Accepted Manuscript

An efficient symbiotic organisms search algorithm with chaotic optimization strategy

for multi-objective task scheduling problems in cloud computing environment

Mohammed Abdullahi, Md Asri Ngadi, Salihu Idi Dishing, Shafi'i Muhammad

Abdulhamid, Barroon Isma'eel Ahmad

PII: S1084-8045(19)30049-9
DOI: https://doi.org/10.1016/j.jnca.2019.02.005
Reference: YJNCA 2309

To appear in: Journal of Network and Computer Applications

Received Date: 9 July 2018

Revised Date: 4 January 2019
Accepted Date: 5 February 2019

Please cite this article as: Abdullahi, M., Ngadi, M.A., Dishing, S.I., Abdulhamid, Shafi'.Muhammad.,
Ahmad, Barroon.Isma'., An efficient symbiotic organisms search algorithm with chaotic optimization
strategy for multi-objective task scheduling problems in cloud computing environment, Journal of
Network and Computer Applications (2019), doi: https://doi.org/10.1016/j.jnca.2019.02.005.

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to
our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and all
legal disclaimers that apply to the journal pertain.
 ACCEPTED MANUSCRIPT

An efficient symbiotic organisms search algorithm with

chaotic optimization strategy for multi-objective task
scheduling problems in cloud computing environment

PT
Mohammed Abdullahia,∗, Md Asri Ngadic , Salihu Idi Dishinga,c , Shafi’i
Muhammad Abdulhamidb , Barroon Isma’eel Ahmada

RI
a Department of Computer Science, Ahmadu Bello University Zaria, Nigeria
b Department of Cyber Security Science, Federal University of Technology Minna, Nigeria
c Department of Computer Science, Faculty of Computing, Universiti Teknologi Malaysia,

SC
81310 Johor Bahru, Malaysia

U
Abstract

AN
In Cloud Computing model, users are charged according to the usage of
resources and desired Quality of Service (QoS). Multi-objective task schedul-
ing problem based on desired QoS is an NP-Complete problem. Due to the
M
NP-Complete nature of task scheduling problems and huge search space pre-
sented by large scale problem instances, many of the existing solution algo-
D

rithms cannot effectively obtain global optimum solutions. In this paper, a

chaotic symbiotic organisms search (CMSOS) algorithm is proposed to solve
TE

multi-objective large scale task scheduling optimization problem on IaaS cloud

computing environment. Chaotic optimization strategy is employed to generate
initial ecosystem(population), and random sequence based components of the
EP

phases of SOS are replaced with chaotic sequence to ensure diversity among
organisms for global convergence. In addition, chaotic local search strategy is
applied to Pareto Fronts generated by SOS algorithms to avoid entrapment in
C

local optima. The performance of the proposed CMSOS algorithm is evaluated

AC

on CloudSim simulator toolkit, using both standard workload traces and syn-

∗ Corresponding author
Email addresses: abdullahilwafu@abu.edu.ng (Mohammed Abdullahi),
dr.asri@utm.my (Md Asri Ngadi), sidishing@abu.edu.ng (Salihu Idi Dishing),
shafii.abdulhamid@futminna.edu.ng (Shafi’i Muhammad Abdulhamid),
sidishing@abu.edu.ng (Barroon Isma’eel Ahmad)

Preprint submitted to Journal of LATEX Templates February 5, 2019

 ACCEPTED MANUSCRIPT

thesized workloads for larger problem instances of up to 5000. Moreover, the

performance of the proposed CMSOS algorithm was found to be competitive

PT
with the existing with the existing multi-objective task scheduling optimiza-
tion algorithms. The CMSOS algorithm obtained significant improved optimal
trade-offs between execution time (makespan) and financial cost (cost) with no

RI
computational overhead. Therefore, the proposed algorithms have potentials to
improve the performance of QoS delivery.

SC
Keywords:
Symbiotic Organisms Search, Metaheuristics Algorithms, Optimization,
NP-Complete, Multi-Objective Task Scheduling, Cloud Computing

U
1. Introduction
AN
To meet up with the increasing computational demand of large scale ap-
plications, Cloud Computing is witnessing high rate deployment of large scale
M
applications in recent times, because Cloud provides elastic and flexible com-
5 pute resources which can be leased on pay-per-use model (Foster et al., 2008).
D

Large scale applications consist of huge number of tasks which are executed
on Infrastructure-as-a-Service clouds. Cloud Computing services are offered in
TE

form of Software as a Service (SaaS), Platform as a Service (PaaS), and Infras-

tructure as a Service (IaaS). SaaS service model delivers applications to end
10 users via Internet and these applications are accessed using client applications
EP

like web browsers. SaaS is usually used for service applications like web-mail,
and document editing applications. PaaS provides application developers with
C

environment for development, testing and hosting of their applications.

Moreover, IaaS provides access to flexible and scalable computing resources
AC

15 for large scale application deployment. With IaaS model, virtualized com-
pute resources called virtual machines (VMs) with pre-configured CPU, storage,
memory, and bandwidth are leased to users by paying for what they use only.
Various VM instances are available to the users at different prices to serve their
various application needs, this gives users the freedom to control compute re-

2
 ACCEPTED MANUSCRIPT

20 source at their disposal. IaaS provides three inherent benefits to users. First,
users lease resource on demand, and charged based on pay-per-usage similar to

PT
basic utilities like electricity, gas, and water. This enables users to shrink or ex-
pand their resource subscription base on the needs of their application. Second,
IaaS Cloud provides direct resource provisioning which improve the performance

RI
25 of user applications. Third, users can demand for leased resources any time and
any where according to the desired level of service. However, determining the

SC
adequate number of resources to execute a set of large scale task on IaaS Cloud
is still an open problem (Thakur and Goraya, 2017; Wu et al., 2015; Zeng et al.,
2015).

U
30 Due to the practical applications and challenges of executing large scale ap-
plications, task scheduling of applications on the large scale have become an
AN
emerging research in cloud computing and have attracted significant attention
of researchers in recent times (Ferdaus et al., 2017). Various heuristics have
M
been applied to solve task scheduling problems which generate optimal solu-
35 tions for small size problems (Chen et al., 2013; Ming and Li, 2012; Mao et al.,
2014; Patel et al., 2015). However, the quality of solutions produced by these
D

techniques degrades woefully as the problem size and number of variables to be

optimized increases. Also, these heuristic methods do not have provisions and
TE

support for meeting various QoS requirements (Hayyolalam and Kazem, 2018;
40 Vakili and Navimipour, 2017; Ghazouani and Slimani, 2017). In contrast, many
EP

cloud users requires certain QoS satisfaction especially for scientific and busi-
ness domain applications. In recent times, attempts have been made to address
task scheduling problems using metaheuristic algorithms like genetic algorithms
C

(GA), particle swarm optimization (PSO), and ant colony optimization (ACO)
to address this problem (Hameed et al., 2014; Wu et al., 2015; Singh and Chana,
AC

45

2016). Utilizing metaheuristic algorithms for solving task scheduling problems

in Cloud have shown promising improvements in achieving efficiency, by reduc-
ing the solution search space. However, metaheuristic algorithms incur high
computational time and in some cases return local optimum solution especially
50 when dealing with large solution space, also, these techniques may suffer from

3
 ACCEPTED MANUSCRIPT

premature convergence and imbalance between local and global search (Tsai and
Rodrigues, 2014; Guzek et al., 2015; Kalra and Singh, 2015; Zhan et al., 2015;

PT
Xue et al., 2016; Meena et al., 2016). These limitations result to sub-optimal
task schedule solutions which affects the performance of service provision in
55 terms of meeting the desired QoS objectives. Furthermore, most of the existing

RI
works fail to capture the essential features of cloud computing like heterogeneity,
elasticity, and dynamism of computing resources there by fail to fulfill user QoS

SC
needs. Hence, there is need for metaheuristic based optimization algorithms
that can efficiently cope with large search space when scheduling large scale
60 applications. Hence, there is scope for further development of task scheduling

U
solutions for further improved solutions. Therefore, this paper presents Chaotic
Multi-Objective Symbiotic Organisms Search (CMSOS) based task scheduling
AN
algorithms for large scale task scheduling optimization on IaaS cloud.
Symbiotic Organisms Search (SOS) algorithm is a recently introduced meta-
M
65 heuristic algorithm in Cheng and Prayogo (2014) and has gathered considerable
interest of researchers from natural computing. SOS was originally proposed
to handle continuous benchmark and engineering problems, which was shown
D

to have a robust performance and has faster convergence speed when com-
pared with GA (Deb et al., 2002), PSO (Kennedy, 2011), Differential Evolution
TE

70 (DE) (Qin et al., 2009), Bees Algorithm (BA) (Pham et al., 2011), and Particle
Bee Algorithm (PBA) (Cheng and Lien, 2012) which are the traditional meta-
EP

heuristic algorithms. SOS have proven to be efficient for optimizing complex

multidimensional search space while handling multi-objective and constrained
optimization problems. Active researches on SOS since its introduction includes
C

75 hybridization, discrete optimization problems, constrained and multi-objective

optimization. Hybridization intends to combine the strengths of SOS like global
AC

search ability and rapid optimization, with other related techniques to address
some of the issues with SOS performance, like entrapment in local optima.
SOS metaheuristic optimization algorithm is based on the interaction be-
80 tween paired of organisms for survival in an ecosystem, it shares some common
features with most of the nature inspired algorithms. The candidate solutions

4
 ACCEPTED MANUSCRIPT

are represented by population of organisms, and mutualism, commensalism and

parasitism operators to direct the search process by candidate solutions. SOS

PT
requires the settings of population size and stopping criterion before the search
85 process starts, in the course of search process selection mechanism is used to
keep better solutions. SOS does not require algorithm specific parameters un-

RI
like PSO that needs inertia weight, social and cognitive factors or GA that
used crossover and mutation. Moreover, inadequate turning of these algorithm

SC
specific parameters could lead to non-optimal solutions. SOS optimization algo-
90 rithm have been recently found to be successful in solving various optimization
problems in a variety of domains like economic dispatch (Dosoglu et al., 2016;

U
Secui, 2016; Guvenc et al., 2016; Sonmez et al., 2016; Tiwari and Pandit, 2016),
power optimization (Banerjee and Chattopadhyay, 2017; Duman, 2016; Zamani

95
AN
et al., 2017; Banerjee and Chattopadhyay, 2016), construction project schedul-
ing (Tran et al., 2016; Cheng et al., 2015), task scheduling (Abdullahi et al.,
M
2016; Abdullahi and Ngadi, 2016; Abdullahi et al., 2017), design optimization
of engineering structures (Tejani et al., 2016; Panda and Pani, 2016; Prayogo
et al., 2017; Nama et al., 2016), transportation (Eki et al., 2015; Vincent et al.,
D

2017), energy optimization (Kanimozhi et al., 2016), wireless communication

(Dib, 2016), and machine learning (Nanda and Jonwal, 2017; Wu et al., 2016).
TE

100

The standard SOS algorithm was proposed to solve unconstrained continuous

optimization problems while task scheduling problem is a discrete optimization
EP

problem.
Our earlier works (Abdullahi et al., 2016; Abdullahi and Ngadi, 2016) consid-
105 ers only single objective task scheduling optimization problems while this paper
C

considers multi-objective task scheduling optimization problem in addition to

the following contributions:
AC

• Pareto based multi-objective SOS algorithm

• Chaotic based ecosystem (population) initialization to improve population

110 diversity and global convergence.

• Replacement of random sequence components of the original SOS with

5
 ACCEPTED MANUSCRIPT

chaotic sequence to ensure global convergence.

• Chaotic Local Search to avoid entrapment of Pareto Front in local optima.

PT
• Performance evaluation of the proposed algorithm against recent multi-
115 objective algorithms.

RI
The structure of the remaining parts of the paper are follows: Review of
related work on existing multi-objective task scheduling techniques are discussed

SC
in Section 2. Section 3 presents the definition of task scheduling problem along
with multi-objective task scheduling formulation. The original SOS algorithm is
120 presented in Section 4. Section 5 describes the concept of chaotic optimization

U
strategy along with chaotic local search technique. The detailed description of

AN
the proposed algorithm is presented in Section 6, performance evaluation and
analysis of the obtained results are presented in Section 7. Finally, conclusion
and suggestions for possible future research are presented in Section 8.
M
125 2. Related work
D

Task scheduling optimization approaches either focused on single objective or

multi-objective. The single objective task scheduling optimization approaches,
TE

only try to optimize either makespan or cost (Hu et al., 2018; Abdullahi et al.,
2016; Abdullahi and Ngadi, 2016; Latiff et al., 2016; Li et al., 2016; Nirmala and
EP

130 Bhanu, 2016; Zhong et al., 2016; Meena et al., 2016; Liu et al., 2016; Tawfeek
et al., 2015; Li et al., 2015; Zuo et al., 2014; Rodriguez and Buyya, 2014; Netjinda
et al., 2014). However, because of the rapid development of Cloud, several QoS
C

objectives needs to be considered which makes task scheduling a multi-objective

optimization problem. The complexity of the multi-objective task optimization
AC

135 formulation arise from the fact that users and providers have different optimiza-
tion goals. Users are mainly concerned with minimizing makespan and cost,
whereas providers want to maximize resource utilization and energy consump-
tion while meeting user QoS requirements. In this situation, task scheduling
have to be solved as a multi-objective optimization problem trying to optimize

6
 ACCEPTED MANUSCRIPT

140 many and yet conflicting objectives, where it is not possible to obtain optimal
solution with regards to all objectives. Therefore, a good trade-offs between the

PT
objectives need to be obtained.
Multi-objective task scheduling optimization challenge is an important con-
sideration because of its direct effect on both cloud service providers and con-

RI
145 sumers (Zhan et al., 2015). In cloud computing platform, task scheduling algo-
rithms must optimize financial cost of leasing compute resources in addition to

SC
execution time (makespan) and other QoS metrics. Generally, cloud providers
offer heterogeneous set of resources (VM instances) at various prices with var-
ied performance. In this way, task scheduling problem needs to be formulated

U
150 as a multi-objective optimization problem that intend to optimize conflicting
objectives such as maksepan and financial cost of task execution. With multi-
AN
objective formulation, there is no single solution which is optimal with respect to
all objectives, but a set of trade-off solutions called Pareto front (Tao et al., 2014;
M
Elhabyan et al., 2018). Multi-objective task scheduling optimization problems
155 are usually solved using aggregation, hierarchical, Pareto, and coevolutionary
multi-swarm approaches. The aggregation (weighted) approach is the common
D

method for solving multi-objective task scheduling problems. The approach as-
sign weights to multiple objectives and sum up the objectives to form single
TE

objective function. For instance, Delavar and Aryan (2014) proposed GA based
160 task scheduling algorithm to optimize makespan, reliability, and load balanc-
EP

ing of applications by putting into consideration the heterogeneous character-

istics of compute resources. Also, Shen et al. (2016) developed GA algorithm
for adaptive scheduling of tasks considering energy consumption and makespan
C

performance. Casas et al. (2016) proposed GA based task scheduling technique

for optimizing makespan and cost. Zuo et al. (2015) proposed ACO based task
AC

165

scheduling algorithm to optimize budget and deadline constrained task schedul-

ing problems, the proposed approach simultaneously makespan and cost within
a given budget and deadline. However, the results of different objectives is
dependent on the values of the assigned weights which may not adequately rep-
170 resent the decision of the user. Moreover, the approach produce only solution

7
 ACCEPTED MANUSCRIPT

which is not adequate for multi-objective decision problems.

The hierarchical approaches optimize task scheduling objectives in a sequen-

PT
tial order, the optimization ordering of the objectives are determined based on
their importance and solution to the objectives are alternately sought based on
175 their ordering. For instance, the approach proposed by Teng et al. (2007) used

RI
sorting strategy, the objective functions are optimized in sequential order. The
optimization of an objective is continuously carried until no further improve-

SC
ment is possible, then next objective is optimized while meeting the constraints
of the previous optimized objectives. Similar approach was used by Zhang et al.
180 (2014) to optimize makespan and cost. However, these approaches are time

U
consuming especially when there are several objectives with constraints, since it
requires several iteration of optimization process. Moreover, the importance of
AN
the objectives is dependent on the problem, and performance of the approach
may be significantly affected by the ranking of the objectives.
M
185 To overcome the drawbacks of both aggregation and hierarchical approaches,
Pareto-based optimization approaches have been put forth for addressing multi-
objective task scheduling problems (Hu et al., 2018; Midya et al., 2018; Tao et al.,
D

2014; Durillo et al., 2014). Pareto approaches finds several optimal trade-off so-
lutions for the objectives for the optimization problem. The concept of Pareto
TE

190 dominance is applied to assign fitness to individuals. The Pareto approach does
not require transforming multiple objectives into single objective formulation,
EP

and generate several trade-off solutions in a single run. Tao et al. (2014) presents
a hybrid GA algorithms to obtain Pareto optimal solutions for makespan and
energy consumption. Pareto optimal trade-offs between makespan, cost, and
C

195 energy consumption was solved using list scheduling heuristics and hybrid PSO
respectively (Fard et al., 2014; Yassa et al., 2013). Similarly, Verma and Kaushal
AC

(2017) presents PSO based multi-objective task scheduling algorithm to ob-

tain optimal trade-offs between makespan, cost, and energy consumption while
meeting deadline and budget constraints respectively. Xu et al. (2014) put
200 forth multi-objective GA for workflow task scheduling problem to simultane-
ously minimize makespan and cost while considering the priorities of the tasks.

8
 ACCEPTED MANUSCRIPT

Moreover, Zhang et al. (2017) proposed multi-objective GA algorithm to obtain

Pareto optimal trade-offs between energy consumption, and reliability for dead-

PT
line constrained task scheduling problems. However, with Pareto task scheduling
205 approaches, it is difficult to select appropriate individual for the next generation
since Pareto dominance is a partial order (Zhan et al., 2013). Therefore, the

RI
solutions obtained may not cover the entire Pareto Front (PF) if the selection
operator fails to keep adequate diversity. Thus, developing multi-objective task

SC
scheduling that effectively assign fitness to individuals while keeping solution to
210 efficiently estimate the entire PF remains challenging research.

U
3. Multi-objective task scheduling problem

AN
Task scheduling problem considered in this paper is to minimize makespan
and financial cost (cost) for executing large scale tasks on IaaS cloud computing
environment. In this section, the IaaS cloud data center model, task execution
M
215 model and task scheduling problem formulation which form the bases of the
proposed algorithm is introduced.
D

3.1. IaaS cloud model

TE

An IaaS cloud data center provides computing resources to users through

virtual machines, an active virtual machine is called an instance. IaaS providers
usually provide various instance series types with wide range of instance types
EP

consisting of different combinations of CPU, memory and bandwidth. In this

study, the CPU capacities are used to determine the estimated execution time of
tasks. It is assumed that IaaS provider offers relatively infinite pool of instances
C

which is described by a set I = {I1 , I2 , I3 , . . .}. The instances are categorised

AC

into series based on the computing needs of the users, for instance, Amazon EC2
currently offer three instance series which are compute intensive, memory inten-
sive and storage intensive instances. The set V = {V1 , V2 , V3 , . . . , Vs , . . . , VS }
describes the type of series offered by an IaaS provider, each series type Vs con-
sists of instance types Vs = {vs1 , vs2 , vs3 , . . . , vsk , . . . , vsK }. IaaS providers describe

9
 ACCEPTED MANUSCRIPT

the CPU capacities different instance types by compute unit (CU). The com-
pute unit of an instance type vsk is denoted as pks which is defined in million

PT
floating point operations per second (MFLOPS), cost per time unit is denoted
as cks , and other features of an instance type include storage space and mem-
ory capacity. The task model considered is a collection of independent tasks

RI
t = {ti |i = 1, 2, 3, ..., n}, there is not precedence constraint between the indi-
vidual tasks. The goal of a task scheduler is to assign given tasks to instance

SC
types to optimize one or more objectives, thus, the aim of this study is to min-
imize makespan and cost under deadline constraint for task execution on IaaS
cloud infrastructure. It is assumed each instance type have sufficient memory

U
and storage to execute the collection of tasks. The execution time e(ti , vsk ) of a
task ti on an instance type vsk is determined as the ratio of task length si to its
compute unit pks as in Equation 1.
AN
si
e(ti , vsk ) = (1)
pks
M
The existing IaaS providers charge users for leased instance per-unit time and
pricing strategies differs from providers. For instance, Amazon EC2 charge
D

220 user per-hour for a leased instance and fractional hours are rounded to full
TE

hour (https://aws.amazon.com/ec2/pricing/) while Microsoft Azure charge per-

minute for an instance usage (https://azure.microsoft.com/en-us/pricing/details/virtual-
machines/linux/). Due to the different pricing strategies by various IaaS providers,
EP

the proposed algorithms are based on a generic pricing model IaaS cloud service
225 provision. Suppose a set P = {P1 , P2 , P3 , . . . , Pt , . . . , Pr } describes the price
models for IaaS cloud service provision, then a function bill(Pt , Vs , vsk ) is de-
C

fined to compute the lease cost of using an instance type vsk of instance series
AC

type Vs using the pricing model Pt . Thus, based the definition of pool of in-
stances, instance series types , instance types and pricing options, IaaS service
230 provision is represented as C = (V, Vs , P ).

10
 ACCEPTED MANUSCRIPT

3.2. Multi-objective task Scheduling formulation

Given a set of independent tasks t = {ti |i = 1, 2, 3, ..., n} and an IaaS cloud

PT
C = (V, Vs , P ), the problem is to produce one or more task schedule S with min-
imum makespan and cost such that the value of the makespan doe not exceed
imposed deadline. Task schedule S = (I, M, makespan, cost) is defined in terms

RI
of a set of leased instances, tasks to instance mapping, makespan, and cost of
execution. The set I = {I1 , I2 , I3 , . . . , In } is the set of leased instance informa-

SC
tion for each task, where Ii is a three turple: Ii =< vsk (ti ), S(ti , vsk ), F (ti , vsk ) >,
where vsk (ti ) is an instance type leased to execute task ti ∈ t with the lease
start time S(ti , vsk ) and lease finish time F (ti , vsk ). For each m(ti , vsk ) ∈ A is

U
a four tuple: m(ti , vsk ) =< ti , vsk (ti ), Sti , Fti >, where ti ∈ T is a task to be
executed on an instance type vsk ∈ I with at a starting execution time Sti and
AN
finishing execution time Fti . The values of makespan and cost are obtained
using Equations 2 and 3.
M
makespan = max{Fti : ti ∈ T } (2)

n
D

X
cost = cks × dF (ti , vsk ) − S(ti , vsk )e (3)
i=1
TE

In this study, only one instance series and one pricing option are considered for
the studied problem, the instance series type is compute intensive. Considering
multiple instance types in a single schedule could be studied in our future work.
EP

The objectives of the task scheduling problem (t, C) in Equation 4.

minimize f = (makespan, cost)T (4)

C

4. Framework of symbiotic organisms search algorithm

AC

In SOS algorithm, potential solutions are represented by a population of

organisms which evolved through successive iterations. Each organism repre-
235 sents a solution for an optimization problem. Initially, the potential solutions
are randomly generated and subsequently the solutions are refined by mutu-
alism, commensalism, and parasitism models of SOS. Mutualism is a kind of

11
 ACCEPTED MANUSCRIPT

relationship between two different species of organisms where both organisms

benefit from the interaction. A classic example of mutualism association is an

PT
240 interaction between bees and flowers. Bees collect nectar from flower for the
production of honey and nectar collection process by Bees enable the trans-
fer of pollen grains which aid pollination. Therefore, the involved organisms

RI
in the interaction mutually benefits from the relationship. In commensalism
relationship, one organism benefits from the interaction while the other is not

SC
245 harmed. A relationship between remora fish and sharks is a typical example of
commensalism association. Remora fish rides on shark for food and shark nei-
ther benefits nor harmed form the relationship. In parasitism relationship, one

U
organism initiates a relationship which benefits itself while the other organism is
harmed. An example of parasitic association is a relationship between anophe-
250
AN
les mosquito and human host. An anopheles mosquito transmits plasmodium
parasite to human host which could cause the death of human host if his/her
M
system cannot fight against the parasite.
In the SOS algorithm evolution, fitter organisms are allowed to proceed to the
next generation of potential solution while the unfitted organisms are discarded.
D

The population of organisms are created in a two d-dimensional search space,

and the positions of each organism is changed based on the models of the three
TE

phases(mutualism, commensalism, and parasitism) of the SOS. Suppose the

position of an ith organism in the solution search space is represented as in
EP

Equation 5.
Xi = (Xi1 , Xi2 , Xi3 , ..., Xid ) (5)

where Xip ∈ [Lp , Up ], p ∈ [1, d], and Lp and Up are the lower and upper bounds
C

of the pth dimension of the search space. At each iteration, the positions of
AC

255 the organisms are updated according to the three phases of the organism as
explained in the following subsections.

4.1. Mutualism phase

Suppose Xi is the ith member of the ecosystem. In this phase, a design
vector Xj is randomly selected from the swarm of organisms to interact with

12
 ACCEPTED MANUSCRIPT

another design Xi (i 6= j) for mutual benefit. The essence of the interaction is

to improve extent of survival of both Xi and Xj in the ecosystem. The new

PT
candidate solutions for Xi and Xj are obtained according to Equations 6 and
7, and the quality of these candidate solution are influenced by Mutual Vector
and Benefit Factors. M V is the mutual relationship vector between Xi and Xj

RI
as defined in Equation 8. Xbest represents the organism with best fitness value.
β1 and β2 represents the benefit factors between organism Xi and Xj . In a mu-

SC
tual relationship, an organism might benefit heavily or lightly while interacting
with a mutual partner. Therefore, β1 and β2 are stochastically obtained are
either 1 or 2. The values 1 and 2 denotes light and heavy benefits respectively.

U
The organism with best fitness value so far is represented by Xbest . By Xbest
interacting with Xi and Xj respectively, the balance between exploitation and
AN
exploration in the search procedure will be maintained to a certain extent. The
new candidate solutions replaced the old ones if their fitness values are better
than those of the old ones. In this case, Xi∗ and Xj∗ replace Xi and Xj respec-
M
tively in the next generation of ecosystem. Otherwise, Xi∗ and Xj∗ are discarded
while Xi and Xj survives to the next generation of the ecosystem. This scenario
D

is captured by Equations 9 and 10.

TE

Xi∗ = Xi + U (0, 1) ∗ (Xbest − M V ∗ β1 ) (6)

EP

Xj∗ = Xj + U (0, 1) ∗ (Xbest − M V ∗ β2 ) (7)

1
MV = (Xi + Xj ) (8)
C

2
where U (0, 1) is a vector of uniformly distributed random numbers between
AC

0 and 1; i = 1, 2, 3, ..., ecosize; j ∈ {1, 2, 3, ..., ecosize|j 6= i}; ecosize is the

number of organisms in the search space.

 X ∗ if f (X ∗ ) > f (X )
i i i
Xi = (9)
∗
 X
i if f (Xi ) ≤ f (Xi )

13
 ACCEPTED MANUSCRIPT


 X∗ if f (Xj∗ ) > f (Xj )
j
Xj = (10)
 X if f (Xj∗ ) ≤ f (Xj )

PT
j

where f (.) denotes the fitness evaluation function.

RI
4.2. Commensalism phase

In commensalism phase, an ith member of the ecosystem randomly selects an

organism Xj for interaction with Xi (i 6= j). In this case, Xi intends to benefit

SC
from Xj , and Xj neither gain or loss from the interaction. The interaction with
Xj and Xbest tries to improve the quality of fitness of design vector Xi and

U
increase the exploitation ability of the algorithm respectively. The interaction
is mathematically modeled by Equation 11. Xbest represents the organism with
AN
best fitness value similar to that of mutualism phase. Xi is updated to Xi∗ as
computed in Equation 11, if the fitness value f (Xi∗ ) is better that of f (Xi ). The
relationship for updating Xi is given by Equation 12.
M
Xi∗ = Xi + U (−1, 1) ∗ (Xbest − Xj ) (11)
D

where U (−1, 1) is a vector of uniformly distributed random numbers between

−1 and 1. i = 1, 2, 3, ..., ecosize; j ∈ {1, 2, 3, ..., ecosize|j 6= i}; ecosize is the
TE

number of organisms in the search space.


 X ∗ if f (X ∗ ) > f (X )
i i i
Xi = (12)
EP

∗
 X
i if f (Xi ) ≤ f (Xi )

260 4.3. Parasitism phase

C

In parasitism phase, an artificial parasite called parasite vector is created by

cloning an ith organism Xi and modify it using randomly generated number.
AC

Then, Xj is randomly selected from ecosystem, and fitness values of parasite

vector and Xj are computed. If the parasite vector is fitter than Xj , then Xj is
replaced by the parasite vector, otherwise Xj survives to the next generation of
ecosystem and parasite vector is discarded. Xj is updated according to relation
in Equation 13. This phase of the increase the exploitation and exploration

14
 ACCEPTED MANUSCRIPT

of the algorithm by randomly removing the inactive solution and introducing

the active ones. Consequently, premature convergence could be avoided and

PT
convergence rate could be improved.

 PV if f (P V ) > f (Xj )
Xj = (13)
 X
j if f (P V ) ≤ f (Xj )

RI
where P V denotes the parasite vector.

SC
5. Chaotic maps

Chaos is a deterministic process which is usually found in dynamic and non-

linear systems, and has high sensitivity to initial conditions and parameters

U
(Kasahara and Yonezawa, 1996). It is characterized by randomness, ergodic-

AN
ity, irregularity and an apparently unpredictable. Chaotic sequences have been
employed in stochastic optimization techniques to provide population diversity
in search space to ensure global convergence and avoidance of local optima
M
entrapment (Wu et al., 2013). Recently, relatively better results have been
obtained by applying chaotic sequence rather than random sequence based op-
D

timization techniques to various real-word optimization problems (Wu et al.,

2013; Le Hoang, 2014; Abdollahzade et al., 2015; Adarsh et al., 2016; Suresh
TE

and Lal, 2017). Overview of chaotic maps commonly applied to optimization

problems can be found in Gandomi and Yang (2014). Logistic chaotic model is
employed to improve the population diversity and reduce the convergence time
EP

of the proposed algorithms because of its success in solving various optimization

problems (Shayeghi and Ghasemi, 2014; Rajagopalan et al., 2015; Secui, 2016).
The logistic chaotic model has relative uniform behaviour and has no cyclic phe-
C

nomenon in the course of iteration as well as better chaotic distribution features

AC

(Wu et al., 2013). The logistic model given as Equation 14 (Alatas, 2010).

yn+1 = λ × yn (1 − yn ) n = 0, 1, 2, 3, ... (14)

where {yn }n=1,2,3,... represents the sets of numbers generated with logistic chaotic
map; λ ∈ [0, 4] is the control parameter of logistic equation; yn ∈ (0, 1) is the
265 nth chaotic number and y0 ∈ (0, 1) and y0 ∈
/ {0.0, 0.25, 0.5, 0.75, 1.0}.

15
 ACCEPTED MANUSCRIPT

Algorithm 1 Symbiotic Organisms Search Algorithm (Cheng and Prayogo,

2014)

PT
Input: Set ecosize, create population of organisms Xi , i = 1, 2, 3, ..., ecosize,
initialize Xi , Set stopping criteria.
Output: Optimal schedule

RI
1: Identify the best organism Xbest
2: while stopping criterion is not met do

SC
3: for i = 1 to ecosize do
4: Mutualism Phase
Xi +Xj
5: MV = 2 . (j 6= i)

U
6: Xi∗ = Xi + U (0, 1) ∗ (Xbest − M V ∗ β1 ) . (β1 , β2 ) : benefit factors
7: Xj∗ = Xj + U (0, 1) ∗ (Xbest − M V ∗ β2 )
8:

9: Xi = Xi∗
AN
if F (Xi∗ ) < F (Xi ) then

10: end if
M
11: if F (Xj∗ ) < F (Xj ) then This could be
12: Xj = Xj∗
D

13: end if
14: Commensalism Phase
TE

15: Xi∗ = Xi + U (−1, 1) ∗ (Xbest − Xj )

16: if F (Xi∗ ) < F (Xi ) then
Xi = Xi∗
EP

17:

18: end if
19: Parasitism Phase
C

20: Create parasite vector

21: if F (parasite vector) < F (Xj ) then
AC

22: Xj = parasitev ector

23: end if
24: Identify the best organism Xbest
25: end for
26: end while

16
 ACCEPTED MANUSCRIPT

5.1. Chaotic local search

The Chaotic Local Search (CLS) strategy have been used to enable global

PT
best individuals to jump out of likely local optima. Thus, CLS is performed on
all solutions in the archive to avoid local Pareto Fronts. For each solution Ai in
the archive A, new feasible solution Ei is generated using logistic chaotic model.

RI
First, Ai is assigned to Ei (ei,1 , ei,1 , ei,1 , ..., ei,D ) and Ei is mapped into initial
vector X 0 = (x01 , x02 , x03 , ..., x0D ) in the range [0, 1] using Equation 15. Then,

SC
the chaotic sequence variable Xj by the iteration of logistic chaotic model as
described in Section 5. Thereafter, the new solution Ei obtained by scaling
the chaotic variable Xj into the original search space according to Equation 16.

U
Later, the objectives of the new solution Ei is evaluated and Ei is added to set
Q. The procedure of CLS is presented as Algorithm 2.

x0j =
AN
ei,j − emin,j
; j = 1, 2, 3, ..., D (15)
emax,j − emin,j
M
where emin,j and emax,j are the minimum and maximum bounds of jth dimen-
sion respectively.
D

ei,j = emin,j + (emax,j − emin,j ) × xj ; j = 1, 2, 3, ..., D (16)

TE

6. Multi-objective symbiotic organisms search for task scheduling op-

EP

timization algorithm

270 The original SOS algorithm was proposed for solving single objective contin-
C

uous optimization problem, but multi-objective task scheduling problem consid-

ered in this paper is a discrete optimization problem, it is virtually impossible to
AC

apply SOS algorithm directly for task scheduling problem on IaaS cloud. Thus,
new set of search operators based on the task scheduling problem features are de-
275 signed which include organism encoding scheme presented in Section 6.1, organ-
ism decoding scheme scheme presented in Section 6.3, ecosystem initialization
using chaotic logistic sequence presented in Section 6.2 to improve ecosystem

17
 ACCEPTED MANUSCRIPT

PT
RI
SC
Algorithm 2 Chaotic Local Search
Input: An archived individuals A
Output: Locally optimized archived individuals E

U
1: for each Ai ∈ A do
2: Let Ei ← Ai
problem.
AN
. Ei = (ei,j |j = 1, 2, 3, ..., D) D is the dimension of the

3: Scale ei,j into the range [0, 1] according to Equation 15 to obtain initial
M
vector X 0
4: Generate a chaotic sequence variable X by the iteration of logistic chaotic
model using Equation 14.
D

5: Obtain the locally optimized solution Ei by scaling the chaotic variable

TE

X into the original search space according to Equation 16.

6: Evaluate all the objectives of Ei
7: end for
EP

 
8: si∗ ← xi∗

9: return Xi∗
C
AC

18
 ACCEPTED MANUSCRIPT

diversity. The random number components of mutualism, and commensalism

operators are replaced with chaotic sequences to improve the global convergence

PT
280 of CMSOS algorithm, the mutualism, commensalism, and parasitism operators
are described in Subsections 6.4.1, 6.4.2, and 6.4.3 respectively.

RI
6.1. Organism encoding

In the proposed algorithm, the population structure of the organisms is rep-

SC
resented as the set of instance types, each organism is an individual in the
285 population that represents a part of the search space. Each coordinate (each
field) in an organism coordinate system is a instance types in the IaaS cloud.

U
In d-dimensional solution search space, a search population of n organisms is
denoted as X = {X1 , X2 , X3 , ..., Xn }. The position of the ith organism is de-

290
AN
noted as Xi = {xi1 , xi2 , xi3 , ..., xid }. To define the solution representation for
the problem, each organism represents a complete task schedule, thus the di-
mension of an organism is same as the number of tasks. The real values are used
M
to represent alternative instance type to be selected. The coordinate system for
determining the position of an organism in the solution search space is depen-
D

dent on the dimension of the organism. As an illustration, an organism depicted

295 in Figure 1 represents a schedule with 7 tasks. The organism is a 7-dimensional
TE

one and its position on the search space is defined by coordinates 1 through 7.
The range of movement of the organisms is determined by the number of
available instances to execute the tasks. Therefore, the value of the coordinates
EP

will be from one to number of available instances. Since the fitness computation
for the selected VM is based on discrete values, the nearest integer value of each
C

coordinate in an organism’s position corresponds to an instance type to execute

the task defined by that specific coordinate. The nearest integer can be obtained
AC

using Equation 17. In this manner, the organisms’s position encodes a task to
instance type mapping. Looking at the example in Figure 1, three instances are
in the resource pool so each coordinate takes a value in the range of 0 to 3. The
value 0.7 of coordinate 1 indicates that task 1 is assigned to instance type 2.
The value 2.8 of coordinate 2 indicates that task 2 is assigned to instance type

19
 ACCEPTED MANUSCRIPT

3. The remaining coordinates follow the same logic.

dxi e (17)

PT
where xi is the value of coordinate number i.

RI
U SC
AN
M
Figure 1: Organism encoding and corresponding task to VM mapping

6.2. Ecosystem initialization

D

In the standard SOS algorithm, uniformly generated random numbers are

TE

300 used to initialize the ecosystem (population) and as a source of randomness in

updating the positions of the organisms. Recently, usage of chaotic maps as
a source of randomness in optimization theory and various fields have gained
EP

wide attention instead of the usual random process. The logistic chaos model
described in Section 5 is used for generation of chaotic sequence.
C

305 6.3. Organism decoding

Corresponding to encoding method, the decoding of an individual organ-

AC

ism into a task schedule is presented as Algorithm 4. At the start, the set of
instances for executing the submitted tasks and set of task to instance type
mappings are initialized to empty sets. Then, the algorithm loops through ev-
310 ery organism’s coordinates contained in vector s to determine task to instance
type assignment related to the current organism’s coordinate and sets I and M

20
 ACCEPTED MANUSCRIPT

Algorithm 3 Individual Organism Encoding

Output: An individual X = (x, s) . x = {x1 , x2 , x3 , ..., xn }; n is the number
of tasks to be scheduled.

PT
1: Initialize a vector y0 . y = {y1 , y2 , y3 , ..., yn }; yi ∈ (0, 1)
2: Generate chaotic sequence y using iteration of logistic chaotic model in

RI
Equation 14
3: Transform the chaotic sequence into the range of parameters of task schedul-

SC
ing model according to: xi = xmax + (xmax − xmin ) × yi ; i = 1, 2, 3, ..., n
4: Transform the organism coordinates into task schedules using Equa-
tion 17: si = dxi e; i = 1, 2, 3, ..., n

U
are filled accordingly. This is achieved by using the organism encoding strategy
AN
explained earlier, that is a coordinate i represents task ti and its value si indi-
cate the code of instance type. At this point, the leased instance type vsj (ti ) to
execute task ti , the starting lease time S(ti , vsj ) and finishing lease time F (ti , vsj )
M
315

of instance type vsj (ti ) are obtained as the elements of the tuple Ii . Then, the
algorithm compute the start time Sti and finish time Fti of task ti .
D

The start time Sti is the available time of instance that task ti is assigned,
which is the ready time S(ti , vssi ) of the instance type with index si . The finish
TE

320 time Fti of task ti is computed according to the total running time and start time
of a task, this is obtained by summing e(i, si ) and Sti . At this point, the three
EP

tuples of task to VM mapping have been computed. Then algorithm update

the time to finish execution of task ti assigned to a VM with index si . The
finishing time Fsi is obtained by summing the processing time e(ti , vssi ) of task
C

325 ti and the start time vssi of an instance with index si . After all the coordinates
have been processed by the algorithm, instances to be leased are contained in
AC

I and their start and stop times as well which are used to compute the cost of
execution according to Equation 3. In addition, task to instance assignments
are contained in M with their start and finish times and these information are
330 be used to compute makespan using Equation 2. At this point, the feasible
schedule an organism is obtained along with the fitness values of the objectives.

21
 ACCEPTED MANUSCRIPT

6.4. Organism position update for task scheduling

PT
The candidate solutions are represented by ecosystem (population) of or-
335 ganisms, while mutualism, commensalism, and parasitism operators to direct

RI
the search process by candidate solutions. Each organism is represented by a
coordinate system in the search space, and organisms keep update of global best

SC
position Xbest which is determined based on the fitness function of the problem
at hand. The fitter organisms are allowed to proceed to the next generation
340 of potential solution while the unfitted organisms are discarded. The fitter or-

U
ganisms are those with good solution while the unfitted organisms holds bad
solution. The positions of the organisms are then update towards the Xbest
AN
locations using mutualism, commensalism and parasitism phases respectively.
The rate of movement of organisms towards the Xbest locations are moderated
by chaotic random sequence to improve global search ability of the organisms.
M
345

The SOS operators are continuously applied on the population of organisms

which represents candidate solutions until the stopping criterion are reached.
D

The mutualism, commensalism, and parasitism operators are described in the

following subsections 6.4.1, 6.4.2, and 6.4.3 respectively.
TE

350 6.4.1. Mutualism operator

Suppose Xi is the ith member of the ecosystem, a design vector Xj is
EP

randomly selected from the ecosystem to interact with another design vector
Xi (i 6= j) for mutual benefit, M V defines the mutual relationship character-
istics as Equation. 18. The essence of the interaction is to improve extent of
C

survival of both Xi and Xj in the ecosystem. The new candidate solutions for
AC

Xi and Xj are obtained according to Equations. 19 and 20 respectively. The

new candidate solutions replaced the old ones if their fitness values are better
than those of the old ones. In this case, Xi∗ and Xj∗ replace Xi and Xj respec-
tively in the next generation of ecosystem. Otherwise, Xi∗ and Xj∗ are discarded
while Xi and Xj survives to the next generation of the ecosystem. This scenario

22
 ACCEPTED MANUSCRIPT

PT
Algorithm 4 Individual Organism Decoding
Input: An individual X = (x, s).

RI
Output: S = (I, M, makespan, cost), I : the set of leased instances; M is the
set of task to instance mappings;
1: Empty sets Ik ; k = 1, 2, 3, ..., m

SC
2: Empty sets Mk ; l = 1, 2, 3, ..., m
3: Initialize makespan ← 0

U
4: Initialize cost ← 0
5: for si ∈ s; i = 1, 2, 3, ..., n do
6:

7:
Let j ← si AN
Obtain a tuple Ii ←< vsj (ti ), S(ti , vsj ), F (ti , vsj ) >
8: Put Ii into set I
M
9: Assign the starting time of task ti as Sti ← F (ti , vsj ); F (ti , vsj ) is the
finishing time of instance type vsj
D

10: Compute estimated execution time of task ti on instance type vsj as

e(ti , vsj ) according to Equation 1
TE

11: Assign the finishing time of task ti as Fi ← e(i, j) + Sti

12: Obtain a tuple m(ti , vsj ) ←< ti , vsj (ti ), Sti , Fti >
13: Put m(ti , vsj ) into set M
EP

14: Update the finishing time of instance type vsj as F (ti , vsj ) ← e(i, j) +
S(ti , vsj ) ; S(ti , vsj ) is the starting time of instance vsj
C

15: end for

16: Compute makespan according to Equation 2.
AC

17: Compute cost according to Equation 3.

18: S = (I, M, makespan, cost)

23
 ACCEPTED MANUSCRIPT

is captured by Equations. 21 and 23.

Xi + Xj
MV ← (18)

PT
2

Xi∗ ← Xi + y1 ∗ (Xbest − M V ∗ β1 ) (19)

RI
Xj∗ ← Xj + y2 ∗ (Xbest − M V ∗ β2 ) (20)

SC

 X∗ if f (Xi∗ ) > f (Xi )
i
Xi = (21)
 X if f (Xi∗ ) ≤ f (Xi )

U
i

 dX ∗ e if f (Xi∗ ) > f (Xi )
si =


i
 dX e AN
i if f (Xi∗ ) ≤ f (Xi )
(22)

 X∗ if f (Xj∗ ) > f (Xj ); sj∗ ← xj∗

 
j
M
Xj = (23)
 X
j if f (Xj∗ ) ≤ f (Xj )
  
 X∗ if f (Xj∗ ) > f (Xj )
j
D

sj = (24)
 dX e
j if f (Xj∗ ) ≤ f (Xj )
TE

where f (.) denotes the fitness evaluation function; y1 and y2 are vectors of
chaotic sequence generated using chaotic logistic model as described in Section 5;
si and sj are the instance type index as defined in Equations 22 and 24
EP

that execute a given task, i = 1, 2, 3, ..., ecosize; j ∈ {1, 2, 3, ..., ecosize|j 6= i};
355 ecosize is the number of organisms in the search space. β1 and β2 denote the
benefit factors.
C

6.4.2. Commensalism operator

AC

In commensalism phase, an ith member of the ecosystem randomly selects an

organism Xj for interaction with Xi (i 6= j). In this case, Xi intends to benefit
from Xj , and Xj neither gain or loss from the interaction. The interaction with
Xj and Xbest tries to improve the quality of fitness of design vector Xi and
increase the exploitation ability of the algorithm respectively. The interaction

24
 ACCEPTED MANUSCRIPT

is modelled by Equation 25. Xbest represents the organism with best fitness
value similar to that of mutualism phase. Xi is updated to Xi∗ as computed in

PT
Equation 25, if the fitness value f (Xi∗ ) is better that of f (Xi ). The relationship
for updating Xi is given by Equation 26.

Xi∗ ← Xi + y1 ∗ (Xbest − Xj )

RI
(25)


 X∗ if f (Xi∗ ) > f (Xi ); si∗ ← xi∗
 

SC
i
Xi = (26)
 X
i if f (Xi∗ ) ≤ f (Xi )

 dX ∗ e if f (Xi∗ ) > f (Xi )

U
i
si = (27)
 dX e
i if f (Xi∗ ) ≤ f (Xi )

AN
where y1 is a vector of chaotic sequence generated using chaotic logistic model
as described in Section 5; i = 1, 2, 3, ..., ecosize; j ∈ {1, 2, 3, ..., ecosize|j 6= i};
360 ecosize is the number of organisms in the search space, si is the instance type
M
index as defined in Equation 27 that executes a given task.

6.4.3. Parasitism operator

D

In parasitism phase, an artificial parasite called parasite vector is created by

TE

cloning the current organism Xi denoted as Xi(cloned) and mutate the randomly
365 selected kth dimension of organism Xi(cloned) according to Equation 28. Then,
Xj is randomly selected from ecosystem, and fitness values of parasite vector
EP

and Xj are computed. If the parasite vector is fitter than Xj , then Xj is

replaced by the parasite vector, otherwise Xj survives to the next generation
of ecosystem and parasite vector is discarded. Xj is updated according relation
C

370 in Equation 29. This phase of the the search procedure to jump out of local
AC

optima by randomly removing the inactive solution and introducing the active
ones.

xik ← xmax + (xmax − xmin ) × y1 (28)

25
 ACCEPTED MANUSCRIPT


 X if f (Xi(cloned) ) > f (Xj )
i(cloned)
Xj = (29)
X if f (Xi(cloned) ) ≤ f (Xj )

PT

j
 
 xi(cloned)  if f (Xi(cloned) ) > f (Xj )
sj =  j (30)
 x if f (Xi(cloned) ) ≤ f (Xj )

RI
where y1 is a vector of chaotic sequence generated according to Equation 14;
Xi(cloned) denotes the parasite vector; xmax and xmin are the minimum and

SC
375 maximum values of the solution range respectively.

6.5. Archive update

U
The effectiveness of the mechanisms for selecting the non-dominated feasible
complete solutions that are contained in the archive facilitates the generation

380
AN
of good Pareto Fronts (PFs). In the course of optimization process, the size of
archive is fixed since the generation of non-dominated feasible complete solutions
M
grows fast. The archive maintains a set of feasible complete solutions, the
capacity of the archive is fixed as ecosize and number of current solution in
the archive is is denoted as e. A new feasible complete solution is added to the
D

archive, if the content of archive is not filled to its capacity. Otherwise, a new
feasible complete solution is added to the archive if it dominates a solution in the
TE

385

archive, in which case the new feasible complete solution replaces the solution it
dominates. To avoid local PFs, Chaotic Local Search (CLS) is performed on each
EP

solution in archive to obtain new solutions, the CLS is described in Section 5.1.
The current solutions in the archive and new generated solutions are combined
390 to obtain 2 × e solutions. Then, non-dominated solutions are determined, if
C

the number of non-dominated solutions are not more than the size of archive,
then all the non-dominated solutions are added as the current content of the
AC

archive and the current size is set as e. Otherwise, fast non-dominated sorting
and crowding distance are performed the combined solutions, the first e less
395 crowded solutions are chosen to be added as the current content of the archive
and e is set as ecosize. The selection procedure based on non-dominated sorting
and crowding distance as depicted in Figure 2.

26
 ACCEPTED MANUSCRIPT

6.6. Current ecosystem update

After each generation, the current organisms are combined with the advanced

PT
400 organisms form combined ecosystem. The combined ecosystem is obviously
larger than the ecosize. Thus, the following techniques are used to select the
ecosystem with the size ecosize for the next generation. First, organisms in the

RI
combined ecosystem is ranked into non-dominated sets (F1 , F2 , F3 , F4 , ...) using
fast non-dominated sorting. The non-dominated organisms belonging to set R1

SC
405 are selected first for addition into the current. If the size of F1 is smaller than the
ecosize, the rest number of organisms are selected from the non-dominated sets
in the order (F2 , F3 , F4 , ...). The procedure continues until the capacity of the

U
ecosize is filled. Suppose Fj is the last set of non-dominated solutions beyond
which no other set can be taken and total size of the sets F1 , F2 , F3 , ..., Fj is more
410
AN
than the ecosize. The optimal ecosystem of ecosize is selected using crowding
distance and selection of solutions are based on descending order of distance.
M
Overview of the procedure is depicted in Figure 2.

Non-dominated sorting
D

(t )
F1 F1
X X (t+1)
F2 F2
TE

Current Ecosystem
ecosystem for next
F3 Crowding distance F3 generation
F4
F4 F4
EP

Reject
Combined
ecosystem

X (new) F5
C

New
ecosystem
... Reject
AC

Fn

Figure 2: Procedure of ecosystem selection for next generation

27
 ACCEPTED MANUSCRIPT

6.7. Implementation of the proposed algorithm

The proposed CMSOS algorithm for large task scheduling optimization prob-

PT
415 lem in IaaS cloud computing environment is presented as Algorithm 5.
Ecosystem generation evolves each ecosystem using mutualism, commensal-
ism, and parasitism operators. The generation of new organisms in mutualism,

RI
commensalism, and parasitism operators are described in Section 6.4.1, 6.4.2,
and 6.4.3 respectively. Algorithm 6 describes the procedure of ecosystem evo-

SC
420 lution.

6.8. Complexity Analysis

U
The time complexity of each phases (Mutualism, Commensalism, and Par-
asititsm) is O(n), where n is the number of tasks. For a given n tasks, the

425
AN
time complexity of each evaluation of an organism is O(n). Therefore, the total
time complexity of the ecosystem evolution is O(egn), where e is the number
of organisms in the ecosystem, and g is the number of generations. Besides
M
the ecosystem evolution, the chaotic local search (CLS) is performed on evolved
organisms, the CLS procedure has a time complexity of O(e). Most of all, the
D

total time complexity of the proposed algorithm is O(egn + e). Hence, the dom-
inant time-consuming part of the proposed algorithm is the ecosystem evolution
TE

430

which has the time complexity of O(egn).

7. Performance evaluation and results analysis

EP

This section describes the experimental design, performance evaluation, and

result analysis of the proposed algorithms.
C

435 7.1. Experimental Design

AC

The proposed algorithm was implemented using CloudSim 3.0.3 (Calheiros

et al., 2011) simulation toolkit for IaaS cloud environment. The choice of
CloudSim simulation toolkit is informed by its support for simulating cloud
computing scenarios which supports modeling and simulation of large scale com-
440 puting environments. CloudSim provides support for modeling data centers,

28
 ACCEPTED MANUSCRIPT

Algorithm 5 Multi-objective Symbiotic Organisms Search for Task Scheduling

PT
Optimization
Input: CMSOS settings: ecosize N , maximum number of generations gmax .
Output: Optimal solutions.

RI
1: Step 1 Initialization: Set g = 0; Initialize y0 for generation of Lo-
gistic chaotic sequence {yn|n=1,2,3,... }; Generate initial ecosystem X (g) =
(g)
{Xi|i=1,2,3,...,ecosize } using Algorithm 3;

SC
2: Step 2 Generation:
3: Step 2.1: Decode organisms in X (0) using Algorithm 4 and evaluate the

U
fitness of each organisms in X (0) ;
4: Step 2.2: Perform non-dominated sorting on individual organisms as de-
AN
scribed in Section 6.6; Randomly select any organism in rank R1 as the
best organism Xbest ; Select non-dominated feasible organisms into external
archive, and set the capacity of external archive to ecosize.
M
5: Step 2.3 Ecosystem Evolution:
6: for g = 1 to gmax do
D

(g)
7: Step 2.3.1: Use Algorithm 6 to generate new ecosystem Xcom .
(g)
8: Step 2.3.2: Perform non-dominated sorting on Xcom ecosystem as de-
TE

scribed in Section 6.6.

9: Step 2.3.3 Archive Update: For each feasible individual organism
0 (g)
X in ecosystem Xcom , replace an organism X ∗ in the archive with the
EP

(g)
organism X 0 in ecosystem Xcom , if X 0 dominates X ∗ ; Perform Chaotic
Local Search (CLS) on current archive as described in Section 5.1.
C

10: Step 2.3.4 Current ecosystem selection: Perform non-dominated

(g)
sorting and crowding distance on ecosystem Xcom as described in Sec-
AC

(g+1)
tion 6.6. Then, select top ecosize organisms into Xcom as the current
ecosystem.
11: end for
12: Step 3: Output the content of archive as the optimal solutions.

29
 ACCEPTED MANUSCRIPT

Algorithm 6 Ecosystem Evolution

(g)
Input: ecosystem X (g) = {Xi|i=1,2,3,...,ecosize }, maximum number of genera-

PT
tions gmax .
Output: Eevol ecosystem.
1: Em ← ∅; Ec ← ∅; El ← ∅; Ecom ← ∅; set g = 0.

RI
2: Step 1 Initialization: Decode the organisms in X (g) using Algorithm 4
and evaluate the fitness of each individual organism; Then, perform non-

SC
dominated sorting on X (g) as described in Section 6.6; Randomly select any
(g)
organism in rank R1 as the best organism Xbest in X (g) ;
3: for i = 1 to N do

U
(g)
4: Mutualism Phase: Randomly select an organism Xj from the
(g)
current generation of organisms to interact with current organism Xi ,
AN
j ∈ {1, 2, 3, ..., N } and j 6= i; Initialize the benefit factors β1 and β2, let
β1 and β2 be assigned a randomly selected number 1 or 2; Compute the
M
mutual vector M V using Equation 18; Update the values of the chaotic
(g) (g)
sequence yni and ynj using Equation 14; Modify organisms Xi and Xj
(g) (g)
using Equations 19 and 20 respectively to obtain Xi(new) and Xj(new) ;
D

(g) (g)
Em = Em ∪ Xi(new) ∪ Xj(new) .
(g)
Commensalism Phase: Randomly select an organism Xj from the
TE

5:
(g)
current generation of organisms to interact with current organism Xi ,
j ∈ {1, 2, 3, ..., N } and j 6= i; Update the values of the chaotic sequence
EP

(g)
yni using Equation 14; Modify organisms Xi using Equation 26 to obtain
(g) (g)
Xi(new) ; Ec = Ec ∪ Xi(new) .
(g) (g)
6: Parasitism Phase: Clone the current organism Xi as Xi(cloned) ; Mu-
C

(g)
tate a randomly selected kth dimension of Xi(cloned) according to Equa-
(g)
tion 28; Randomly select an organism Xj from the current generation of
AC

(g) (g)
organisms; Decode Xi(cloned) and Xj organisms using Algorithm 4; Evalu-
(g) (g) (g) (g)
ate the fitness of Xi(cloned) and Xj as F (Xi(cloned) ) and F (Xj ); Update
(g)
Xj for next generation of organisms according to Equations 29 and 30;
(g)
Ep = Ep ∪ Xj(new) .
7: end for
8: Eevol = Em ∪ Ec ∪ Ep
30
9: return Eevol
 ACCEPTED MANUSCRIPT

physical machine hosts, VMs, cloud service brokers, and scheduling systems.
In the experiment, an IaaS cloud provider with a single data center, 2 hosts,

PT
and 20 VMs of different configurations (Rodriguez and Buyya, 2014; Li et al.,
2016). The configurations of the data center and the hosts are presented in
445 Table 1. The VM configurations are based on the current Amazon EC2 offer-

RI
ings (https://aws.amazon.com/ec2/pricing/) as presented in Table 2. The VM
processing capacity of VMs in MFLOPS based on the work of Ostermann et al.

SC
(2009). The workload parameters for tasks are presented in Table 3.

Table 1: Experimental Settings

U
Cloud Entity Parameter Value
Datacenter Number 1

RAM
AN
Number 2
2048000 MB
Storage 1000000 MB
M
Host Bandwidth 1000000000 Mb/s
Operating System Linux
D

Architecture x86
VMM Xen
TE

Number 20
VM
Bandwidth 0.1 GB/s
EP

The Parallel Workloads Archive, whose data is the focus of this paper, is a
450 repository of such logs; it is accessible at URL www. cs.huji.ac.il/labs/parallel/workload/.
C

The archived logs (see Table 4) contain accounting data about the jobs that ex-
ecuted on parallel supercomputers, clusters, and grids, which is necessary in
AC

order to evaluate schedulers for such systems. These logs have been used in
many hundreds of research papers since the archive was started in 1999.
455 Both standard parallel workload traces and synthetic workloads are used to
evaluate the performance of the proposed algorithms. It is assumed that tasks
are independent that is no precedence constraints between tasks, and execution

31
 ACCEPTED MANUSCRIPT

Table 2: Configurations and Types of VMs

Name vCPU Processing Memory (GiB) SSD Stor- Cost per hour ($)

PT
capacity age (GB)
(MFLOPS)

RI
c3.large 2 8 800 3.75 2 × 16 0.105
c3.xlarge 4 17 600 7.5 2 × 40 0.210
c3.2xlarge 8 35 200 15 2 × 80 0.420

SC
c3.4xlarge 16 70 400 30 2 × 160 0.840
c3.8xlarge 32 140 800 60 2 × 320 1.680

U
Table 3: Workload Settings

Parameter
Length
AN Value
[5000, 50 000] MFLOPS
File size [10, 100] GB
M
Memory [10, 100] GB
D

of tasks are non-preemptive. The parallel workloads used for evaluation are
NASA Ames iPSC/860 and HPC2N; the workloads are accessible through the
TE

460 URL http://www.cs.huji.ac.il/labs/parallel/workload/. NASA Ames iPSC/860

and HPC2N set log are some of the popular standard formatted workloads
for evaluating the performance of distributed systems (Feitelson et al., 2014;
EP

Wang et al., 2016; Alla et al., 2017). The information about the log are shown
in Table 4. The synthetic workloads are generated using normal and uniform
C

465 distribution. Uniform distribution depicts more medium size tasks, and fewer
small and large size tasks. Uniform distribution depicts an equal number of
AC

large, medium, and small size tasks. The larger instances will enable us to
gain insight into the scalability of performance of the algorithms with large
problem sizes. Besides the standard workload traces, the synthetic workloads
470 are generated using normal and uniform distribution respectively. Normally
distributed workloads distribution depicts medium size tasks, and fewer small

32
 ACCEPTED MANUSCRIPT

and large size tasks. Uniformly distributed workloads depicts equal number of
large, medium, and small size tasks.

PT
Table 4: Logs in the Parallel Workloads

Log Period Months PEs1 Users Jobs Util.2 File3

RI
NASA iPSC 10/93-12/93 3 128 69 42,264 0.47 NASA-iPSC-1993-3.swf
HPC2N 07/02-01/06 42 240 258 527,371 0.70 HPC2N-2002-2.swf

SC
1
was nodes or CPUs in old logs, today it typically represents cores.
2
is the system utilization, i.e. the fraction of the resources that were allocated to jobs. It is not com-
puted for SHARCNET because this is a grid system, and the constituent clusters became available

U
at different times.
3
AN
File names include a version number, as most logs were re-converted to swf when errors were found
or new considerations were introduced.
M
7.1.1. Performance Metrics
475 This study used makespan, cost (financial cost), Hypervolume, and Percent-
D

age change as performance metrics to evaluate the proposed algorithms against

similar task scheduling algorithms in the literature. Makespan is the latest fin-
TE

ish time of VMs, minimal makespan implies that users pay moderate cost for
their task execution since cloud service offering is based on per-use-model and
480 users are charged per unit time of VM usage usually per-hour. Cost is the cost
EP

of leasing VMs from IaaS cloud providers.

Makespan also referred to as total execution time is the latest finish time
of all the VMs used in executing the collection of tasks as defined in Equa-
C

tion 31(Netjinda et al., 2014).

AC

makespan = max{vmtime
j ; j = 1, 2, 3, ..., m} (31)

485 where m is the number of VMs; vmtime

j is the total execution time of VM j.
Cost is the sum of the product of VM cost by its makespan rounded to the

33
 ACCEPTED MANUSCRIPT

closest integer as define in Equation 32 Zheng and Sakellariou (2013).

m
X

PT
cost = dvmtime
j evmcost
j (32)
j=1

where m is the number of VMs; vmtime

j is the total execution time of VM j;

RI
vmcost
j is the financial cost for leasing VM j per unit time.
490 Hypervolume (HV) indicator (Zhu et al., 2016; Zhang et al., 2017) is the
most popular performance metric for this sake. HV is obtained by computing

SC
the volume of the objective function space between the obtained non-dominated
solutions and a reference point, by providing an insight between the convergence
and diversity of the solution sets. HV is obtained as a union of all the found

U
495 hypercubes according to Equation 33. To obtain the HV values, each algorithm
AN
is run on all the workload instances for 30 independent runs and solutions ob-
tained by each algorithm for the 30 runs are merged to form a reference set, then,
non-dominated solutions are selected to the reference set to form true Pareto
M
Front (PF) and results dominated by true PF are discarded (Zitzler et al., 2003).
500 Then, the makespan and cost are normalized, a reference point (1.1, 1.1) is used
D

to compute the values of HV (Ishibuchi et al., 2010; Zhu et al., 2016).

TE

|R|
!
[
HV = volume vj (33)
j=1

The percentage change for each proposed algorithm is computed with respect
EP

to the compared algorithms from the literature as in Equation 34. This provides
an insight on the extent of the performance of the proposed algorithm against
505 the existing algorithms in the literature (Vincent et al., 2017).
C

(Zprop − Zlit )
AC

gap(%) = (34)
Zlit
where Zprop is the solution obtained by the proposed algorithm and Zlit is the
solution obtained one the algorithm are reported in the literature. A negative
indicates that proposed algorithm is better.

34
 ACCEPTED MANUSCRIPT

The results of the proposed MSOS algorithm are compared with EMS-C
510 (Zhu et al., 2016), ECMSMOO (Yao et al., 2016), and BOGA (Zhang et al.,

PT
2017) multi-objective task scheduling algorithms using the same workload trace
(Table 1) and on the same test bed (Tables 2 and 3). The compared algorithms
are chosen so as to compare the proposed techniques against the recent tech-

RI
niques in the area. Besides, the goal of compared algorithms are identical to
515 the goal of the proposed techniques. For fair comparison, the stopping condi-

SC
tion for compared algorithms and proposed are taken to be same. Each of the
algorithms are used for solving the workload instances over 30 independent runs
(Zhu et al., 2016).

U
Table 5: Parameter Settings for Compared Algorithms

Algorithm AN
Parameter
Crossover rate Pc
Value
1.0
EMS-C
Mutation rate Pm 1/n
M
Crossover rate Pc 0.5
BOGA Mutation rate Pm 0.5
D

Social learning factor c1 2

Personal learning factor c2 2
TE

ECMSMOO Variable inertia weight ω 0.9-0.4

EP

7.2. Results Analysis and Discussion

520 Section 7.2.1 presents the results analysis and discussion of CMSOS algo-
rithm.
C

7.2.1. Comparison of CMSOS results with compared algorithms

AC

This section begin with discussion on the benefit of using chaotic optimiza-
tion strategy within the proposed algorithm. The proposed algorithm with
525 (denoted as CMSOS) and without (denoted as MSOS) chaotic optimization
strategy. To ensure a fair comparison between the CMSOS and MSOS, the
initial ecosystem (population), number of generations, stopping criteria and

35
 ACCEPTED MANUSCRIPT

hardware resources are same for the workload instances. Both algorithms (CM-
SOS and MSOS) are run for 30 runs over all workload instances. Thereafter,

PT
530 the performance of the proposed CMSOS is assessed by comparing it against
other algorithms. To ensure a fair comparison, the parameter values and the
termination condition of EMS-C, ECMSMOO, and BOGA are fixed same as

RI
CMSOS.
The non-dominated solutions for workloads instances of 5000 sizes are pre-

SC
535 sented in Figure 3a to Figure 4b. As it can be observed from the figures, CMSOS
performs remarkably better than EMS-C, ECMSMOO, and BOGA algorithms.
The remarkable performance of CMSOS algorithm is attributed to the global

U
convergence of underlying MSOS algorithm, the incorporation of chaotic op-
timization strategy into MSOS ensures diversity among the organisms which
540

dle large search space.

AN
further enables the algorithm to achieve better convergence and effectively han-
M
The Hypervolume improvements for CMSOS algorithm over the compared
algorithms are given in Figure 5a to Figure 6b. From the figures, it can be ob-
served that CMSOS algorithm have a significant Hypervolume improvement over
D

545 the EMS-C, ECMSMOO, and BOGA algorithms for all the workload instances.
CMSOS obtain performance improvement over EMS-C ranging 8.72% to 19.55%
TE

across the workloads, while the performance improvement over ECMSMOO is

between 11.51% to 23.70%. Moreover, the percentage improvement over BOGA
EP

is between 9.52% to 28.72%. Besides, CMSOS showed noticeable improvement

550 of 5.43% to 14.41% over the MSOS without the incorporation of chaotic se-
quence which showcased the effectiveness of chaotic optimization strategy.
C

Furthermore, the computational time (in seconds) taken by CMSOS to ob-

AC

tain the best result is compared with EMS-C, ECMSMOO, and BOGA for all
555 the tested workload instances. Table 6 shows the computational times (in sec-
onds) of CMSOS compared to EMS-C, ECMSMOO, and BOGA. It is clear
from Table 6 that the computational time of CMSOS is lower that of EMS-C,
ECMSMOO, and BOGA for all the tested workload instances. The reported

36
 ACCEPTED MANUSCRIPT

250
●
●
●
● ● CMSOS
● MSOS
EMS−C
200

PT
●
●
● ECMSMOO
● ●
BOGA
●
150

●
Cost ($)

● ●

RI
● ●
● ●●
100

●●
●● ● ●
●
●●
●● ●

SC
●●
● ●●
●●●
50

● ●
●● ● ●
●●● ●●
●
●
●●
0

U
0 500 1000 1500 2000 2500
AN Makespan

(a) NASA
M
500

●
●
● ● CMSOS
D

● ● MSOS
EMS−C
400

● ECMSMOO
TE

●
● ● ●
BOGA
●
300
Cost ($)

● ● ●●
● ●
EP

●●
200

●
●
●●
●●●
●● ● ●
● ●●
●
100

●● ●
●●●
●
AC 0

0 500 1000 1500 2000 2500

Makespan

(b) HSPC2N

37

Figure 3: Obtained non-dominated solutions by CMSOS for Real Parallel Workloads

 ACCEPTED MANUSCRIPT

750 1000
●

●
●
●
●●
● CMSOS
●
● ●
● MSOS
●
●●
● EMS−C

PT
● ● ECMSMOO
● ● ●
●
BOGA
● ●
600

● ●●
●●
Cost ($)

●●● ●● ●
●

RI
●
● ● ●●
●
●
300

●
●● ●● ●
●
●● ●
● ●
● ●●
●

SC
● ●
●●●
150

●●
●●
●● ●
●
0

U
0 1000 2000 3000 4000 5000
AN Makespan

(a) Random
M
750 1000

●
●
● ● CMSOS
D

●●
● ●●
● ●● MSOS
●● ●
● EMS−C
● ● ECMSMOO
TE

●
●● ● ●
BOGA
●
● ●
600

● ●●
●●
Cost ($)

●●● ●● ●
●
●● ● ●
● ●●
EP

●
●
300

●
●● ●● ●
● ●
●● ●
● ●
●● ●●
●
● ●
●●●
150

● ●●
●●
C

●● ●
●
●
●●
AC 0

0 1000 2000 3000 4000 5000

Makespan

(b) Uniform

38

Figure 4: Obtained non-dominated solutions by CMSOS for Synthetic Workloads

 ACCEPTED MANUSCRIPT

PT
20 CMSOS
HV improvement of CC−CMSOS (%)

EMS−C
ECMSMOO
BOGA

15

RI
10

SC
5

U
1000 2000 3000 4000 5000
Number of Tasks
AN
(a) NASA
M
25 CMSOS
HV improvement of CC−CMSOS (%)

EMS−C
ECMSMOO
BOGA
20
D

15
TE

10

0
EP

1000 2000 3000 4000 5000

Number of Tasks
C

(b) HSPC2N
AC

Figure 5: Convergence and Diversity Performance of CMSOS on Real Parallel Workloads

39
 ACCEPTED MANUSCRIPT

PT
CMSOS
HV improvement of CC−CMSOS (%)

25
EMS−C
ECMSMOO
BOGA

20

RI
15

10

SC
5

U
1000 2000 3000 4000 5000
Number of Tasks
AN(a) Random
M
CMSOS
HV improvement of CC−CMSOS (%)

EMS−C

20 ECMSMOO
BOGA

15
D

10
TE

0
EP

1000 2000 3000 4000 5000

Number of Tasks
C

(b) Uniform
AC

Figure 6: Convergence and Diversity Performance of CMSOS on Synthetic Workloads

40
 ACCEPTED MANUSCRIPT

results reveal that the proposed CMSOS produces better quality solutions with
560 lower computational time as compared to other algorithms. This proves that

PT
CMSOS is an effective and efficient solution method for solving large scale task
scheduling optimization problems.

Table 6: Running times of CC-CMSOS algorithm and compared algorithms

RI
Instances EMS-C ECMSMOO BOGA CMSOS CC-CMSOS

SC
1000 14.50 17.95 18.17 13.03 0.31
2000 36.34 36.72 37.01 35.35 1.48
HPC2N 3000 58.65 59.31 59.89 56.53 2.82

U
4000 72.86 73.66 73.74 70.10 5.12
5000 85.06 87.22 91.46 84.85 8.03
1000
2000
14.21
29.51
AN 22.78
38.97
23.53
41.99
12.32
28.14
0.44
1.26
NASA 3000 48.13 48.62 49.80 42.49 5.02
M
4000 61.00 67.91 68.34 51.52 6.38
5000 75.25 77.08 90.83 73.50 7.08
D

1000 18.41 18.53 25.41 7.81 0.74

2000 31.57 33.08 36.98 30.20 2.37
TE

Uniform 3000 38.44 39.29 44.23 38.00 3.85

4000 53.96 67.25 71.81 44.91 5.14
EP

5000 75.13 92.84 96.18 72.77 9.72

1000 12.27 12.87 20.81 6.80 1.06
2000 30.52 43.44 60.84 21.16 2.72
C

Random 3000 70.80 72.44 78.54 66.70 3.39

4000 86.70 91.37 93.69 81.88 6.14
AC

5000 97.65 104.66 110.41 96.26 8.35

The results showed that CMSOS algorithm gain better convergence and so-
lution diversity, thus leading to global solution. Better convergence is derived
565 from global convergence of MSOS algorithm. The principal feature that ensures

41
 ACCEPTED MANUSCRIPT

global convergence in both CMSOS and MSOS algorithms is introduction of

chaotic maps for generating initial solutions and replacement of random number

PT
components of the SOS algorithm which increases diversity among organisms
which represents candidate solutions. Another reason for better global con-
570 vergence of the CMSOS algorithm is the commensalism association exhibited

RI
by organisms which encourages elitism among organisms. The commensalism
mechanism and chaotic sequence strategy improves local search and global con-

SC
vergence of the proposed algorithm. The application of chaotic local search
strategy on Pareto Front tried to avoid possible entrapment in local optima.
575 Furthermore, the archive maintenance when the archive is filled to capacity

U
and ecosystem selection using non-dominated sorting and crowding distance im-
proves the coverage of the Pareto optimal front in the course of the optimization
AN
process. Overall, the above revealing results justifies the benefit of incorporat-
ing chaotic optimization strategy within the proposed algorithm. So, the use of
M
580 chaotic optimization strategy can efficiently enhance the search performance to
obtain better solutions for all tested workload instances.
D

8. Conclusion and future work

TE

In this paper, a multi-objective symbiotic organisms search algorithm with

a chaotic optimization strategy for addressing task scheduling problem is pro-
585 posed. The experimental results of both the standard and synthetic workload
EP

instances indicates the appropriateness of the proposed algorithm for producing

task schedules. The proposed algorithm consistently produced task schedules
with better makespan and cost with respect to the compared algorithms, for all
C

the workload instances studied. Chaotic optimization was employed to generate

AC

590 initial ecosystem (population) for effective ecosystem diversity to ensure better
global convergence. Moreover, new operators for the phases of SOS were de-
signed to further ensure global solutions for task scheduling problem. Finally,
chaotic local search was hybridized with the proposed algorithm to empower
CMSOS with the exploitative ability to complement the explorative power of

42
 ACCEPTED MANUSCRIPT

595 underlining SOS algorithm. The proposed algorithm can be extended to han-
dle other QoS requirements like reliability and security for very large workload

PT
instances.

Acknowledgement

RI
This work is supported by UTM/RUG/15H99 RMC Universiti Teknologi
600 Malaysia.

SC
References

U
Abdollahzade M, Miranian A, Hassani H, Iranmanesh H. A new hybrid en-
hanced local linear neuro-fuzzy model based on the optimized singular spec-

605
AN
trum analysis and its application for nonlinear and chaotic time series fore-
casting. Information Sciences 2015;295:107–25.
M
Abdullahi M, Ngadi MA. Hybrid symbiotic organisms search optimization al-
gorithm for scheduling of tasks on cloud computing environment. PloS one
D

2016;11(6):e0158229.

Abdullahi M, Ngadi MA, Dishing SI. Chaotic symbiotic organisms search for
TE

610 task scheduling optimization on cloud computing environment. In: Student

Project Conference (ICT-ISPC), 2017 6th ICT International. IEEE; 2017. p.
EP

1–4.

Abdullahi M, Ngadi MA, et al. Symbiotic organism search optimization based

task scheduling in cloud computing environment. Future Generation Com-
C

615 puter Systems 2016;56:640–50.

AC

Adarsh B, Raghunathan T, Jayabarathi T, Yang XS. Economic dispatch using

chaotic bat algorithm. Energy 2016;96:666–75.

Alatas B. Chaotic bee colony algorithms for global numerical optimization.

Expert Systems with Applications 2010;37(8):5682–7.

43
 ACCEPTED MANUSCRIPT

620 Alla HB, Alla SB, Ezzati A, Mouhsen A. A novel architecture with dynamic
queues based on fuzzy logic and particle swarm optimization algorithm for

PT
task scheduling in cloud computing. In: Advances in Ubiquitous Networking
2. Springer; 2017. p. 205–17.

Banerjee S, Chattopadhyay S. Optimization of three-dimensional turbo code

RI
625 using novel symbiotic organism search algorithm. In: India Conference (IN-
DICON), 2016 IEEE Annual. IEEE; 2016. p. 1–6.

SC
Banerjee S, Chattopadhyay S. Power optimization of three dimensional turbo
code using a novel modified symbiotic organism search (msos) algorithm.

U
Wireless Personal Communications 2017;92(3):941–68.

630
AN
Calheiros RN, Ranjan R, Beloglazov A, De Rose CAF, Buyya R. CloudSim:
a toolkit for modeling and simulation of cloud computing environments and
evaluation of resource provisioning algorithms. Software: Practice and Expe-
M
rience 2011;41(1):23–50.

Casas I, Taheri J, Ranjan R, Wang L, Zomaya AY. Ga-eti: An enhanced

D

635 genetic algorithm for the scheduling of scientific workflows in cloud envi-
ronments. Journal of Computational Science 2016;doi:http://doi.org/10.
TE

1016/j.jocs.2016.08.007.

Chen H, Wang F, Helian N, Akanmu G. User-priority guided min-min schedul-

EP

ing algorithm for load balancing in cloud computing. In: Parallel Comput-
640 ing Technologies (PARCOMPTECH), 2013 National Conference on. Banga-
lore,India: IEEE; 2013. p. 1–8.
C

Cheng MY, Lien LC. Hybrid artificial intelligence–based pba for benchmark
AC

functions and facility layout design optimization. Journal of Computing in

Civil Engineering 2012;26(5):612–24.

645 Cheng MY, Prayogo D. Symbiotic organisms search: a new metaheuristic opti-
mization algorithm. Computers & Structures 2014;139:98–112.

44
 ACCEPTED MANUSCRIPT

Cheng MY, Prayogo D, Tran DH. Optimizing multiple-resources leveling in

multiple projects using discrete symbiotic organisms search. Journal of Com-

PT
puting in Civil Engineering 2015;30(3):04015036.

650 Deb K, Pratap A, Agarwal S, Meyarivan T. A fast and elitist multiobjective

genetic algorithm: Nsga ii. IEEE transactions on evolutionary computation

RI
2002;6(2):182–97.

SC
Delavar AG, Aryan Y. Hsga: a hybrid heuristic algorithm for workflow schedul-
ing in cloud systems. Cluster computing 2014;17(1):129–37.

655 Dib NI. Design of linear antenna arrays with low side lobes level using symbiotic

U
organisms search. Progress In Electromagnetics Research B 2016;68:55–71.

AN
Dosoglu MK, Guvenc U, Duman S, Sonmez Y, Kahraman HT. Symbiotic organ-
isms search optimization algorithm for economic/emission dispatch problem
in power systems. Neural Computing and Applications 2016;:1–17.
M
660 Duman S. Symbiotic organisms search algorithm for optimal power flow problem
based on valve-point effect and prohibited zones. Neural Computing and
D

Applications 2016;:1–15.
TE

Durillo JJ, Nae V, Prodan R. Multi-objective energy-efficient workflow

scheduling using list-based heuristics. Future Generation Computer Systems
2014;36:221–36.
EP

665

Eki R, Vincent FY, Budi S, Redi AP. Symbiotic organism search (sos) for
solving the capacitated vehicle routing problem. World Academy of Science,
C

Engineering and Technology, International Journal of Mechanical, Aerospace,

AC

Industrial, Mechatronic and Manufacturing Engineering 2015;9(5):850–4.

670 Elhabyan R, Shi W, St-Hilaire M. A pareto optimization-based approach to

clustering and routing in wireless sensor networks. Journal of Network and
Computer Applications 2018;114:57–69.

45
 ACCEPTED MANUSCRIPT

Fard HM, Prodan R, Fahringer T. Multi-objective list scheduling of workflow

applications in distributed computing infrastructures. Journal of Parallel and

PT
675 Distributed Computing 2014;74(3):2152–65.

Feitelson DG, Tsafrir D, Krakov D. Experience with using the parallel workloads
archive. Journal of Parallel and Distributed Computing 2014;74(10):2967–82.

RI
Ferdaus MH, Murshed M, Calheiros RN, Buyya R. An algorithm for network

SC
and data-aware placement of multi-tier applications in cloud data centers.
680 Journal of Network and Computer Applications 2017;98:65–83.

Foster I, Zhao Y, Raicu I, Lu S. Cloud computing and grid computing 360-degree

U
compared. In: 2008 Grid Computing Environments Workshop. Austin, Texas:
Ieee; 2008. p. 1–10.
AN
Gandomi AH, Yang XS. Chaotic bat algorithm. Journal of Computational
685 Science 2014;5(2):224–32.
M
Ghazouani S, Slimani Y. A survey on cloud service description. Journal of
Network and Computer Applications 2017;91:61–74.
D

Guvenc U, Duman S, Dosoglu MK, Kahraman HT, Sonmez Y, Yılmaz C. Ap-

TE

plication of symbiotic organisms search algorithm to solve various economic

690 load dispatch problems. In: INnovations in Intelligent SysTems and Applica-
tions (INISTA), 2016 International Symposium on. Sinaia, Romania: IEEE;
EP

2016. p. 1–7.

Guzek M, Bouvry P, Talbi EG. A survey of evolutionary computation for

C

resource management of processing in cloud computing [review article]. IEEE

AC

695 Computational Intelligence Magazine 2015;10(2):53–67.

Hameed A, Khoshkbarforoushha A, Ranjan R, Jayaraman PP, Kolodziej J,

Balaji P, Zeadally S, Malluhi QM, Tziritas N, Vishnu A, et al. A survey
and taxonomy on energy efficient resource allocation techniques for cloud
computing systems. Computing 2014;:1–24.

46
 ACCEPTED MANUSCRIPT

700 Hayyolalam V, Kazem AAP. A systematic literature review on qos-aware service

composition and selection in cloud environment. Journal of Network and

PT
Computer Applications 2018;.

Hu H, Li Z, Hu H, Chen J, Ge J, Li C, Chang V. Multi-objective scheduling

for scientific workflow in multicloud environment. Journal of Network and

RI
705 Computer Applications 2018;114:108–22.

SC
Ishibuchi H, Hitotsuyanagi Y, Tsukamoto N, Nojima Y. Many-objective test
problems to visually examine the behavior of multiobjective evolution in a
decision space. In: International Conference on Parallel Problem Solving

U
from Nature. Kraków, Poland: Springer; 2010. p. 91–100.

710
AN
Kalra M, Singh S. A review of metaheuristic scheduling techniques in cloud
computing. Egyptian Informatics Journal 2015;16(3):275–95.

Kanimozhi G, Rajathy R, Kumar H. Minimizing energy of point charges on a

M
sphere using symbiotic organisms search algorithm. International Journal on
Electrical Engineering and Informatics 2016;8(1):29.
D

715 Kasahara Y, Yonezawa Y. The properties of complex evolution in chaos gen-

TE

eration process. In: Evolutionary Computation, 1996., Proceedings of IEEE

International Conference on. Nagoya University, JAPAN: IEEE; 1996. p. 874–
9.
EP

Kennedy J. Particle swarm optimization. In: Encyclopedia of machine learning.

720 Springer; 2011. p. 760–6.
C

Latiff MSA, Madni SHH, Abdullahi M, et al. Fault tolerance aware scheduling
AC

technique for cloud computing environment using dynamic clustering algo-

rithm. Neural Computing and Applications 2016;:1–15.

Le Hoang S. Optimizing municipal solid waste collection using chaotic particle

725 swarm optimization in gis based environments: a case study at danang city,
vietnam. Expert systems with applications 2014;41(18):8062–74.

47
 ACCEPTED MANUSCRIPT

Li X, Xu J, Yang Y. A chaotic particle swarm optimization-based heuristic for

market-oriented task-level scheduling in cloud workflow systems. Computa-

PT
tional intelligence and neuroscience 2015;2015:81.

730 Li Z, Ge J, Yang H, Huang L, Hu H, Hu H, Luo B. A security and cost aware

scheduling algorithm for heterogeneous tasks of scientific workflow in clouds.

RI
Future Generation Computer Systems 2016;65:140–52.

SC
Liu L, Zhang M, Buyya R, Fan Q. Deadline-constrained coevolutionary genetic
algorithm for scientific workflow scheduling in cloud computing. Concurrency
735 and Computation: Practice and Experience 2016;29(5).

U
Mao Y, Chen X, Li X. Max–min task scheduling algorithm for load balance in

AN
cloud computing. In: Proceedings of International Conference on Computer
Science and Information Technology. Kunming, China: Springer; 2014. p.
457–65.
M
740 Meena J, Kumar M, Vardhan M. Cost effective genetic algorithm for workflow
scheduling in cloud under deadline constraint. IEEE Access 2016;4:5065–82.
D

Midya S, Roy A, Majumder K, Phadikar S. Multi-objective optimization tech-

TE

nique for resource allocation and task scheduling in vehicular cloud architec-
ture: A hybrid adaptive nature inspired approach. Journal of Network and
745 Computer Applications 2018;103:58–84.
EP

Ming G, Li H. An improved algorithm based on max-min for cloud task schedul-

ing. In: Recent Advances in Computer Science and Information Engineering.
C

Springer; 2012. p. 217–23.

AC

Nama S, Saha A, Ghosh S. Improved symbiotic organisms search algorithm

750 for solving unconstrained function optimization. Decision Science Letters
2016;5(3):361–80.

Nanda SJ, Jonwal N. Robust nonlinear channel equalization using wnn trained
by symbiotic organism search algorithm. Applied Soft Computing 2017;.

48
 ACCEPTED MANUSCRIPT

Netjinda N, Sirinaovakul B, Achalakul T. Cost optimal scheduling in iaas for

755 dependent workload with particle swarm optimization. The Journal of Su-

PT
percomputing 2014;68(3):1579–603.

Nirmala SJ, Bhanu SMS. Catfish-pso based scheduling of scientific workflows

in iaas cloud. Computing 2016;98(11):1091–109.

RI
Ostermann S, Iosup A, Yigitbasi N, Prodan R, Fahringer T, Epema D. A per-
formance analysis of ec2 cloud computing services for scientific computing. In:

SC
760

International Conference on Cloud Computing. Munich, Germany: Springer;

2009. p. 115–31.

U
Panda A, Pani S. A symbiotic organisms search algorithm with adaptive penalty
function to solve multi-objective constrained optimization problems. Applied
765 Soft Computing 2016;46:344–60.
AN
Patel G, Mehta R, Bhoi U. Enhanced load balanced min-min algorithm for
M
static meta task scheduling in cloud computing. Procedia Computer Science
2015;57:545–53.
D

Pham D, Ghanbarzadeh A, Koc E, Otri S, Rahim S, Zaidi M. The bees

algorithm–a novel tool for complex optimisation. In: Intelligent Production
TE

770

Machines and Systems-2nd I* PROMS Virtual International Conference (3-14

July 2006). 2011. .
EP

Prayogo D, Cheng MY, Prayogo H. A novel implementation of nature-inspired

optimization for civil engineering: A comparative study of symbiotic organ-
775 isms search. Civil Engineering Dimension 2017;19(1):36–43.
C

Qin AK, Huang VL, Suganthan PN. Differential evolution algorithm with strat-
AC

egy adaptation for global numerical optimization. IEEE transactions on Evo-

lutionary Computation 2009;13(2):398–417.

Rajagopalan A, Sengoden V, Govindasamy R. Solving economic load dispatch

780 problems using chaotic self-adaptive differential harmony search algorithm.
International Transactions on Electrical Energy Systems 2015;25(5):845–58.

49
 ACCEPTED MANUSCRIPT

Rodriguez MA, Buyya R. Deadline based resource provisioningand scheduling

algorithm for scientific workflows on clouds. IEEE Transactions on Cloud

PT
Computing 2014;2(2):222–35.

785 Secui DC. A modified symbiotic organisms search algorithm for large scale
economic dispatch problem with valve-point effects. Energy 2016;113:366–84.

RI
Shayeghi H, Ghasemi A. A modified artificial bee colony based on chaos theory

SC
for solving non-convex emission/economic dispatch. Energy Conversion and
Management 2014;79:344–54.

790 Shen Y, Bao Z, Qin X, Shen J. Adaptive task scheduling strategy in cloud:

U
when energy consumption meets performance guarantee. World Wide Web
2016;:1–19.
AN
Singh S, Chana I. A survey on resource scheduling in cloud computing: Issues
and challenges. Journal of Grid Computing 2016;14(2):217–64.
M
795 Sonmez Y, Kahraman HT, Dosoglu MK, Guvenc U, Duman S. Symbiotic organ-
isms search algorithm for dynamic economic dispatch with valve-point effects.
D

Journal of Experimental & Theoretical Artificial Intelligence 2016;:1–21.

TE

Suresh S, Lal S. Multilevel thresholding based on chaotic darwinian particle

swarm optimization for segmentation of satellite images. Applied Soft Com-
puting 2017;55:503–22.
EP

800

Tao F, Feng Y, Zhang L, Liao T. Clps-ga: A case library and pareto solution-
based hybrid genetic algorithm for energy-aware cloud service scheduling.
C

Applied Soft Computing 2014;19:264–79.

AC

Tawfeek MA, El-Sisi A, Keshk A, Torkey FA. Cloud task scheduling based on
805 ant colony optimization. Int Arab J Inf Technol 2015;12(2):129–37.

Tejani GG, Savsani VJ, Patel VK. Adaptive symbiotic organisms search (sos)
algorithm for structural design optimization. Journal of Computational De-
sign and Engineering 2016;3(3):226–49.

50
 ACCEPTED MANUSCRIPT

Teng S, Lee LH, Chew EP. Multi-objective ordinal optimization for simulation
810 optimization problems. Automatica 2007;43(11):1884–95.

PT
Thakur A, Goraya MS. A taxonomic survey on load balancing in cloud. Journal
of Network and Computer Applications 2017;.

RI
Tiwari A, Pandit M. Bid based economic load dispatch using symbiotic organ-
isms search algorithm. In: Engineering and Technology (ICETECH), 2016

SC
815 IEEE International Conference on. Tamil Nadu, India: IEEE; 2016. p. 1073–8.

Tran DH, Cheng MY, Prayogo D. A novel multiple objective symbiotic or-
ganisms search (mosos) for time–cost–labor utilization tradeoff problem.

U
Knowledge-Based Systems 2016;94:132–45.

820
AN
Tsai CW, Rodrigues JJ. Metaheuristic scheduling for cloud: A survey. IEEE
Systems Journal 2014;8(1):279–91.
M
Vakili A, Navimipour NJ. Comprehensive and systematic review of the service
composition mechanisms in the cloud environments. Journal of Network and
Computer Applications 2017;81:24–36.
D

Verma A, Kaushal S. A hybrid multi-objective particle swarm optimization for

TE

825 scientific workflow scheduling. Parallel Computing 2017;62:1–19.

Vincent FY, Redi AP, Yang CL, Ruskartina E, Santosa B. Symbiotic organism
EP

search and two solution representations for solving the capacitated vehicle
routing problem. Applied Soft Computing 2017;52:657–72.

Wang B, Song Y, Sun Y, Liu J. Managing deadline-constrained bag-of-tasks

C

830 jobs on hybrid clouds with closest deadline first scheduling. KSII Transactions
AC

on Internet & Information Systems 2016;10(7).

Wu F, Wu Q, Tan Y. Workflow scheduling in cloud: a survey. The Journal of

Supercomputing 2015;71(9):3373–418.

51
 ACCEPTED MANUSCRIPT

Wu H, Zhou Y, Luo Q, Basset MA. Training feedforward neural networks

835 using symbiotic organisms search algorithm. Computational Intelligence and

PT
Neuroscience 2016;2016.

Wu Q, Law R, Wu E, Lin J. A hybrid-forecasting model reducing gaussian noise

based on the gaussian support vector regression machine and chaotic particle

RI
swarm optimization. Information Sciences 2013;238:96–110.

SC
840 Xu Y, Li K, Hu J, Li K. A genetic algorithm for task scheduling on heteroge-
neous computing systems using multiple priority queues. Information Sciences
2014;270:255–87.

U
Xue B, Zhang M, Browne W, Yao X. A survey on evolutionary computation

845
AN
approaches to feature selection 2016;20(4):606 –26.

Yao G, Ding Y, Jin Y, Hao K. Endocrine-based coevolutionary multi-swarm

for multi-objective workflow scheduling in a cloud system. Soft Computing
M
2016;:1–14.

Yassa S, Chelouah R, Kadima H, Granado B. Multi-objective approach for

D

energy-aware workflow scheduling in cloud computing environments. The

TE

850 Scientific World Journal 2013;2013.

Zamani MKM, Musirin I, Suliman SI. Symbiotic organisms search technique

for svc installation in voltage control. Indonesian Journal of Electrical Engi-
EP

neering and Computer Science 2017;6(2):318–29.

Zeng L, Veeravalli B, Zomaya AY. An integrated task computation and data

C

855 management scheduling strategy for workflow applications in cloud environ-

AC

ments. Journal of Network and Computer Applications 2015;50:39–48.

Zhan ZH, Li J, Cao J, Zhang J, Chung HSH, Shi YH. Multiple populations
for multiple objectives: A coevolutionary technique for solving multiobjective
optimization problems. IEEE Transactions on Cybernetics 2013;43(2):445–63.

52
 ACCEPTED MANUSCRIPT

860 Zhan ZH, Liu XF, Gong YJ, Zhang J, Chung HSH, Li Y. Cloud computing
resource scheduling and a survey of its evolutionary approaches. ACM Com-

PT
puting Surveys (CSUR) 2015;47(4):63.

Zhang F, Cao J, Li K, Khan SU, Hwang K. Multi-objective scheduling of many

tasks in cloud platforms. Future Generation Computer Systems 2014;37:309–

RI
865 20.

SC
Zhang L, Li K, Li C, Li K. Bi-objective workflow scheduling of the energy
consumption and reliability in heterogeneous computing systems. Information
Sciences 2017;379:241–56.

U
Zheng W, Sakellariou R. Budget-deadline constrained workflow planning for
870
AN
admission control. Journal of grid computing 2013;11(4):633–51.

Zhong Z, Chen K, Zhai X, Zhou S. Virtual machine-based task scheduling algo-

rithm in a cloud computing environment. Tsinghua Science and Technology
M
2016;21(6):660–7.

Zhu Z, Zhang G, Li M, Liu X. Evolutionary multi-objective workflow schedul-

D

875 ing in cloud. IEEE Transactions on Parallel and Distributed Systems

TE

2016;27(5):1344–57.

Zitzler E, Thiele L, Laumanns M, Fonseca CM, Da Fonseca VG. Performance

assessment of multiobjective optimizers: An analysis and review. IEEE Trans-
EP

actions on evolutionary computation 2003;7(2):117–32.

880 Zuo L, Shu L, Dong S, Zhu C, Hara T. A multi-objective optimization scheduling

C

method based on the ant colony algorithm in cloud computing. IEEE Access
AC

2015;3:2687–99.

Zuo X, Zhang G, Tan W. Self-adaptive learning pso-based deadline constrained

task scheduling for hybrid iaas cloud. IEEE Transactions on Automation
885 Science and Engineering 2014;11(2):564–73.

53
 ACCEPTED MANUSCRIPT

Abdullahi Mohammed received his PhD in Computer Science

from Universiti Teknologi Malaysia. He received his M.Sc. degree
in Computer Science from Ahmadu Bello University, Zaria Nigeria
and B.Tech degree in Mathematics with Computer from Federal
University of Technology, Minna Nigeria. He is a lecturer in
Ahmadu Bello University, Zaria Nigeria. His research interests
include algorithm design for distributed systems, big data analytics,

PT
machine learning, and and large scale optimization using nature
inspired algorithms. He is a member of IEEE and ACM.

RI
Md Asri Bin Ngadi received PhD in Computer Science from Aston
University, Birmingham UK and B.Sc in Computer Science from
Universiti Teknologi Malaysia. His research interests is related to

SC
Wireless Computer, Cloud Computing, and Network Security.
Currently he is appointed as Associate Professor at Universiti
Teknologi Malaysia and a member of IEEE and ACM.

U
Salihu Idi Dishing is a PhD research student at the Faculty of
AN
Computing, Universiti Teknologi Malaysia (UTM), in Malaysia. He
holds an MSc degree from the Robert Gordon University (RGU),
Aberdeen, United Kingdom, and a BSc in Mathematics with Computer
science from Ahmadu Bello University (ABU), Zaria, Nigeria, where
M

he is a faculty member with the department of Computer Science, in

Faculty of Physical Sciences. His research interests include:
Distributed (Cloud, Grid, and Fog) systems modeling and simulation,
D

Nature-inspired algorithms, and Machine Learning.

TE

Shafi’i Muhammad ABDULHAMID received his PhD in

Computer Science from Universiti Teknologi Malaysia. He received
his M.Sc. degree in Computer Science from Bayero University Kano,
EP

Nigeria and B.Tech. degree in Mathematics/Computer Science from

the Federal University of Technology Minna, Nigeria. His current
research interest is scheduling in Grid and Cloud Computing. He is a
member of IEEE and a member of Nigerian Computer Society
C

(NCS).
AC

Barroon Isma'eel Ahmad obtained his BSc. from Usmanu

Danfodiyo University Sokoto and MSc. From Ahmadu Bello
University Zaria both in Computer Science in Nigeria, his PhD from
International Islamic University Malaysia in Information
Technology. He is currently working with Department of Computer
Science at Ahmadu Bello University Zaria, Nigeria. His research
interests are in Health Informatics, Mobile and Pervasive
Computing, Embedded Systems, and eLearning. He is a member of
IEEE, ACM, AIS, IACSIT, NCS, and AITP.