specific numbers of tasks and virtual
16 Calculate fitness value for each particle (MS, TET, DI, or machines, a Matlab program generates random task lengths and
TEC) random VMs execution rates within the ranges (1000–4000) in
17 Find best particle so far. MI and (1000–5000) in MIPS respectively. In addition, the power
18 Calculate the new velocity value according to equation consumption of each VM is set randomly between (200–1000)
(v) watt. The random power consumption values are related to VMs
19 Update the particles according to the maximum velocity execution rates, i.e., the more execution rate a VM has the more
value by power it consumes. In the simulation, a VM can have one of two
20 setting the corresponding VM location to 1 and 0 for the rest. power consumption states, active and idle states. In active state,
21 End a VM consumes 100% of its total power consumption denoted by
22 Find the global best particle a as defined by equation (8). On the other hand, in an idle state
23 End
24 Set the allocation matrix to the global best particle
Table 2
Cloud system parameters.
In algorithm 3, where the PSO is initialized heuristically, step 12 call Parameters Values
algorithm 1 or 2 to initialize the PSO using MCT or LJFP algorithms
Number of tasks 200–1000
respectively. The PSO iteration loop starts at step 14 and ends at 23. Size of tasks (MI) 1000–4000
The algorithm processes each particle in the submain loop from Number of VMs 40–200
steps 15 to 21. Within the submain loop, the fitness value for each Execution rate of VMs (MIPS) 1000–5000
particle (each solution represented by task to VM allocation matrix) Power consumption of VMs (Watt) 200–1000
Percentage of power consumption of idle to active states 0.6–0.7
is calculated. The fitness value depends on one of the four metrics
6
S.A. Alsaidy, A.D. Abbood and M.A. Sahib Journal of King Saud University – Computer and Information Sciences xxx (xxxx) xxx
Fig. 5. Performance comparison between scheduling algorithms in scenario 1 using makespan (MS).
7
S.A. Alsaidy, A.D. Abbood and M.A. Sahib Journal of King Saud University – Computer and Information Sciences xxx (xxxx) xxx
Fig. 6. Performance comparison between algorithms in scenario 1 using total execution time (TET).
Fig. 7. Performance comparison between scheduling algorithms in scenario 1 using degree of imbalance (DI).
Fig. 8. Performance comparison between scheduling algorithms in scenario 1 using total energy consumption (TEC).
8
S.A. Alsaidy, A.D. Abbood and M.A. Sahib Journal of King Saud University – Computer and Information Sciences xxx (xxxx) xxx
Table 4
Makespan performance comparison (scenario 1).
Nt Algorithm
MCT SJFP LJFP MinMin MaxMin PSO SJFP-PSO LJFP-PSO MCT-PSO
200 2.2 6.5 3.5 2.7 2.4 4.0 4.9 3.3 2.2
400 4.2 11.3 8.3 6.6 5.9 7.5 7.4 7.0 4.2
600 9.4 14.6 12.0 10.7 9.9 10.4 11.1 11.9 8.5
800 12.1 20.4 17.6 16.0 14.9 15.2 14.2 16.4 12.1
1000 18.0 25.7 22.8 21.2 20.0 20.7 21.8 22.8 17.6
Table 5
Total execution time performance comparison (scenario 1).
Nt Algorithm
MCT SJFP LJFP MinMin MaxMin PSO SJFP-PSO LJFP-PSO MCT-PSO
200 159.6 215.5 184.6 155.3 172.4 155.9 158.7 176.8 155.8
400 326.7 390.0 361.7 349.8 346.6 328.4 332.2 332.6 326.7
600 535.8 610.0 582.0 571.0 562.4 520.9 527.8 517.0 535.8
800 716.5 805.0 776.2 771.4 753.2 690.4 709.6 722.0 716.5
1000 972.3 1093.5 1058.2 1051.2 1025.6 931.1 966.4 959.9 972.3
Table 6
Degree of imbalance performance comparison (scenario 1).
Nt Algorithm
MCT SJFP LJFP MinMin MaxMin PSO SJFP-PSO LJFP-PSO MCT-PSO
200 0.9 2.7 1.2 1.7 0.6 2.4 2.7 1.2 0.9
400 0.6 2.5 1.7 1.3 1.0 2.0 2.0 1.4 0.6
600 1.1 1.9 1.5 1.3 1.1 1.8 1.7 1.5 1.0
800 1.1 2.1 1.7 1.5 1.4 1.7 1.9 1.5 1.1
1000 1.2 1.9 1.7 1.5 1.4 1.8 1.9 1.7 1.2
Table 7
Total energy consumption performance comparison (scenario 1).
Nt Algorithm
MCT SJFP LJFP MinMin MaxMin PSO SJFP-PSO LJFP-PSO MCT-PSO
200 123 287 173 141 130 189 222 173 122
400 242 534 412 341 314 368 368 372 242
600 476 683 581 527 496 514 523 580 440
800 600 919 810 749 709 685 677 674 599
1000 836 1120 1014 952 910 935 929 1005 822
The convergence curves of the PSO, SJFP-PSO, LJFP-PSO, and it starts at MS = 100. As can be seen from Fig. 13, the MCT-PSO out-
MCT-PSO are depicted in Fig. 13. In this experiment, the number performs all other algorithms.
of tasks and virtual machines are fixed to be 1000 and 20 respec- The performance of the proposed MCT-PSO and LJFP-PSO are
tively. The makespan as an objective is plotted with respect to compared with four well-known and recent metaheuristic algo-
the number of iterations. Only the first 20 iterations are shown rithms. These comparative algorithms are; multi-objective
in the plot as the convergence behavior of all algorithms is at antlion optimizer (MALO) (Abualigah and Diabat, 2020), moth
steady after iteration 20. search based on differential evolution (MSDE) (Elaziz et al.,
As shown in Fig. 13, the PSO, SJFP-PSO, LJFP-PSO, and MCT-PSO 2019), whale optimization algorithm (WOA) (Mirjalili and
converge to MS equal to 87, 81, 80, and 67 sec. respectively. The Lewis, 2016), and discrete symbiotic organism search algorithm
heuristic initialized PSO algorithms start with a MS values equal (DSOS) (Abdullahi et al., 2016). In order to obtain fair compar-
to their corresponding heuristic initialization. In this case, the ison, the same simulation parameters used with the chosen
SJFP-PSO, LJFP-PSO, and MCT-PSO start with 124, 121, and 114 algorithms are applied. To present the scalability of the proposed
respectively. Since the PSO initialize its particles randomly, it can algorithms in the comparison, the number of tasks is varied
start with any MS greater than or equal to its optimal value, here from 100 to 1000 tasks with a step of 100. The tasks length
9
S.A. Alsaidy, A.D. Abbood and M.A. Sahib Journal of King Saud University – Computer and Information Sciences xxx (xxxx) xxx
Fig. 10. Performance comparison between scheduling algorithms in scenario 2 using total execution time (TET).
Fig. 11. Performance comparison between scheduling algorithms in scenario 2 using degree of imbalance (DI).
10
S.A. Alsaidy, A.D. Abbood and M.A. Sahib Journal of King Saud University – Computer and Information Sciences xxx (xxxx) xxx
Fig. 12. Performance comparison between scheduling algorithms in scenario 2 using total energy consumption (TEC-Joule).
Table 8
Makespan performance comparison (scenario 2).
Nt Algorithm
MCT SJFP LJFP MinMin MaxMin PSO SJFP-PSO LJFP-PSO MCT-PSO
40 25.9 29.0 27.6 27.5 25.1 20.4 20.6 22.0 25.9
80 10.6 16.4 14.3 14.8 11.8 14.9 16.4 13.9 9.7
120 5.5 10.8 8.5 7.8 6.1 9.3 10.5 8.5 4.9
160 3.7 8.5 6.2 4.9 4.1 7.1 7.8 6.2 3.7
200 2.7 5.6 4.5 3.7 3.2 7.2 5.6 4.5 2.7
Table 9
Total execution time performance comparison (scenario 2).
Nt Algorithm
MCT SJFP LJFP MinMin MaxMin PSO SJFP-PSO LJFP-PSO MCT-PSO
40 547.0 571.0 567.8 568.3 558.4 475.6 494.5 504.6 486.8
80 522.8 600.2 583.2 579.4 556.5 526.6 519.1 510.7 518.7
120 433.3 510.4 484.2 469.4 459.8 436.5 445.7 454.3 427.0
160 417.6 504.2 466.3 440.8 442.1 415.8 430.0 445.8 417.6
200 410.1 476.8 461.2 430.8 442.7 443.8 465.5 453.6 410.1
Table 10
Degree of imbalance performance comparison (scenario 2).
Nt Algorithm
MCT SJFP LJFP MinMin MaxMin PSO SJFP-PSO LJFP-PSO MCT-PSO
40 1.4 1.6 1.5 1.5 1.3 1.2 1.3 1.1 1.0
80 0.9 1.7 1.4 1.6 1.1 1.8 1.7 1.4 0.8
120 0.8 2.1 1.5 1.4 0.9 2.1 2.1 1.5 0.7
160 0.7 2.3 1.5 1.2 0.7 2.5 2.2 1.5 0.7
200 0.7 1.8 1.2 1.7 0.6 2.7 1.8 1.2 0.7
defined by million instructions (MI) are generated uniformly comparison results between the proposed and the comparative
between 1000 and 2000. Moreover, the number of virtual machi- algorithms. The comparison is based on the performance of the
nes is kept constant at Nv = 20. The processing power of the algorithms in balancing the load among the virtual machines
VMs ranges from 1000 to 10,000 MIPS. Fig. 14 shows the defined by DI.
11
S.A. Alsaidy, A.D. Abbood and M.A. Sahib Journal of King Saud University – Computer and Information Sciences xxx (xxxx) xxx
Table 11
Total energy consumption performance comparison (scenario 2).
Nt Algorithm
MCT SJFP LJFP MinMin MaxMin PSO SJFP-PSO LJFP-PSO MCT-PSO
40 461 504 484 482 449 393 386 400 362
80 392 558 498 511 426 495 542 478 365
120 340 585 478 444 367 460 571 478 311
160 318 628 477 395 344 522 615 477 318
200 304 531 439 375 336 598 497 418 304
From Fig. 14, it can be observed that the MCT-PSO and LJFP- only at high number of tasks (Nt = 1000), the comparative algo-
PSO attain optimum load balance by scoring minimum DI for rithms overcome the proposed algorithms and record less DI val-
Nt = 100 to 500. At Nt = 600 to 900 the proposed algorithms have ues except for DSOS algorithm. Table 12, summarize the
almost comparable DI values to their comparatives. However, numerical comparison results.
12
S.A. Alsaidy, A.D. Abbood and M.A. Sahib Journal of King Saud University – Computer and Information Sciences xxx (xxxx) xxx
Table 12
Summary of numerical comparison results.
Nt Algorithms
MCT-PSO LJFP-PSO MALO MSDE WOA DSOS
100 0.5571 0.8227 0.9301 0.8649 1.0857 1.7500
200 0.5261 0.7047 0.9561 0.9046 1.1122 1.3511
300 0.5641 0.7906 0.9798 0.9321 1.0624 1.3526
400 0.7422 0.7969 0.9020 0.8469 0.9105 1.6200
500 0.7463 0.8929 0.9529 1.0228 0.9576 1.7522
600 0.9477 1.0424 0.9263 1.0017 1.0312 1.7141
700 0.9560 0.9903 0.9667 0.9548 1.1451 1.7545
800 1.0534 1.2015 0.9216 0.9325 0.9658 1.8025
900 1.1315 1.0407 0.9726 0.9895 0.9965 1.8055
1000 1.3754 1.5681 0.9701 1.0010 1.0865 1.8654
6. Conclusion Belgacem, A., Beghdad-Bey, K., Nacer, H., Bouznad, S., 2020. Efficient dynamic
resource allocation method for cloud computing environment. Cluster Comput
23 (4), 2871–2889. https://doi.org/10.1007/s10586-020-03053-x.
In cloud computing, adopting an efficient task scheduling algo- Chaudhary, D., Kumar, B., 2018. Cloudy GSA for load scheduling in cloud computing.
rithm is vital for both users and providers. In this paper, heuristic Appl. Soft Comput. 71, 861–871. https://doi.org/10.1016/j.asoc.2018.07.046.
Devaraj, A.F.S., Elhoseny, M., Dhanasekaran, S., Lydia, E.L., Shankar, K., 2020.
algorithms are used to assist PSO algorithm for task scheduling. Hybridization of firefly and improved multi-objective particle swarm
The particles of the PSO are heuristically initialized using longest optimization algorithm for energy efficient load balancing in cloud computing
job to fastest processor (LJFP) and minimum completion time environments. J. Parallel Distrib. Comput. 142, 36–45. https://doi.org/10.1016/j.
jpdc.2020.03.022.
(MCT) algorithms. Starting the search process with LJFP and MCT
Elaziz, M.A., Xiong, S., Jayasena, K.P.N., Li, L., 2019. Task scheduling in cloud
selected based solutions can significantly influence the conver- computing based on hybrid moth search algorithm and differential evolution.
gence speed and performance. Simulation results demonstrate Knowl.-Based Syst. 169, 39–52. https://doi.org/10.1016/j.knosys.2019.01.023.
the effective performance of the proposed LJFP-PSO and MCT-PSO Fanian, F., Bardsiri, V.K., Shokouhifar, M., 2018. A new task scheduling algorithm
using firefly and simulated annealing algorithms in cloud computing. Int. J. Adv.
algorithms. Also, comparison simulation results revealed that the Comput. Sci. Appl. 9, 195–202. https://doi.org/10.14569/IJACSA.2018.090228.
proposed heuristic initialized PSO outperforms recent task Golchi, M.M., Saraeian, S., Heydari, M., 2019. A hybrid of firefly and improved
scheduling algorithms in terms of convergence and load balancing. particle swarm optimization algorithms for load balancing in cloud
environments: performance evaluation. Comput. Netw. 162, 106860. https://
It is worth noting that the proposed heuristic initialization of the doi.org/10.1016/j.comnet.2019.106860.
PSO population creates initial particles that all start the search pro- Jena, U.K., Das, P.K., Kabat, M.R., 2020. Hybridization of meta-heuristic algorithm for
cess from a single starting position. This starting point is created load balancing in cloud computing environment. J. King Saud Univ. – Comput.
Inf. Sci. 1–11. https://doi.org/10.1016/j.jksuci.2020.01.012.
using a heuristic algorithm. As a future work, it is possible to inves- Kaur, A., Kaur, B., 2019. Load balancing optimization based on hybrid Heuristic-
tigate the use of different heuristic methods to create different Metaheuristic techniques in cloud environment. J. King Saud Univ. – Comp. Inf.
starting search points for the PSO or any other metaheuristic Sci. https://doi.org/10.1016/j.jksuci.2019.02.010.
Keshanchi, B., Souri, A., Navimipour, N.J., 2017. An improved genetic algorithm for
algorithms.
task scheduling in the cloud environments using the priority queues: formal
verification, simulation, and statistical testing. J. Syst. Softw. 124, 1–21. https://
Declaration of Competing Interest doi.org/10.1016/j.jss.2016.07.006.
Kumar, M., Sharma, S.C., 2018. PSO-COGENT: Cost and energy efficient scheduling in
cloud environment with deadline constraint. Sustainable Comput. Inf. Syst. 19,
The authors declare that they have no known competing finan- 147–164. https://doi.org/10.1016/j.suscom.2018.06.002.
cial interests or personal relationships that could have appeared Mahato, D.P., Singh, R.S., Tripathi, A.K., Maurya, A.K., 2017. On scheduling
to influence the work reported in this paper. transactions in a grid processing system considering load through Ant Colony
Optimization. Appl. Soft Comput. 61, 875–891. https://doi.org/10.1016/j.
asoc.2017.08.047.
References Mansouri, N., Mohammad Hasani Zade, B., Javidi, M.M., 2019. Hybrid task
scheduling strategy for cloud computing by modified particle swarm
Abdullahi, M., Ngadi, M.A., Abdulhamid, S.M., 2016. Symbiotic organism search optimization and fuzzy theory. Comput. Ind. Eng. 130, 597–633. https://doi.
optimization based task scheduling in cloud computing environment. Future org/10.1016/j.cie.2019.03.006.
Generat. Comp. Syst. 56, 640–650. https://doi.org/10.1016/j.future.2015.08.006. Mirjalili, S., Lewis, A., 2016. The whale optimization algorithm. Adv. Eng. Softw. 95,
Abualigah, L., Diabat, A., 2020. A novel hybrid antlion optimization algorithm for 51–67. https://doi.org/10.1016/j.advengsoft.2016.01.008.
multi-objective task scheduling problems in cloud computing environments. Mishra, S.K., Sahoo, B., Parida, P.P., 2020. Load balancing in cloud computing: A big
Cluster Comput. 5. https://doi.org/10.1007/s10586-020-03075-5. picture. J. King Saud Univ. – Comp. Inf. Sci. 32 (2), 149–158. https://doi.org/
Adhikari, M., Nandy, S., Amgoth, T., 2019. Meta heuristic-based task deployment 10.1016/j.jksuci.2018.01.003.
mechanism for load balancing in IaaS cloud. J. Network Comp. Appl. 128, 64–77. Sardaraz, M., Tahir, M., 2019. A hybrid algorithm for scheduling scientific workflows
https://doi.org/10.1016/j.jnca.2018.12.010. in cloud computing. IEEE Access 7, 186137–186146. https://doi.org/10.1109/
Al-Arasi, R., Saif, A., 2020. Task scheduling in cloud computing based on ACCESS.2019.2961106.
metaheuristic techniques: A review paper. EAI Endorsed Trans. Cloud Syst. 6, Strumberger, I., Bacanin, N., Tuba, M., Tuba, E., 2019. Resource scheduling in cloud
162829. https://doi.org/10.4108/eai.13-7-2018.162829. computing based on a hybridized whale optimization algorithm. Appl. Sci. 9.
Arunarani, A.R., Manjula, D., Sugumaran, V., 2019. Task scheduling techniques in https://doi.org/10.3390/app9224893.
cloud computing: a literature survey. Future Generat. Comp. Syst. 91, 407–415.
https://doi.org/10.1016/j.future.2018.09.014.
13