A Hyper-Heuristic Scheduling Algorithm For Cloud

This article has been accepted for publication in a future issue of this journal, but has not been
fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TCC.2014.2315797, IEEE Transactions on Cloud Computing
IEEE TRANSACTIONS ON CLOUD COMPUTING, VOL. X, NO. X,JANUARY XXXX
A Hyper-Heuristic Scheduling Algorithm for Cloud

Chun-Wei Tsai, Wei-Cheng Huang, Meng-Hsiu Chiang, Ming-Chao Chiang, and Chu-Sing Yang
AbstractRule-based scheduling algorithms have been widely

used on many cloud computing systems because they are simple
and easy to implement. However, there is plenty of room to
improve the performance of these algorithms, especially by
using heuristic scheduling. As such, this paper presents a novel
heuristic scheduling algorithm, called hyper-heuristic scheduling
algorithm (HHSA), to find better scheduling solutions for cloud
computing systems. The diversity detection and improvement
detection operators are employed by the proposed algorithm to
dynamically determine which low-level heuristic is to be used in
finding better candidate solutions. To evaluate the performance of
the proposed method, this study compares the proposed method
with several state-of-the-art scheduling algorithms, by having all
of them implemented on CloudSim (a simulator) and Hadoop
(a real system). The results show that HHSA can significantly
reduce the makespan of task scheduling compared with the other
scheduling algorithms evaluated in this paper, on both CloudSim
and Hadoop.
Index TermsCloud computing, evolutionary algorithm,
scheduling, and Hadoop.
I. I NTRODUCTION
NTIL now we are still actively looking for possible

solutions to enhance the performance of information
systems for computation, analysis, and storage. Since distributed and parallel computing was widely used to enhance
the performance of a variety of computer systems, several
strategies have been presented for different approaches and
congenital restrictions in different eras. No matter which
consideration it is for, how to efficiently manage computer
resources is definitively a vital research issue. Among them,
scheduling [1], [2], [3] is essential in the success of enhancing
the performance of a computer system.
With the advance of computer and internet technologies,
paradigms of cloud computing [4], [5], [6], [7] have been
successfully used on several information systems in recent
years. Although cloud computing systems nowadays provide
a better way to carry out the submitted tasks in terms
of responsiveness, scalability, and flexibility, most job and
task scheduling problems on cloud computing systems are
still either NP-hard or NP-complete [8]. According to our
observations, most rule-based scheduling algorithms (e.g.,
C. W. Tsai is with the Department of Applied Informatics and Multimedia,
Chia Nan University of Pharmacy & Science, Tainan 71710, Taiwan, e-mail:
[email protected].
W. C. Huang is with the Department of Computer Science and Engineering, National Sun Yat-sen University, Kaohsiung 80424, Taiwan, e-mail:
[email protected].
M. H. Chiang is with the Department of Computer Science and Engineering, National Sun Yat-sen University, Kaohsiung 80424, Taiwan, e-mail:
[email protected].
M. C. Chiang is with the Department of Computer Science and Engineering,
National Sun Yat-sen University, Kaohsiung 80424, Taiwan, e-mail: [email protected].
C. S. Yang is with the Department of Electrical Engineering, National
Cheng Kung University, Tainan 70101, Taiwan, e-mail:csyang.ee.ncku.edu.tw.
exhaustive and deterministic algorithms) are widely used on

todays cloud computing systems because they are simple and
easy to implement. Unfortunately, these rule-based scheduling algorithms are inappropriate for large-scale or complex
scheduling problems because the results of these scheduling
strategies are usually far from optimal. This implies that there
is plenty of room to improve the scheduling algorithms for
cloud computing systems.
A promising research direction, which applies modern
heuristics [9] to scheduling on cloud computing systems,
has attracted several researchers from different research domains. The extended version of heuristics combines two or
more heuristics into a single heuristic algorithm to leverage
their strengths for scheduling, called hybrid heuristic-based
scheduling algorithm. Different from most hybrid heuristicbased scheduling algorithms that employed only one or two
heuristics, the proposed algorithm is a hyper-heuristic-based
algorithm that integrates several heuristic algorithms. The
basic idea of the proposed algorithm is to leverage the
strengths of heuristic algorithms, such as simulated annealing
[10], genetic algorithm [11], particle swarm optimization [12],
and ant colony optimization [13], by integrating them into a
single algorithm. The main contributions of the paper can be
summarized as follows:
1) A novel high-performance hyper-heuristic algorithm is
proposed for scheduling on cloud computing systems to
reduce the makespan.
2) Two detection operatorsone for diversity detection and
one for improvement detectionare proposed for the
proposed algorithm to determine the timing to employ
the low-level heuristic algorithm.
3) The proposed algorithm can be applied to both sequencedependent (see also Section V-A1) and sequenceindependent (see also Section V-B1) scheduling problems.
4) To evaluate its performance on both a cloud simulator
and a real cloud system, the proposed algorithm and
several other scheduling algorithms (for the purpose of
comparison) are implemented on both the CloudSim [14]
and Hadoop.
5) A detailed analysis of the parameter settings is also given
to show the impact they may have on the performance of
the proposed algorithm.
The remainder of the paper is organized as follows. Section II begins with a brief introduction to the traditional
scheduling, followed by the scheduling on cloud computing.
In addition to the scheduling problem and its relevant technologies, Section III provides a brief discussion on heuristic
algorithms and compares the differences between hybridheuristic and hyper-heuristic algorithms. Section IV describes
2168-7161 (c) 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
in detail the proposed algorithm and gives a simple example.

The simulation results, on both CloudSim and Hadoop, are
discussed in Section V. Conclusions and future trends are
drawn in Section VI.
II. S CHEDULING
A. Scheduling problems
1) Traditional Scheduling: The scheduling problem [15],
[16] can be defined as follows: Find an optimal solution to
schedule a given set of tasks T = {T1 , T2 , . . . , Tn } to a
given set of machines M = {M1 , M2 , . . . , Mm } subject to a
predefined set of constraints and measurements. For instance,
one of the widely used measurements for the scheduling
problem is the so-called makespan Cmax (s), which is defined
as the completion time of the last task. Mathematically, the
scheduling problem is to
minimize f (s) = Cmax (s),
(1)
where s is a candidate solution, and, letting Cj denote the

completion time of job j, Cmax (s) = maxj Cj is the completion time of the last job. Moreover, with advance of computer
systems, the scheduling problem has evolved from single
processor, single tasking, to multiprocessor, multitasking, then
to large scale distributed system nowadays. To measure the
solutions for different kinds of scheduling problems, makespan
is no longer the only choice. Instead, a number of measurements have been developed. Examples are maximum lateness,
maximum tardiness, and maximum flowtime [1].
Several traditional scheduling algorithms have been presented, such as earliest due date first (EDD) [17], critical path
method (CPM) [18], project evaluation and review technique
(PRET) [19], dynamic programming [20], and branch-andbound [21]. These methods are generally much easier to implement than metaheuristic and hybrid scheduling algorithms
because they are all designed based on one or a few particular
rules (ways) to manage and arrange the tasks, but metaheuristic
and hybrid scheduling algorithms are not. Some of them
even guarantee to find the optimal solution for the scheduling
problem in question, but they are all suffered from the fact
of not being able to find the optimal solution in a reasonable
time, especially when the problem becomes either too complex
or too large.
In fact, it is the other two important disadvantages that
take the researchers into consideration of modifying the search
strategy of these traditional scheduling algorithms. First, most
full search algorithms, such as the well-known branch-andbound and dynamic programming, are normally time consuming because the number of checks they have to perform is
very large. Second, in the case of the deterministic algorithms
(such as EDD), although they are very fast and easy to implement, they are easily falling into local optima. The classical
exhaustive algorithms are generally good enough to provide
the optimal solution quickly for small scheduling problems
[22], but not large scheduling problems [23]. In this paper, by
the small scheduling problems, we mean problems for which
all the solutions can be checked in a reasonable time by using
classical exhaustive algorithms running on modern computer
systems. In contrast, by the large scheduling problems, we

mean problems for which not all the solutions can be checked
in a reasonable time by using the same algorithms running
on the same computer systems. These observations make it
easy to understand that exhaustive algorithms will take a
prohibitive amount of time to check all the candidate solutions
for large scheduling problems because the number of candidate
solutions is simply way too large to be checked in a reasonable
time. As a result, researchers have paid their attention to the
development of scheduling algorithms that are efficient and
effective, such as heuristics.
2) Scheduling on Cloud: With computing systems being
shifted to cloud-based systems progressively, one of the main
characteristics is that it works on a pay-as-you-use basis [24].
Several studies attempted to define the scheduling problem on
cloud systems as the workflow problem, which can be further
classified into two levels: service-level (platform layer and
static scheduling) and task-level (unified resource layer and
dynamic scheduling) [25]. Different from grid computing,
the user can install their programs on the virtual machines
(VMs) and determine how to execute their programs on the
cloud computing system. For these reasons, although both
grid computing and cloud computing are heterogeneous, the
key issues they face are very different. A good example is
the cost and latency of data transfer on these environments.
That is why some studies added more considerations to their
definitions of scheduling on cloud. For instance, a couple
of studies [26], [27] used directed acyclic graph (DAG) to
define the scheduling problem on cloud. The basic idea is to
use the vertices of a DAG to represent a set of tasks and
the edges between the vertices to represent the dependencies
between the tasks. Then, the scheduling problem on cloud can
be formulated for the solution s as follows [27]:
minimize f (s) = Cmax (s) +
n X
m
X
i=1 j=1
subject to Cmax (s) U (s),

C(s) B(s),
Cij ,
(2)
where f (s) is the objective function; Cmax (s) the completion

time of the last job (also called makespan); n the number of
tasks; m the number of machines; Cij the cost of processing
the ith task on the
Pmjth machine; U (s) the number of overdue
tasks; C(s) =
j=1 Cij the total cost of s; and B(s) the
restriction on the budget for the tasks of s. Moreover, since the
study described in [27] also takes into account the maximum
expenditure for the provisioning of services or applications on
cloud, such as QoS, the budget is also considered as part of
the scheduling problem for cloud when necessary.
Recently, Hadoop [28] was widely used in deploying cloud
computing system. In addition to being used by several international companies (e.g., IBM, Google, AOL, eBay, Yahoo,
and Amazon) [29], the features provided by the open-source
environment and MapReduce framework have made it a very
popular cloud computing platform [30]. Its scalability for
large-scale data and systems has also been found in recent
researches, such as big data approach [31] and smart grid
[32]. The basic idea is so that the users of such a system
can submit jobs to the Hadoop cluster, which will then split
the jobs into tasks of a given size and provide a number of
slots to run tasks in parallel. The processing speed of these
slots depends on the types of jobs submitted to the cluster.
The definitions of parallel machines for scheduling can be
divided into three kinds: (1) identical machines in parallel, (2)
machines in parallel with different speeds, and (3) unrelated
machines in parallel [1]. Since a Hadoop cluster generally
consist of machines with different speeds, its task scheduling
problem can be considered as the second kind of parallel
machines for scheduling.
B. Scheduling Methods for Cloud
To develop more efficient scheduling algorithms for cloud,
some recent reviews were given in [33], [34], which encompass non-metaheuristic and metaheuristic scheduling algorithms, such as earliest-finish-time-based algorithm and
genetic algorithm. Also given in [33] was a comparison
sheet to classify the scheduling algorithms discussed there by
methods, parameters, and factors of scheduling. In addition,
the method presented in [35] uses feedback information to
estimate the earliest finish time to dynamically adjust the
resource allocation.
Among researches on non-metaheuristic scheduling algorithms, the earliest start time, latest start time, and makespan
[36], [37] are usually used to evaluate the scheduling results;
however, to precisely describe cloud environment, additional
measurements (objective functions) are also used in describing
and measuring the scheduling on cloud. For instance, the cost
of workflow and the deadline defined by a user were used in
[38] while the price-based model was used in [39]. Besides
the additional measurements, another study [36] combined the
static and dynamic scheduling algorithms into a cloud system
to adjust its resource allocation strategies on the fly. The main
advantages of this kind of algorithms are: (1) they are simple
and easy to implement; (2) some rule-based deterministic
algorithms can find acceptable solutions quickly; and (3) most
of them are compatible to each other; thus, some studies have
tried to integrate two or more non-metaheuristic algorithms
to solve the scheduling problem. Theoretically, one of the
major disadvantages of these algorithms is in that the results
obtained by these algorithms may be far from optimal or even
acceptable. In practice, our observation shows that because
they are simple and easy to implement on a cloud system,
several cloud systems use traditional scheduling algorithms
to manage the computing resources, such as Hadoop, which
uses first in first out (FIFO) [40] as the default scheduling
algorithm.
Although the FIFO scheduler of Hadoop can achieve a
high utilization as far as a cluster is concerned, it provides
no fairness at all. Even worse, it sometimes leads to a long
response time, especially when the number of jobs to be run
is beyond the available computing resources. For a concrete
example, consider the following scenario: A very long job
arrives the job queue followed by a short job with the same
priority. The FIFO scheduler of Hadoop will schedule the first
job and allocate all the resources it needs. This implies that
the second job must wait until the first job is completed. The
result is obviously a very long response time for the second
job. For a cluster that has enough capacity to run all the jobs
at the same time, most schedulers can satisfy the demand of
the input jobs and tasks. It means that each job can be run and
completed before the deadline. Assume a high map-capacity
Hadoop cluster has 200 maps. Most of the schedulers can
finish the jobs when there are 10 input jobs each of which
use 20 maps, i.e., when there are total 10 20 = 200 tasks
to be assigned to the 200 map slots. When the input jobs and
tasks are too large to be run by the cluster (system) at the same
time, the overall computation time will differ because different
scheduling strategies may decide to schedule the jobs to be run
and the resources to be allocated differently. Moreover, most,
if not all, of the schedulers shipped with Hadoop today do not
take into consideration the so-called makespanthe objective
HHSA described herein is trying to improve.
To solve this problem, Facebook developed Hadoop Fair
Scheduler (HFS) and improved it with Delay Scheduling [41]
to get better data locality and throughput. HFS calculates each
tasks fair share, sorts them in increasing order, and then
picks the one with least fair share when there are empty slots.
In the scenario above, HFS will schedule the shorter job first
when it receives low fair share, thus finishing the shorter job
before the longer one completes, by sacrificing only a little
delay for a dramatically faster response time of the second job.
On the other hand, Yahoo developed Capacity Scheduler [42]
to deal with the same scheduling problem. By guaranteeing the
capacity of the queues for different jobs, Capacity Scheduler
can also achieve fairness, but it requires a more sophisticated
control over the cluster. In summary, these two schedulers
are not focused on pursuing the optimal solution; rather,
they try to enhance the response time of each job. As such,
this paper is aimed at demonstrating that by estimating and
predicting the processing speed and completion time of each
machine and each job, HHSA can generate feasible solutions
that optimize the makespan not taken into consideration by
the other schedulers. In addition, our experiments show that
HHSA can eventually provide better solutions in terms of
makespan.
In addition to non-metaheuristic scheduling algorithms, an
alternative is the heuristic scheduling algorithms [9] that can
be used to provide better scheduling plans than do the rulebased scheduling algorithms for cloud computing. As shown
in Fig. 1(a), the basic idea of heuristics is to use three
key operatorstransition, evaluation, and determination
to search for the possible solutions on the convergence
process. Note that in Fig. 1(a), t denotes the iteration number;
tmax the maximum number of iterations or the stop criteria.
More precisely, the transition operator creates the solution
s, by using methods which could be either perturbative or
constructive or both [43]; the evaluation operator measures the
fitness of s by using a predefined measurement; and then the
determination operator determines the next search directions
based on the s from the transition operator and the evaluation
operator. Because heuristics use tactical guess to find the
possible solutions, it has a much better chance to find a better
result than do most rule-based deterministic algorithms.
Input
H
H1
LLH Selection
Transition
H1
H2
Input
TED
Input
Evaluation
t=t+1
Determination
t = tmax ?
No
Yes
Fig. 1.
TED
t=t+1
Hn
t = tmax ?
Yes
Output
Output
(a)
(b)
No
No
Hn
Hi
t=t+1
Accepted?
Yes
t = tmax ?
No
Yes
Output
(c)
Outline of (a) heuristics, (b) hybrid-heuristics, and (c) hyper-heuristics.
III. I NTEGRATION OF H EURISTIC A LGORITHMS

As we mentioned in Section II-B, heuristic algorithms are
able to provide better results than rule-based algorithms for
complex scheduling problems on cloud computing systems.
Moreover, as far as this paper is concerned, by the complex
scheduling problems, we mean scheduling problems that are
NP-complete [8], thus having a huge solution space for which
all the solutions cannot be checked by any known exhaustive
algorithm in a reasonable time, not only in a traditional
information system (non-cloud virtualized cluster) but also in
a cloud system. Each heuristic algorithm, however, has its pros
and cons as far as scheduling is concerned. For example, ant
colony optimization may provide a better scheduling solution
than other scheduling algorithms, but it takes much more
computation time than the others. An alternative to create a
scheduling algorithm is by integrating two or more heuristic
algorithms or combining heuristic algorithms with rule-based
algorithms, Generally speaking, an integration that invokes
two or more heuristic algorithms at each iteration reflects the
complementary advantages of these algorithms to find a better
result than a single heuristic algorithm does. As shown in
Fig. 1(b), the basic idea of hybrid-heuristic algorithm is to
combine heuristic algorithms to perform the transition (T),
evaluation (E), and determination (D) at each iteration, where
Hi denotes one of the heuristic algorithms. This kind of
integration may compensate for the intrinsic weak points of
specific heuristic algorithms. For instance, the search diversity
of tabu search may be increased by integrating it with genetic
algorithm or other heuristics. However, a critical problem is
that although the hybrid-heuristic may have a higher chance
to find a better result than a single heuristic does, it generally takes a longer computation time than heuristics at each
iteration of the convergence process.
Recently, another way to combine two or more heuristic algorithms, called hyper-heuristic algorithm [44], has attracted many researchers from different domains. As shown
in Fig. 1(c), one of the heuristic algorithms in the pool
of candidate heuristics {H1 , H2 , . . . , Hn } will be selected
by the low-level heuristic (LLH) selection operator as the
heuristic algorithm to be performed, which is referred to as
Hi in Fig. 1(c). For this kind of heuristics, low-level heuristic
selection and acceptance are the two essential operators to
determine which heuristic algorithm is to be selected and
whether the selected heuristic algorithm will be continuously
used or not. For this reason, several LLH selection operators

have been presented in different studies [45] to choose the
Hi . Different from the LLH selection operator that plays
the role of determining the heuristic algorithm, the acceptance
operator plays the role of determining the timing to select a
new heuristic algorithm, such as when the Hi gets stuck with
a solution for a certain number of iterations or when the Hi
falls into local optimum.
In spite of all these extensions, there still lacks a simple and
efficient design for scheduling on cloud today. For example, although most rule-based scheduling algorithms are very simple,
they are unable to find the optimal or approximate solution in
a reasonable time because most scheduling problems are NPcomplete. Now, let us look back at the heuristic algorithms.
Like the hybrid-heuristic, hyper-heuristic also attempts to use
two or more heuristics on the convergence process. Unlike
the hybrid-heuristic, the basic idea of hyper-heuristic is to
use one and only one heuristic algorithm at each iteration.
With these characteristics, hyper-heuristic algorithms can then
maintain a high search diversity to increase the chance of
finding better solutions at later iterations while not increasing
the computation time.
The time complexities of heuristic, hybrid-heuristic, and
hyper-heuristic are definitely an important issue, especially
a comparison between the time complexities of these approaches. As described in [46], the expected running time of
heuristic is O(M log M ) where M indicates the number of
subsolutions of each solution of the problem. Tseng and Yang
in a later study [47] pointed out that the time complexity
of heuristic is O(`N M 2 ) where ` indicates the number of
iterations the algorithm performs; N the population size; and
M the number of subsolutions. In [48], the time complexity of
hybrid-heuristic algorithm for scheduling was given. For each
iteration, the computation of makespan takes O(mn), simulated annealing takes O(mn2 ), and NEH (Nawaz, Enscore,
and Ham) takes O(mn3 ), where m denotes the number of
machines; and n the number of jobs. Thus, the time complexity
of hybrid-heuristic algorithm is O(`mn3 ). This is due to
the fact that the NEH operator takes O(mn3 ). Putting these
observations together, the time complexity of hybrid-heuristic
algorithm can be obtained as O(`M N 2 ), assuming that (1)
m n of the hybrid-heuristic algorithm equals M of the
heuristic algorithm, which is usually true for the scheduling
2
3
4
5
6
7
8
9
10
11
12
13
14
Set up the parameters.

Input the scheduling problem.
Initialize the population of solutions Z = {z1 , z2 , ..., zN }.
Randomly select a heuristic algorithm Hi from the candidate pool H.
While the termination criterion is not met
Update the population of solutions Z by using the selected algorithm Hi .
F1 = Improvement Detection(Z).
F2 = Diversity Detection(Z).
If (Hi , F1 , F2 )
Randomly select a new Hi .
Z = Perturb(Z).
End.
End.
Output the best so far solution as the final solution.
23
SA
22
21
best so far makespan
20
19
18
17
16
15
0
200
400
600
23
Fig. 2.
800
1000
Iteration
(a) Best so far makespan of SA.
Hyper-heuristic scheduling algorithm (HHSA).
GA
22
problems,1 and (2) the time complexity of all the other

operators of heuristic algorithms are bounded from above by
N M 2 . Similar to the hybrid-heuristic algorithm, the hyperheuristic algorithm also uses two or more heuristic algorithms,
but the time complexity of each operator is usually bounded
from above by O(M N 2 ). As a result, the time complexity
of the hyper-heuristic algorithm is also O(`M N 2 ) if all the
operators (or heuristic algorithms) used are bounded from
above by M N 2 .
IV. T HE P ROPOSED A LGORITHM
A. Hyper-Heuristic Scheduling Algorithm
The basic idea of the proposed algorithm is to use the
diversity detection and improvement detection operators to
balance the intensification and diversification in the search
of the solutions during the convergence process. The proposed algorithm, called hyper-heuristic scheduling algorithm
(HHSA), is as shown in Fig. 2. Line 1 sets up the parameters
max and ni , where max denotes the maximum number
of iterations the selected low-level heuristic algorithm is to
be run; ni the number of iterations the solutions of the
selected low-level heuristic algorithm are not improved. Line 2
reads in the tasks and jobs to be scheduled, i.e., the problem
in question. Line 3 initializes the population of solutions
Z = {z1 , z2 , ..., zN }, where N is the population size. On
line 4, a heuristic algorithm Hi is randomly selected from
the candidate pool H = {H1 , H2 , ..., Hn }.
As far as the proposed algorithm described herein is
concerned, the low-level heuristic candidate pool consists
of simulated annealing [10], genetic algorithm [11], particle
swarm optimization [12], and ant colony optimization [13].
The selected hyper-heuristic algorithm (LLH) Hi will then be
performed repeatedly until the termination criterion is met,
as shown in lines 513. More precisely, the selected LLH
Hi will evolve the solution Z for max iterations by using
the determine function (Hi , F1 , F2 ), as defined in Eq. (3),
to balance the intensification and diversification of the search
directions, which in turn rely on the information provided by
the improvement detection operator denoted F1 (as shown in
line 8) and by the diversity detection operator denoted F2 (as
1 This
means that the solution is encoded as a matrix of size m by n, i.e., a

matrix with m rows and n columns, where each row corresponds to a machine;
each column corresponds to a job. Therefore, the number of subsolutions of
each solution is m n = M .
21
20
19
18
17
16
15
0
200
400
600
800
1000
Iteration
(b) Best so far makespan of GA.

Fig. 3. Results of applying heuristic algorithms to the j30 dataset for the
workflow problem on CloudSim.
shown in line 7) to decide whether to select a new LLH or

not. For SSBHA (single-solution-based heuristic algorithms),
only F1 is used whereas for PBHA (population-based heuristic
algorithms), both F1 and F2 are used.
8
>
<false
(H, F1 , F2 ) = false
>
:true
if H S and F2 = true,
if H P and F1 = true and F2 = true,
otherwise.
(3)
where S denotes the set of SSBHAs; P the set of PBHAs.

When the determine function (Hi , F1 , F2 ) returns a true,
the proposed algorithm will randomly select a new heuristic
algorithm Hi and then return the solution Z to the perturbation
operator to fine-tune the result, as shown in lines 911. Note
that the single-solution-based heuristic algorithms employ one
and only one search direction at each iteration on the convergence process while the population-based heuristic algorithms
employ multiple search directions at each iteration.
1) The Improvement Detection Operator: A simple random
method is used to select the low-level heuristic Hi from the
candidate pool H. According to our observation, the best so
far makespan (BSFMK) for both SA and GA could continue
to improve the results at early iterations (e.g., less than 200
iterations), but it is difficult to improve the results at later
iterations (e.g., after 800 iterations), especially when the search
directions converge to a small number of directions, as shown
in Fig. 3.
From these observations, a simple approach to detecting
when to select a new LLH is as given below.
(
F1 =
false
true
BSFMK is not improved after tni iterations,

otherwise.
(4)
This approach only checks to see whether the solutions found

by Hi , i.e., BSFMK, are being improved or not. If the selected
Hi cannot improve the BSFMK after a row of ni iterations,
the improvement detection operator will return a false value
to the high level hyper center to indicate that it should pick

up a new LLH. Note that the improvement detection operator
will return a false value in three cases: the maximum number
of iterations max is reached, the number of iterations ni
the solutions are not improved is reached, and when the stop
condition is reached.
2) The Diversity Detection Operator: In addition to the improvement detection operator, the diversity detection operator
is used by HHSA to decide when to change the low-level
heuristic algorithm Hi . Here is how it works. The diversity
of the initial solution D(Z0 ) will be used as a threshold ,
i.e., = D(Z0 ). The diversity of the current solution D(Z)
is computed as the average of the task distances between
individual solutions which are defined as follows: If a task
in two different individuals is assigned to the same VM, the
task distance is 0; otherwise, the task distance is 1.
If the diversity of the current solution D(Z) is less than the
threshold (the diversity of the initial solution), this operator
will return false, and HHSA will then randomly select a new
LLH, as given below.
Low level heuristics

SA
GA
PSO
ACO
Hi
N
population-based?
Diversity detection
Improvement detection
Perturbation
Fig. 4.
Example showing how HHSA works.

Data Center
Host
Host
VM
VM
VM
VM
VM
VM
VM
VM
VM
VM
(a)
task
virtual task
(
true D(Z) > ,
F2 =
false otherwise.
1
(5)
s
Note that is defined as 3 where and denotes,

respectively, the average distance and the standard deviation
of the distances of the initial solution.
3) The Perturbation Operator: In addition to the perturbation operators of Hi itself, the proposed algorithm will perturb
the solutions obtained by Hi before they are passed on to the
newly selected LLH. This means that the candidate solutions
created by the low-level heuristic Hi can be perturbed by
the simulated annealing mechanism we presented herein, by
assigning a different temperature to each individual to balance
the intensification and diversification of the search. More
precisely, the perturbation temperature Tk is used to determine
the probability of randomly changing the subsolutions, such as
changing the first subsolution of the solution 0001 from 0 to
1, thus obtaining the solution 1001. This mechanism implies
that the higher the temperature, the higher the opportunity to
escape from the local search space to find better solutions.
More precisely, the perturbation temperature Tk is defined as
8
f (zbest ) f (zk )
max
>
<Tmax
f
(zbest ) f (zworst )
max
Tk =
>
:Tmax max 1
max
N
if Hi P,
(6)
otherwise,
where Tk (1 k N ) denotes the temperature assigned

to the kth individual, Tmax the maximum temperature, max
the maximum number of iterations LLH will be performed,
and the number of iterations LLH is actually performed.
Also, f () is the fitness function; zk , zbest , and zworst are,
respectively, the finesses of the kth solution, the best fitness of
the population, and the worst fitness of the population. If the
low-level heuristic is an SSBHA, it implies that no population
information will be available. In this case, the perturbation
operator will only use the number of iterations to decide the
temperature of the current best individual from the population
of LLH which will be passed to the newly selected LLH.
(b)
Fig. 5. Example showing how CloudSim works by assigning tasks to VMs.
(a) virtual machines. (b) workflow tasks.
B. An Example
A simple example is given in Fig. 4 to depict how HHSA
works. As this example shows, there are two cases to deal with:
one is SSBHAs while the other is PBHAs. If the randomly
selected algorithm is an SSBHA, such as simulation annealing
(SA), then it will be evolved by using the improvement
detection operator to check if there is an improvement in the
fitness of the current solution. On the contrary, if a PBHA
is chosen, it is the diversity of the current solutions (i.e., the
solutions in the population) that will be checked by using the
diversity detection operator before the improvement detection
operator is applied. After that, the perturbation operator will
be applied to either the current solution or the population
before they are passed on to the newly selected LLH. This
process will be repeated until the stop condition is reached. As
shown in Fig. 4, the hyper-heuristic algorithm contains many
low-level heuristic algorithms. After the candidate low-level
heuristic algorithms to be in the pool are decided, the proposed
algorithm will randomly select a low-level heuristic algorithm
Hi from the pool. Then, if Hi is a PBHA, the proposed
algorithm will first perform the diversity detection operator.
Finally, for both the PBHA and SSBHA, the improvement
detection and perturbation operators will be performed. The
process will be repeated until a new low-level heuristic algorithm is selected.
TABLE I
PARAMETER SETTINGS OF THE FIVE CLOUD SCHEDULING ALGORITHMS
FOR W ORKFLOW S CHEDULING ON C LOUD S IM
Algorithm
SA
GA
ACO
PSO
HHSA
Parameters
temperature t = 10
temperature reduce rate tr = 0.9
mutation probability m = 0.95
pheromone updating fact = 0.1
choosing probability q0 = 0.85
related influence weights = = 1
inertia weight = 0.8
acceleration coefficient c1 = c2 = 1.0
max iteration of low-level algorithm max = 50
non-improved iteration threshold ni = 5
V. R ESULTS
In this section, two kinds of scheduling problems are used
to evaluate the performance of the proposed algorithm. One
is the workflow scheduling problem; the other is the Hadoop
map-task scheduling problem.
A. Workflow Scheduling on CloudSim
1) Datasets and Parameter Settings: CloudSim [14] is a
tool for emulating a cloud computing environment. In this
study, CloudSim is used as the emulator in solving the workflow scheduling problem. As shown in Fig. 5(a), CloudSim can
be used to construct a data center with a set of virtual machines
as the resource. Each task of a workflow can be started by
assigning it to a VM once its parent tasks are completed. For
example, as shown in Fig. 5(b), task 3 will be started once
its parent tasks 0 and 1 are completed. Besides, the starting
and ending tasks, s and e, are virtual task and are not to be
assigned to any VM. For more details and other workflow
examples, readers are referred to [49] and [50].
The empirical analysis was conducted on an ASUS PC with
2.67 GHz Intel i7-920 CPU and 4GB of memory running
Fedora 12 with Linux 2.6.31 and using CloudSim to construct
four virtual machines in a data center. The workflow datasets
from the studies described in [51], [52] and PSPLIB [53]
are employed to compare the performance of the proposed
algorithm and other scheduling algorithms evaluated in this
paper. Among these datasets, the e-Economic, fMRI, and
protein annotation have, respectively, 10, 15, and 15 tasks
[51], [52] while the j30, j60, and j90 have, respectively, 30,
60, and 90 tasks [53]. More precisely, fMRI is a dataset for
evaluating workflow that can be processed in parallel while
e-Economic is a dataset for evaluating workflow that can be
processed sequentially. For a scheduling problem, the dataset
of [53] contains serial and parallel tasks. Each simulation was
carried out 30 runs. The maximum number of iterations each
run is set equal to 1,000. The population size is set equal to
50 for PBHAs. In HHSA, the maximum number of iterations
of low-level algorithm max is set equal to 50, and the nonimproved iteration threshold ni is set equal to 5. The other
parameter settings of the scheduling algorithms are as shown
in Table I.
2) Simulation Results of Workflow Scheduling: To evaluate
the performance of HHSA for the workflow scheduling problem, we compare it with two traditional rule-based algorithms
and four heuristic algorithms, namely, min-min [54], maxmin [54], simulated annealing [55], genetic algorithm [56],
particle swarm optimization [57], and ant colony optimization
[58]. For the datasets e-Economic, e-Protein, and fMRI, the
results in Table II show that heuristics usually can find better
results than traditional scheduling algorithms (i.e., min-min,
max-min, and FIFO) in terms of makespan. In addition, the
results of ACO are similar to those of the proposed algorithm
for the first three datasets shown in Table II. For the large
datasets, j30, j60, and j90, the results of four heuristics (SA,
GA, PSO, and ACO) are better than min-min, max-min, and
FIFO. Moreover, the results also show that HHSA outperforms
all the other scheduling algorithms, namely, min-min, maxmin, FIFO, SA, GA, PSO, and ACO for the last three datasets
shown in Table II.
The results of best-so-far makespan given in Fig. 6 show
that the min-min and max-min get almost exactly the same
makespan in all iterations. For the SA, the higher the temperature which will get a larger fine-tune rate for the current
solution, the poorer the best-so-far makespan compared to
the other non-traditional algorithms at the early iterations. On
the other hand, with a lower temperature, SA will converge
quickly. For the population-based heuristic algorithms GA,
PSO, and ACO, the search directions will converge in about
500 iterations. Fig. 6 also shows that HHSA converges faster
than the other heuristic algorithms, and it usually takes no
more than 200 iterations. In addition, because HHSA is able to
automatically choose the suitable low-level heuristic algorithm
and use the perturbation method to improve the results, the end
results of HHSA are usually better than the other scheduling
algorithms, especially for complex and large-scale problems.
A closer look at these simulation results shows that the maxmin algorithm does not scale well as the number of tasks
increases. For example, for small datasets, the results depicted
in Fig. 6(a), (b), and (c) show that max-min can provide a
result that is close to the other heuristic scheduling algorithms
in terms of the makespan (more precisely, the differences are
no more than 3). However, as the number of tasks increases,
the results described in Fig. 6(d), (e), and (f) show that the
differences between max-min and other heuristic scheduling
algorithms are also increased. For instance, Fig. 6(d) shows
that the differences between max-min and other heuristic
scheduling algorithms are more than five; Fig. 6(e) and (f)
show that the differences between them are more than ten. On
the other hand, for the min-min algorithm, the results given in
Fig. 6(d), (e) and (f) show that min-min is closer to the other
heuristic scheduling algorithm than max-min for large-scale
datasets. Fig. 6 also provides the convergence information
of the scheduling algorithms compared in this study. The
results given in Fig. 6(a) show that the search process of
these scheduling algorithms will converge to a stable state
very quickly when the dataset is small. A good example is
the results of ACO and HHSA, which show that the result at
iteration 100 is very close to the result at iteration 1,000. But
the results depicted in Fig. 6(b) show that the search process
of ACO and HHSA can still find a better result after iteration
100, meaning that the heuristic scheduling algorithms still have
a chance to find a better result at later iterations when the
TABLE II
S IMULATION RESULTS OF THE WORKFLOW SCHEDULING ALGORITHMS EVALUATED IN THIS STUDY
dataset
min min
max min
FIFO
SA
GA
Average
5.91
5.50
7.00
5.17
5.21
Best
5.91
5.50
7.00
5.16
5.16
e-Economic
Worst
5.91
5.50
7.00
5.25
5.41
S.D.
0.00
0.00
0.00
0.02
0.09
Average
8.75
8.75
9.38
7.87
7.76
Best
8.75
8.75
9.38
7.62
7.50
e-Protein
Worst
8.75
8.75
9.38
8.25
8.25
S.D.
0.00
0.00
0.00
0.17
0.27
Average
10.0
10.0
10.7
9.74
9.64
Best
10.0
10.0
10.7
9.50
9.50
fMRI
Worst
10.0
10.0
10.7
10.0
10.0
S.D.
0.00
0.00
0.00
0.16
0.14
Average
18.7
20.3
20.5
16.7
15.8
Best
18.7
20.3
20.5
16.0
15.0
j30
Worst
18.7
20.3
20.5
17.5
16.5
S.D.
0.00
0.00
0.00
0.35
0.35
Average
37.5
41.1
41.5
35.9
33.2
Best
37.5
41.1
41.5
34.0
31.5
j60
Worst
37.5
41.1
41.5
38.8
34.6
S.D.
0.00
0.00
0.00
0.97
0.78
Average
56.0
61.9
54.0
55.9
50.8
Best
56.0
61.9
54.0
53.1
48.5
j90
Worst
56.0
61.9
54.0
59.8
53.7
S.D.
0.00
0.00
0.00
1.71
1.20
N.B.: The units of measurement for each column is in seconds. S.D. represents
10
5.8
5.6
MIN_MIN
MAX_MIN
SA
GA
PSO
ACO
HHSA
11
11.5
MIN_MIN
MAX_MIN
SA
GA
PSO
ACO
HHSA
10.5
6.2
11
MIN_MIN
MAX_MIN
SA
GA
PSO
ACO
HHSA
6.4
PSO
ACO
HHSA
5.26
5.16
5.16
5.16
5.16
5.16
5.75
5.16
5.16
0.13
0.00
0.00
7.83
7.50
7.50
7.50
7.50
7.50
8.5
7.50
7.50
0.28
0.00
0.00
9.66
9.50
9.50
9.50
9.50
9.50
10.0
9.50
9.50
0.15
0.00
0.00
16.2
15.8
15.4
15.6
15.5
15.0
17.0
16.2
15.8
0.39
0.18
0.15
34.2
32.7
31.9
32.3
31.6
31.0
36.5
33.2
32.5
1.08
0.34
0.46
51.2
48.9
48.0
49.0
48.2
47.0
54.6
49.8
48.7
1.44
0.33
0.46
the standard deviation.
9.5
9
8.5
10.5
10
5.4
8
5.2
9.5
7.5
5
200
400
600
800
1000
200
400
Iteration
800
1000
24
200
400
50
21
20
19
18
800
1000
(c) fMRI
MIN_MIN
MAX_MIN
SA
GA
PSO
ACO
HHSA
45
22
600
Iteration
(b) e-Protein
MIN_MIN
MAX_MIN
SA
GA
PSO
ACO
HHSA
23
MIN_MIN
MAX_MIN
SA
GA
PSO
ACO
HHSA
70
65
40
60
55
35
17
50
16
15
30
0
200
400
600
800
1000
200
400
Iteration
(d) j30
Fig. 6.
Iteration
(a) e-Economic
600
600
Iteration
(e) j60
800
1000
200
400
600
800
1000
Iteration
(f) j90
Convergence analysis by best-so-far makespan.
dataset becomes large. In Fig. 6(c) to (f), this kind of results

becomes quite clear. In addition, based on the observations of
the convergence of these scheduling algorithms, we can easily
know their performance. For example, for small datasets, the
convergence speeds of ACO and HHSA described in Fig. 6(a),
(b), and (c) are quite close to each other. However, as the
number of tasks increases, the convergence speeds of ACO
and HHSA are as shown in Fig. 6(d), (e), and (f). They show
that HHSA has a higher chance to find a better result than
ACO does because it keeps finding a better result after the
ACO has reached a stable state.
B. Hadoop map-task scheduling
1) Datasets and Parameter Settings: Another distributed
system, Hadoop, provides slots as the resources to be allocated
to the independent jobs, as shown in Fig. 7. Each job will

be divided into multiple independent map tasks according to
a predefined map size, and the number of reduce tasks is
given when submitting the jobs to the Hadoop cluster. Each
scheduler is carried out for 30 runs, with Hadoop restarted
at the end of each run. The maximum number of iterations
each run is set equal to 50; the population size is set equal
to 10 for PBHAs. The parameter settings of the scheduling
algorithms of Hadoop are slightly different from those used in
the simulations on CloudSim and are as shown in Table III. By
the nature of most evolutionary algorithms, the computation of
fitness (makespan in this case) is required. In order to compute
the makespan, the algorithm needs to know the job size and
the processing speed. For the simulations, both attributes are
given in the dataset. But on Hadoop users can only acquire the
Map
Task
Job
Reduce
slot
TABLE IV
DATASET GENERATED FOR BENCHMARK
slot
1
slot
slot
slot
2
# Maps
1
2
3-20
21-60
% Jobs in Benchmark
56%
14%
16%
14%
# Jobs in Benchmark
28
7
8
7
slot
slot
Fig. 7.
Job executing on Hadoop.
TABLE III
PARAMETERS OF THE FIVE CLOUD SCHEDULING ALGORITHMS
EVALUATED
Algorithm
SA
GA
ACO
PSO
HHSA
Parameters
temperature t = 10
temperature reduce rate tr = 0.9
mutation probability m = 0.95
trail persistence = 0.8
choosing probability q0 = 0.85
relative importance = = 1
inertia weight = 0.8
acceleration coefficient c1 = c2 = 1.0
the max number of iterations of LLH max = 50
the max number of non-improved iterations ni = 10
size of each map task. The size of each reduce task remains
unknown. Since the makespan of each reduce task cannot be
predicted, only map task scheduling is applicable. Therefore,
for the proposed algorithm, our focus is on scheduling tasks
to map slots.
As mentioned previously, in addition to the size of each
map task, we also need the processing speed of each task. To
obtain this critical attribute, the algorithm first checks if the
job exists in the record of processing speed. If not, one of the
tasks in the job will be prioritized and scheduled immediately.
After the map task is finished, we can divide the size of the
map task by the processing time to get the speed which then
can be used to calculate the processing time of the rest of the
map tasks and to predict the total makespan.
In order to simulate the production environment, the job
inter-arrival times in the experiment are roughly exponentially
distributed with a mean of 14 seconds, as randomly sampled
from a Facebook trace [41]. The datasets for the grep jobs on
Hadoop are strings each of size 100 bytes and each on a line
by itself randomly generated using a Python script. The sizes
and submission periods are also randomly generated based on
the statistics described in [41]. Fifty jobs are submitted in the
experiment, and the total submission schedule is 13 minutes
long. The sizes and the distributions of the jobs are as listed
in Table IV. In this study, we also tried submitting larger jobs
in a shorter period. More precisely, in the test, jobs to sort two
1GB, two 3GB, two 5GB, and one 10GB files are submitted
in one minute.
In this study, Hadoop was running on both single-node
(machine) and multi-node setups in a virtualized environment.
The single-node setup is an Intel i7-3820 with 64G RAM,
which runs Gentoo Linux with kernel version 3.7.10. Hadoop
was compiled from source version 1.1.3 using sun-jdk-1.6
toolchain. By properly configuring the map and task capacities
of Hadoop, a single machine can still provide multiple working
slots. For example, a machine with 8 map task capacities

is equivalent to a cluster consisting of 4 machines each of
which provide 2 map task capacities. Meanwhile, each slot
can be treated as an independent virtual machine in the
Cloud environment, which can also be mapped to the virtual
machines of CloudSim in the simulation.
For the multi-node setup, the Hadoop cluster was running on
four virtual machines hosted on a cluster with AMD Opteron
6128 processors running XenServer. Each virtual machine has
eight cores and 16GB RAM and runs CentOS with kernel
version 2.6.32. Hadoop was compiled from source version
1.1.3 using sun-jdk-1.6 toolchain. One of the four virtual
machines is configured as the worker and job tracker node
for scheduling jobs. The other three are configured as worker
nodes. Each node is configured with 4 map slots, implying
that there are total 16 map slots.
This study uses the map/reduce grep program distributed
with Hadoop for the benchmark. The datasets are randomly
generated so that each line contains 100 bytes of data. To
evaluate the performance of HHSA, the same datasets are fed
to FIFO and Fair Scheduler [41]. For larger datasets, TeraSort
[59], which is also provided by Hadoop, is used to generate
and sort the data. Moreover, another dataset, retrieved from
the log of 1998 World Cup website [60] provided by Internet
Traffic Archive, for analyzing the web user access behavior is
used to evaluate the performance of the proposed algorithm
when it is applied to a real complex scheduling problem. In
this experiment, the data are used by a map/reduce web log
analyzer to calculate the reference counts of URLs.
2) Simulation Results of Hadoop map-task scheduling:
The experimental results of Hadoop are as shown in Table V
and Table VI. For the grep jobs, the three schedulers give
the similar results in both the single-node and multi-node
setup. The results show that for the grep jobs, the Hadoop
cluster was under-loaded, meaning that when new jobs arrived,
previous jobs and tasks had already finished, thus making the
scheduling results, i.e., the performance, of three schedulers
similar to each other. This is justified by the average task
waiting time depicted in Fig. 10, to be discussed in later
section.
When facing a burst, HHSA can make a better scheduling
decision. For the TeraSort jobs, HHSA outperforms the other
two schedulers. In the single-node setup, HHSA beats Fair
Scheduler and FIFO by 17.19% and 30.75%, respectively, on
average. In the multi-node setup, HHSA again outperforms
the other two schedulers. On average, HHSA is 11.36%
faster than FIFO and 4.56% faster than Fair Scheduler. As
the results show, by optimizing the makespan, HHSA can
provide a better scheduling decision in terms of the best-so-far
makespan. The results of web log analyzer show that HHSA
performs marginally better than FIFO, by 2.62% and 0.25%,

respectively, in the single-node and multi-node setup. This is
due to the fact that the web log analyzing job takes O(n)
time; thus, the time costs for all the tasks are very close to
each other. Under this circumstance, FIFO can easily produce
a good schedule that is close to that produced by the proposed
algorithm. On the other hand, the time cost of each TeraSort
task varies greatly, thus largely increasing the difficulty of
scheduling. In this case, HHSA outperforms the other two
schedulers significantly.
The performance of FIFO and HHSA on CloudSim and
Hadoop is as shown in Table VII. For both CloudSim and
Hadoop, a set of jobs the size of which are close to those
of TeraSort benchmarks are generated and used. Then, the
size of each task is randomly generated so that it falls in
the range of 64MB 30%; each job will be divided into a
set of smaller tasks of the same size though the processing
time of each task is still not necessarily the same. When
most tasks are of the same size, FIFO, which uses greedy
strategy to scheduling tasks, can achieve a high utilization and
good makespan. However, when the tasks are of different size,
FIFO will become the worst method. Moreover, the simulation
results also show that HHSA outperforms FIFO, especially
when the problem is complex. The results are consistent with
those of Table II, where the sizes of jobs in the datasets for
CloudSim are more diverse than those in this experiment.
C. Analysis
10
TABLE V
P ERFORMANCE OF HHSA ON H ADOOP WITH S INGLE -N ODE ( IN
SECONDS )
Task Type
grep
Log Analyzer
TeraSort
Average
Best
Worst
S.D.
Average
Best
Worst
S.D.
Average
Best
Worst
S.D.
FIFO
858
857
862
1.00
1065
1042
1094
14.93
904
838
1024
43.54
Fair Scheduler
861
859
864
1.41
1040
1037
1048
2.65
756
710
833
29.42
TABLE VI
P ERFORMANCE OF HHSA ON H ADOOP WITH M ULTI -N ODE ( IN SECONDS )
Task Type
grep
Log Analyzer
TeraSort
Average
Best
Worst
S.D.
Average
Best
Worst
S.D.
Average
Best
Worst
S.D.
FIFO
851
848
861
2.23
795
752
832
21.07
1180
1077
1326
65.12
Fair Scheduler
852
843
861
3.87
802
749
876
26.67
1096
1027
1249
46.97
HHSA
853
848
861
3.16
793
741
829
22.56
1046
932
1178
66.02
TABLE VII
P ERFORMANCE OF HHSA ON C LOUD S IM AND H ADOOP ( IN SECONDS )
Platform
The maximum number of iterations of LLH max and the

maximum number of non-improved iterations ni set up for
the workflow scheduling problem may not be suitable for the
Hadoop map-task scheduling problem. As such, to evaluate the
performance of HHSA for the Hadoop map-task scheduling
problem, four datasets with a varying number of independent
tasks, namely, 25 tasks, 50 tasks, 100 tasks, and 200 tasks, are
used. Here, our focus is on the average makespan of 30 runs
with max in the range of 20 to and ni in the range of 1 to
where is the maximum number of iterations of HHSA.
This experiment is aimed at measuring the performance of
HHSA when applying to the Hadoop map-task scheduling
problem with the population size fixed at 50 for the PBHAs.
This eventually takes into consideration all the combinations
of the parameter settings shown in Table VIII.
1) The maximum number of iterations of LLH: Fig. 8 shows
the makespans for a varying maximum number of iterations
of LLH max in the range [20, ] and a varying maximum
number of iterations {250, 500, 750, 1000} for each
dataset. It is worth noting that the maximum number of nonimproved iterations ni is set equal to max . This test is aimed
to help us understand whether the value of max will influence
the scheduling result. As the results in Fig. 8 show, the
larger the number of iterations, the better the makespan. For
instance, 250 iterations give a worse result than the other three
cases. The results further show that the lower the maximum
number of iterations of low-level algorithm max , the better
the makespan. Apparently, as the value of max approaches the
maximum number of iterations, it implies that the probability
HHSA
859
850
866
5.19
1037
1018
1048
6.78
626
531
829
66.98
CloudSim
Hadoop (Multi-Node)
Average
Best
Worst
S.D.
Average
Best
Worst
S.D.
FIFO
989
989
989
0.00
1180
1077
1326
65.12
HHSA
952
937
980
12.29
1046
932
1178
66.02
TABLE VIII
PARAMETER SETTINGS FOR THE ANALYSIS
Parameter
max number of iterations ()
max number of iterations of LLH (max )
max number of non-improved iterations (ni )
Settings
250, 500, 750, 1000
[20, ]
[1, ]
of using LLH is getting smaller. In the worse case, only one

of the LLHs will be used. That is, HHSA will degenerate to
one of the low-level heuristics, which in turn will reduce the
chance of balancing the intensification and diversification in
the search of solutions. In addition, Fig. 8(b), (c) and (d) show
that for problems larger than that given in Fig. 8(a), a similar
result is obtained; that is, decreasing the value of max for the
proposed algorithm can provide better results in most cases in
terms of makespan. Consequently, as Fig. 8 shows, the smaller
the value of max , the better the makespan.
2) The maximum number of non-improved iterations:
Furthermore, Fig. 9 shows the makespans for a varying maximum number of non-improved iterations ni
{1, 2, , } and a varying maximum number of iterations
{250, 500, 750, 1000} for each dataset. In order to
26.05
50.2
250
500
750
1000
26
25.95
250
500
750
1000
50.1
25.9
50
25.85
49.9
makespan
makespan
11
25.8
25.75
49.8
25.7
49.7
25.65
49.6
25.6
49.5
25.55
25.5
49.4
0
200
400
600
LLH max iterations
800
1000
(a) Average makespan of 25 tasks

75.8
1000
166
makespan
makespan
800
250
500
750
1000
166.5
75.4
75.2
165.5
75
165
74.8
164.5
74.6
164
0
200
400
600
LLH max iterations
800
1000
(c) Average makespan of 100 tasks

Fig. 8.
400
600
LLH max iterations
167
250
500
750
1000
75.6
200
(b) Average makespan of 50 tasks
200
400
600
LLH max iterations
800
1000
(d) Average makespan of 200 tasks
Average makespan for a varying number of tasks and max .

26.1
50.3
250
500
750
1000
26
50.1
50
makespan
25.9
makespan
250
500
750
1000
50.2
25.8
25.7
49.9
49.8
49.7
49.6
25.6
49.5
25.5
49.4
0
200
400
600
non-improve iterations
800
1000
(a) Average makespan of 25 tasks

76.2
75.8
800
1000
250
500
750
1000
166.5
166
makespan
75.6
makespan
400
600
167
250
500
750
1000
76
200
(b) Average makespan of 50 tasks
75.4
75.2
75
165.5
165
74.8
164.5
74.6
74.4
164
0
200
400
600
800
1000
(c) Average makespan of 100 tasks

Fig. 9.
200
400
600
800
1000
(d) Average makespan of 200 tasks
Average makespans for a varying number of tasks and ni .
eliminate the influence the maximum number of iterations

of the selected LLH may have, the value of max is always
set equal to the maximum number of iterations . When the
maximum number of non-improved iterations is set equal to
a small value, it means that if the best-so-far makespan of
the selected LLH is not improved within the given number
of iterations, the selected LLH will be replaced by another
LLH. This implies that a smaller value of ni will increase
the diversity of the search directions of the selected LLH, thus
improving the quality for not only the small problem shown
in Fig. 9(a) but also the larger and complex problems depicted
in Fig. 9(b), (c), and (d). In contrast, a larger value of ni will
increase the speed of convergence. Since the diversity and the
quality of makespan may not be improved at the same time,
our observation shows that a suitable value of ni can not only
get a better mean makespan, it also uses LLH more efficiently.
3) The average task waiting time: The average task waiting

times for grep and TeraSort in the single-node and multinode setup on Hadoop are as shown in Fig. 10. The small
waiting times of grep given in Fig. 10(a) imply that the
cluster is under-loaded and thus most tasks can be scheduled
immediately. These results also explain that for small datasets
in a cluster that is under-loaded, the performance of these
scheduling algorithms is pretty much the same. On the other
hand, the results of TeraSort described in Fig. 10(b) indicate
that for large datasets in a cluster that is fully loaded, compared
to FIFO, HHSA can reduce the average task waiting time of
TeraSort by 29.04%2 and 12.49%,3 respectively, in the singlenode and multi-node setup. This means that by optimizing the
makespan, tasks in the queue can be completed earlier; thus,
2 [(374.90
3 [(432.60
266.03)/374.90] 100 = 29.04%.

378.56)/432.60] 100 = 12.49%.
Time in seconds
single-node
multi-node
6
5.00
5.10
4.80
2.00
1.93
1.27
0
FIFO
Fair
HHSA
Scheduler
(a) grep
500
single-node
multi-node
Time in seconds
432.60
400
381.93
374.90
378.57
303.97
300
266.03
200
FIFO
Fair
HHSA
Scheduler
(b) TeraSort
Fig. 10.
Average task waiting times for grep and TeraSort.
incoming tasks can be scheduled sooner and the average task

waiting time can be further reduced. Moreover, the results of
HHSA for TeraSort on Hadoop show that HHSA takes only
0.0144 milliseconds on average, meaning that the computation
time of HHSA is negligible in most cases.
4) The cloud rental payment: In terms of makespan, on
Hadoop, HHSA is about 11.36% faster than FIFO on average.
The multi-node setup used in this paper is based on the
Standard On-Demand Extra Large Instances of Amazon
Elastic MapReduce service [61], which costs 0.12 US dollars
per hour. Though the proposed algorithm described herein
is aimed at optimizing the makespan instead of the budget.
The reality is that by reducing the makespan, it also reduces
the total cost of cloud service. According to the pricing and
taking into account the improvement of the makespan, the
proposed method can reduce 0.12 11.36% = 0.014 US
dollars per hour for each standard on-demand extra large
instance. Although HHSA takes a longer computation time
than the other schedulers, the difference is negligible because
it is compensated by a better makespan, which will save much
more time than HHSA takes.
D. Summary
As mentioned in Section II-A2, the characteristics of a
cloud system are quite different from those of the information
systems available today. A typical characteristic of a cloud
system is in that the number of tasks it has to handle is much
larger than that an information system available today has to
handle, thus making it hard to schedule all the tasks. Another
typical characteristic that affects the design of scheduling
algorithms for a cloud system is in that a cloud system usually
contains a number of virtual machines (or nodes) that add more
uncertainties for task scheduling. Still another characteristic is
in that a cloud system usually has many heterogeneous tasks to
be scheduled, thus making it much more difficult to schedule
than tasks on the information systems available today.
To solve these problems, a high-performance scheduling
algorithm which can provide a better scheduling result in
12
a reasonable time is needed. This explains why a hyperheuristic-based scheduling algorithm is presented in this paper,
which uses various search strategies to solve the scheduling
problem of a cloud for providing a better result than traditional scheduling algorithms and other heuristic scheduling
algorithms do. Because the design of the proposed algorithm
is not for a particular cloud system, it can be applied to
many applications that require task scheduling, such as science computing on Cloud. But the high latency inherent in
the heuristic-based algorithms will make them unsuitable for
the low-level scheduling of IaaS. Moreover, to the best of
our knowledge, there exists no general and efficient way to
automatically determine the number of resources that is used
for each application for a very simple reason. The number
of resources required may differ for different input to the
application.
Compared to most traditional rule-based deterministic
scheduling algorithms (e.g., min-min, max-min, and FIFO)
that use only one search direction, the proposed algorithm
uses multiple search directions at each iteration during the
convergence process, thus having a higher chance to find a
better result than the rule-based deterministic algorithms do.
The results depicted in Sections V-A2 and V-B2 and Table VI
show that in terms of makespan, the proposed algorithm outperforms the traditional scheduling algorithms in most cases,
for not only the single-node but also the multi-node datasets.
Compared to the other heuristic scheduling algorithms, the
proposed algorithm, again in terms of makespan, can still
provide a better result than the other heuristic scheduling
algorithms, in terms of not only the results of CloudSim
but also the results of Hadoop, as shown in Sections V-A2
and V-B2. One of the reasons that can be easily justified is
that HHSA leverages the strengths of many search strategies
provided by different heuristic scheduling algorithms (lowlevel algorithms), thus making it more diverse in terms of
the directions for which it searches during the convergence
process, especially when compared to the other heuristic
scheduling algorithms.
For the stability issue, HHSA is not designed for interconnectivity and locality because they will degrade the performance of Hadoop. As the authors of [62] pointed out, if
the nodes of a virtualized Hadoop cluster actually reside in
the same physical node, the overhead incurred for transferring
data between these nodes is generally smaller than transferring
the data between physical nodes. Hence, if the overhead of
connection is too large but the scheduling problem is too small,
the performance of the proposed algorithm will be degraded;
otherwise, the performance of the proposed algorithm will be
better than the other scheduling algorithms, especially when
the scheduling problem becomes too complex or too large.
VI. C ONCLUSIONS
This study presents a high-performance hyper-heuristic algorithm to find better scheduling solutions for cloud computing systems. The proposed algorithm uses two detection
operators to automatically determine when to change the lowlevel heuristic algorithm and a perturbation operator to finetune the solutions obtained by each low-level algorithm to
further improve the scheduling results in terms of makespan.

As the simulation results show, the proposed algorithm can
not only provide better results than the traditional rule-based
scheduling algorithms, it also outperforms the other heuristic
scheduling algorithms, in solving the workflow scheduling and
Hadoop map-task scheduling problems on cloud computing
environments. In addition, the simulation results show further
that the proposed algorithm converges faster than the other
heuristic algorithms evaluated in this study for most of the
datasets. In brief, the basic idea of the proposed hyperheuristic algorithm is to leverage the strengths of all the lowlevel algorithms while not increasing the computation time,
by running one and only one low-level algorithm at each
iteration. This is fundamentally different from the so-called
hybrid heuristic algorithm, which runs more than one lowlevel algorithm at each iteration, thus requiring a much longer
computation time. Our experiences show that the proposed
algorithm can be applied to different cloud computing systems
to enhance the performance of scheduling problems. In the
future, our focus will be on finding more effective detection
operators and perturbation method to enhance the performance
of the proposed algorithm. We will also attempt to apply
HHSA to different research domains.
ACKNOWLEDGMENT
The authors would like to thank the editors and anonymous
reviewers for their valuable comments and suggestions on the
paper that greatly improve the quality of the paper.
R EFERENCES
[1] M. Pinedo, Scheduling: Theory, Algorithms and Systems. Prentice Hall,
Inc., 1995.
[2] P. Chretienne, E. G. Coffman, J. K. Lenstra, and Z. Liu, Eds., Scheduling
Theory and Its Applications. John Wiely & Sons Ltd, 1995.
[3] A. Allahverdi, C. T. Ng, T. C. E. Cheng, and M. Y. Kovalyov, A survey
of scheduling problems with setup times or costs, European Journal of
Operational Research, vol. 187, no. 3, pp. 9851032, 2008.
[4] M. Armbrust, A. Fox, R. Griffith, A. D. Joseph, R. Katz, A. Konwinski,
G. Lee, D. Patterson, A. Rabkin, I. Stoica, and M. Zaharia, A view
of cloud computing, Communications of the ACM, vol. 53, no. 4, pp.
5058, 2010.
[5] I. Foster, Y. Zhao, I. Raicu, and S. Lu, Cloud computing and grid
computing 360-degree compared, in Proceedings of the Grid Computing Environments Workshop, 2008, pp. 110.
[6] K. Bilal, M. Manzano, S. Khan, E. Calle, K. Li, and A. Zomaya, On
the characterization of the structural robustness of data center networks,
IEEE Transactions on Cloud Computing, vol. 1, no. 1, pp. 6477, 2013.
[7] X. Lei, X. Liao, T. Huang, H. Li, and C. Hu, Outsourcing large matrix
inversion computation to a public cloud, IEEE Transactions on Cloud
Computing, vol. 1, no. 1, pp. 7887, Jan 2013.
[8] M. R. Garey and D. S. Johnson, Computer and Intractability: A Guide
to the Theory of NP-Completeness. New York: W.H. Freeman and
Company, 1979.
[9] C. W. Tsai and J. Rodrigues, Metaheuristic scheduling for cloud: A
survey, IEEE Systems Journal, vol. 8, no. 1, pp. 279297, 2014.
[10] S. Kirkpatrick, C. D. Gelatt, and M. P. Vecchi, Optimization by
simulated annealing, Science, vol. 220, no. 4598, pp. 671680, 1983.
[11] J. H. Holland, Adaptation in Natural and Artificial Systems: An Introductory Analysis with Applications to Biology, Control, and Artificial
Intelligence. University of Michigan Press, 1975.
[12] J. Kennedy and R. C. Eberhart, Particle swarm optimization, in
Proceedings of the IEEE International Conference on Neural Networks,
1995, pp. 19421948.
[13] M. Dorigo and L. M. Gambardella, Ant colony system: A cooperative
learning approach to the traveling salesman problem, IEEE Transactions on Evolutionary Computation, vol. 1, no. 1, pp. 5366, 1997.
13
[14] R. N. Calheiros, R. Ranjan, C. A. F. De Rose, and R. Buyya, Cloudsim:

A novel framework for modeling and simulation of cloud computing infrastructures and services, Computing Research Repository, vol.
abs/0903.2525, 2009.
[15] J. Bazewicz, K. H. Ecker, E. Pesch, G. Schmidt, and J. Weglarz,
Scheduling Computer and Manufacturing Processes. Springer-Verlag
New York, Inc., 2001.
[16] Y. T. J. Leung, Handbook of Scheduling: Algorithms, Models and
Performance Analysis. Chapman & Hall/CRC, 2004.
[17] K. M. Elsayed and A. K. Khattab, Channel-aware earliest deadline due
fair scheduling for wireless multimedia networks, Wireless Personal
Communications, vol. 38, no. 2, pp. 233252, 2006.
[18] W. H. Kohler, A preliminary evaluation of the critical path method
for scheduling tasks on multiprocessor systems, IEEE Transactions on
Computers, vol. C-24, no. 12, pp. 12351238, 1975.
[19] R. Ingalls and D. Morrice, PERT scheduling with resources using qualitative simulation graphs, in Proceedings of the Simulation Conference,
vol. 1, 2000, pp. 362370.
[20] H.-J. Schutz and R. Kolisch, Approximate dynamic programming
for capacity allocation in the service industry, European Journal of
Operational Research, vol. 218, no. 1, pp. 239250, 2012.
[21] D. Shi and T. Chen, Optimal periodic scheduling of sensor networks: A
branch and bound approach, Systems & Control Letters, vol. 62, no. 9,
pp. 732738, 2013.
[22] S. M. Johnson, Optimal two- and three-stage production schedules with
setup times included, Naval Research Logistics Quarterly, vol. 1, no. 1,
pp. 6168, 1954.
[23] M. R. Garey, D. S. Johnson, and R. Sethi, The complexity of flowshop
and jobshop scheduling, Mathematics of Operations Research, vol. 1,
no. 2, pp. 117129, 1976.
[24] D. Daniel and S. J. Lovesum, A novel approach for scheduling service
request in cloud with trust monitor, in Proceedings of the International
Conference on Signal Processing, Communication, Computing and Networking Technologies, 2011, pp. 509513.
[25] Z. Wu, X. Liu, Z. Ni, D. Yuan, and Y. Yang, A market-oriented
hierarchical scheduling strategy incloud workflow systems, The Journal
of Supercomputing, pp. 138, 2011.
[26] B. Saovapakhiran, G. Michailidis, and M. Devetsikiotis, Aggregateddag scheduling for job flow maximization in heterogeneous cloud
computing, in Proceedings of the IEEE Global Telecommunications
Conference, 2011, pp. 16.
[27] M. Rahman, X. Li, and H. Palit, Hybrid heuristic for scheduling
data analytics workflow applications in hybrid cloud environment, in
Proceedings of the IEEE International Symposium on Parallel and
Distributed Processing Workshops, 2011, pp. 966974.
[28] Apache Hadoop. [Online]. Available: http://hadoop.apache.org
[29] Y. M. Huang and Z. H. Nie, Cloud computing with Linux and Apache
Hadoop. [Online]. Available: http://www.ibm.com/developerworks/aix/
library/au-cloud apache/
[30] B.
David,
Open
source
cloud
computing
with
Hadoop. [Online]. Available: http://www.linuxjournal.com/content/
open-source-cloud-computing-hadoop
[31] A. ODriscoll, J. Daugelaite, and R. D. Sleator, Big data, Hadoop
and cloud computing in genomics, Journal of Biomedical Informatics,
vol. 46, no. 5, pp. 774781, 2013.
[32] Y. Song, M. Wu, and L. Yang, Study of smart grid marketing system
architecture based on Hadoop platform of cloud computing, Journal of
Communication and Computer, vol. 9, pp. 741743, 2012.
[33] A. Bala and I. Chana, A survey of various workflow scheduling algorithms in cloud environment, in Proceedings of the National Conference
on Information and Communication Technology, 2011, pp. 2630.
[34] N. Kaur, T. S. Aulakh, and R. S. Cheema, Comparison of workflow
scheduling algorithms in cloud computing, International Journal of
Advanced Computer Science and Applications, vol. 2, no. 10, pp. 8186,
2011.
[35] J. Li, M. Qiu, Z. Ming, G. Quan, X. Qin, and Z. Gu, Online
optimization for scheduling preemptable tasks on IaaS cloud systems,
Journal of Parallel and Distributed Computing, vol. 72, no. 5, pp. 666
677, 2012.
[36] J. Li, M. Qiu, J. Niu, Y. Chen, and Z. Ming, Adaptive resource
allocation for preemptable jobs in cloud systems, in Proceedings
of the International Conference on Intelligent Systems Design and
Applications, 2010, pp. 3136.
[37] C. Lin and S. Lu, Scheduling scientific workflows elastically for cloud
computing, in Proceedings of the IEEE International Conference on
Cloud Computing, 2011, pp. 746747.
[38] S. Abrishami and M. Naghibzadeh, Deadline-constrained workflow

scheduling in software as a service cloud, Scientia Iranica, vol. 19,
no. 3, pp. 680689, 2012.
[39] Y. C. Lee, C. Wang, A. Y. Zomaya, and B. B. Zhou, Profit-driven
scheduling for cloud services with data access awareness, Journal of
Parallel and Distributed Computing, vol. 72, no. 4, pp. 591602, 2012.
[40] K. Kousalya and P. Balasubramanie, Ant algorithm for grid scheduling
powered by local search, International Journal of Open Problems in
Computer Science and Mathematics, vol. 1, no. 3, pp. 222240, 2008.
[41] M. Zaharia, D. Borthakur, J. Sen Sarma, K. Elmeleegy, S. Shenker, and
I. Stoica, Delay scheduling: a simple technique for achieving locality
and fairness in cluster scheduling, in Proceedings of the European
conference on Computer systems, 2010, pp. 265278.
[42] Yahoo!, Hadoop capacity scheduler. [Online]. Available:
http://hadoop.apache.org/docs/r2.2.0/hadoop-yarn/hadoop-yarn-site/
CapacityScheduler.html
[43] H. H. Hoos and T. Stutzle, Stochastic Local Search: Foundations &
Applications. Elsevier / Morgan Kaufmann, 2004.
[44] P. Cowling and G. Kendall, A hyperheuristic approach to scheduling a
sales summit, in Proceedings of the Practice and Theory of Automated
Timetabling III, 2001, pp. 176190.
[45] P. I. Cowling, G. Kendall, and E. Soubeiga, A hyperheuristic approach
to scheduling a sales summit, in Proceedings of the Practice and Theory
of Automated Timetabling, 2000, pp. 176190.
[46] B. K. Ambati, J. Ambati, and M. M. Mokhtar, Heuristic combinatorial
optimization by simulated Darwinian evolution: A polynomial time
algorithm for the traveling salesman problem, Biological Cybernetics,
vol. 65, no. 1, pp. 3135, 1991.
[47] L. Y. Tseng and S. B. Yang, A genetic approach to the automatic
clustering problem, Pattern Recognition, vol. 34, no. 2, pp. 415424,
2001.
[48] D. Laha and U. Chakraborty, An efficient hybrid heuristic for makespan
minimization in permutation flow shop scheduling, The International
Journal of Advanced Manufacturing Technology, vol. 44, no. 5-6, pp.
559569, 2009.
[49] L. Ramakrishnan, C. Koelbel, Y.-S. Kee, R. Wolski, D. Nurmi, D. Gannon, G. Obertelli, A. YarKhan, A. Mandal, T. M. Huang, K. Thyagaraja,
and D. Zagorodnov, VGrADS: Enabling e-science workflows on grids
and clouds with fault tolerance, in Proceedings of the Conference on
High Performance Computing Networking, Storage and Analysis, 2009,
pp. 47:147:12.
[50] M. Mao and M. Humphrey, Auto-scaling to minimize cost and meet
application deadlines in cloud workflows, in Proceedings of the International Conference for High Performance Computing, Networking,
Storage and Analysis, 2011, pp. 112.
[51] Y. Zhao, M. Wilde, I. Foster, J. Voeckler, T. Jordan, E. Quigg, and
J. Dobson, Grid middleware services for virtual data discovery, composition, and integration, in Proceedings of the Workshop on Middleware
for Grid Computing, 2004, pp. 5762.
[52] A. OBrien, S. Newhouse, and J. Darlington, Mapping of scientific
workflow within the e-protein project to distributed resources, in
Proceedings of the UK e-Science All Hands Meeting, UK, 2004, pp.
404409.
[53] R. Kolisch and A. Sprecher, PSPLIB - a project scheduling problem
library: OR software - ORSEP operations research software exchange
program, European Journal of Operational Research, vol. 96, no. 1,
pp. 205216, 1997.
[54] J. Blythe, S. Jain, E. Deelman, Y. Gil, and K. Vahi, Task scheduling
strategies for workflow-based applications in grids, in Proceedings of
the IEEE International Symposium on Cluster Computing and the Grid,
2005, pp. 759767.
[55] S. Benedict and V.Vasudevan, Scheduling of scientific workflows using
simulated annealing algorithm for computational grids, in Proceedings
of the International Journal of Soft Computing, vol. 2, no. 5, 2007, pp.
606611.
[56] J. Yu and R. Buyya, A budget constrained scheduling of workflow
applications on utility grids using genetic algorithms, in Proceedings
of the Workflows in Support of Large-Scale Science, 2006, pp. 110.
[57] T. Chen, B. Zhang, X. Hao, and Y. Dai, Task scheduling in grid based
on particle swarm optimization, in Proceedings of the International
Symposium on Parallel and Distributed Computing, 2006, pp. 238245.
[58] W. N. Chen and J. Zhang, An ant colony optimization approach to a grid
workflow scheduling problem with various QoS requirements, IEEE
Transactions on Systems, Man, and Cybernetics, Part C: Applications
and Reviews, vol. 39, no. 1, pp. 2943, 2009.
[59] O. OMalley, TeraByte Sort on Apache Hadoop. [Online]. Available:
http://sortbenchmark.org/YahooHadoop.pdf
14
[60] I. T. Archive, World Cup 98. [Online]. Available: http://ita.ee.lbl.gov/

html/contrib/WorldCup.html
[61] Amazon, Amazon elastic mapreduce. [Online]. Available: http:
//aws.amazon.com/elasticmapreduce
[62] H. S. Shin, K. H. Kim, C. Y. Kim, and S. I. Jung, The new approach
for inter-communication between guest domains on virtual machine
monitor, in Proceeding of the International Symposium on Computer
and information sciences, 2007, pp. 16.
Chun-Wei Tsai received the M.S. degree in Management Information System from National Pingtung University of Science and Technology, Pingtung, Taiwan in 2002, and the Ph.D. degree in
Computer Science from National Sun Yat-sen University, Kaohsiung, Taiwan in 2009. He was also
a postdoctoral fellow with the Department of Electrical Engineering, National Cheng Kung University,
Tainan, Taiwan before joining the faculty of Applied
Geoinformatics and then the faculty of Information
Technology, Chia Nan University of Pharmacy &
Science, Tainan, Taiwan in 2010 and 2012, respectively, where he is currently
an Assistant Professor. His research interests include information retrieval,
evolutionary computation, internet technology, and combinatorial optimization.
Wei-Cheng Huang received the B.S. degree in
Computer Science and Information Engineering
from National Pingtung Institute of Commerce,
Pingtung, Taiwan, in 2010, and the M.S. degree
in Computer Science and Engineering from National Sun Yat-sen University, Kaohsiung, Taiwan,
in 2012. He is currently pursuing the Ph.D. degree
in Computer Science and Engineering from National
Sun Yat-sen University. His current research interests include evolutionary computation, clustering,
scheduling, and cloud computing.
Meng-Hsiu Chiang received the B.S. degree in
Computer Science and Engineering from National
Sun Yat-sen University, Kaohsiung, Taiwan, in 2010.
He is currently pursuing the M.S. degree in Computer Science and Engineering from National Sun
Yat-sen University. His current research interests
include scheduling, distributed systems, and cloud
computing.
Ming-Chao Chiang received the B.S. degree in
Management Science from National Chiao Tung
University, Hsinchu, Taiwan, ROC in 1978 and
the M.S., M.Phil., and Ph.D. degrees in Computer
Science from Columbia University, New York, NY,
USA in 1991, 1998, and 1998, respectively. He
has over 12 years of experience in the software
industry encompassing a wide variety of roles and
responsibilities in both large and start-up companies
before joining the faculty of the Department of
Computer Science and Engineering, National Sun
Yat-sen University, Kaohsiung, Taiwan, ROC in 2003, where he is currently an
Associate Professor. His current research interests include image processing,
evolutionary computation, and system software.
Chu-Sing Yang received the B.S. degree in Engineering Science, and the M.S. and Ph.D. degrees in
Electrical Engineering from National Cheng Kung
University, Tainan, Taiwan in 1976, 1984 and 1987,
respectively. He joined the faculty of the Department of Electrical Engineering, National Sun Yatsen University, Kaohsiung, Taiwan, as an Associate
Professor in 1988. Since 1993, he has been a professor of the Department of Computer Science and
Engineering, National Sun Yat-sen University. He
was a Chair of Department of Computer Science
and Engineering, National Sun Yat-sen University from August 1995 to July
1999; a director of Computer Center, National Sun Yat-sen University from
August 1998 to October 2002; and a Program Chair of ICS-96 and Program
Co-Chair of ICPP-2003. He joined the faculty of the Department of Electrical
Engineering, National Cheng Kung University, Tainan, Taiwan, as a Professor
in 2006. His research interests include web mining, embedded system, network
management, and wireless communication.

A Hyper-Heuristic Scheduling Algorithm For Cloud

Uploaded by

Copyright:

Available Formats

A Hyper-Heuristic Scheduling Algorithm For Cloud

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

A Hyper-Heuristic Scheduling Algorithm For Cloud

Uploaded by

Copyright:

Available Formats

This article has been accepted for publication in a future issue of this journal, but has not been

A Hyper-Heuristic Scheduling Algorithm for Cloud

AbstractRule-based scheduling algorithms have been widely

NTIL now we are still actively looking for possible

exhaustive and deterministic algorithms) are widely used on

in detail the proposed algorithm and gives a simple example.

where s is a candidate solution, and, letting Cj denote the

systems. In contrast, by the large scheduling problems, we

subject to Cmax (s) U (s),

where f (s) is the objective function; Cmax (s) the completion

Outline of (a) heuristics, (b) hybrid-heuristics, and (c) hyper-heuristics.

III. I NTEGRATION OF H EURISTIC A LGORITHMS

used or not. For this reason, several LLH selection operators

Set up the parameters.

(a) Best so far makespan of SA.

Hyper-heuristic scheduling algorithm (HHSA).

problems,1 and (2) the time complexity of all the other

means that the solution is encoded as a matrix of size m by n, i.e., a

best so far makespan

(b) Best so far makespan of GA.

shown in line 7) to decide whether to select a new LLH or

where S denotes the set of SSBHAs; P the set of PBHAs.

BSFMK is not improved after tni iterations,

This approach only checks to see whether the solutions found

to the high level hyper center to indicate that it should pick

Low level heuristics

Example showing how HHSA works.

Note that is defined as 3 where and denotes,

where Tk (1 k N ) denotes the temperature assigned

best so far makespan

best so far makespan

best so far makespan

Convergence analysis by best-so-far makespan.

dataset becomes large. In Fig. 6(c) to (f), this kind of results

to the independent jobs, as shown in Fig. 7. Each job will

Job executing on Hadoop.

slots. For example, a machine with 8 map task capacities

performs marginally better than FIFO, by 2.62% and 0.25%,

The maximum number of iterations of LLH max and the

of using LLH is getting smaller. In the worse case, only one

(a) Average makespan of 25 tasks

(c) Average makespan of 100 tasks

(b) Average makespan of 50 tasks

(d) Average makespan of 200 tasks

Average makespan for a varying number of tasks and max .

(a) Average makespan of 25 tasks

(b) Average makespan of 50 tasks

(c) Average makespan of 100 tasks

(d) Average makespan of 200 tasks

Average makespans for a varying number of tasks and ni .

eliminate the influence the maximum number of iterations

3) The average task waiting time: The average task waiting

266.03)/374.90] 100 = 29.04%.

Average task waiting times for grep and TeraSort.

incoming tasks can be scheduled sooner and the average task

further improve the scheduling results in terms of makespan.

[14] R. N. Calheiros, R. Ranjan, C. A. F. De Rose, and R. Buyya, Cloudsim:

[38] S. Abrishami and M. Naghibzadeh, Deadline-constrained workflow

[60] I. T. Archive, World Cup 98. [Online]. Available: http://ita.ee.lbl.gov/