This article was downloaded by: [78.56.45.

90]
On: 18 June 2014, At: 00:21
Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House,
37-41 Mortimer Street, London W1T 3JH, UK

IETE Technical Review

Publication details, including instructions for authors and subscription information:
http://www.tandfonline.com/loi/titr20

Resource Allocation and Scheduling in Cloud

Computing: Policy and Algorithm
a b c b
Tinghuai Ma , Ya Chu , Licheng Zhao & Otgonbayar Ankhbayar
a
Jiangsu Engineering Centre of Network Monitoring, Nanjing University of Information
Science & Technology, Nanjing, 210044, China
b
School of Computer and Software, Nanjing University of Information Science & Technology
University, Nanjing, 210044, China
c
National Information Meteorological Center, China Meteorology Administration, Beijing,
China
Published online: 20 May 2014.

To cite this article: Tinghuai Ma, Ya Chu, Licheng Zhao & Otgonbayar Ankhbayar (2014) Resource Allocation and Scheduling in
Cloud Computing: Policy and Algorithm, IETE Technical Review, 31:1, 4-16

To link to this article: http://dx.doi.org/10.1080/02564602.2014.890837

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”) contained
in the publications on our platform. However, Taylor & Francis, our agents, and our licensors make no
representations or warranties whatsoever as to the accuracy, completeness, or suitability for any purpose of the
Content. Any opinions and views expressed in this publication are the opinions and views of the authors, and
are not the views of or endorsed by Taylor & Francis. The accuracy of the Content should not be relied upon and
should be independently verified with primary sources of information. Taylor and Francis shall not be liable for
any losses, actions, claims, proceedings, demands, costs, expenses, damages, and other liabilities whatsoever
or howsoever caused arising directly or indirectly in connection with, in relation to or arising out of the use of
the Content.

This article may be used for research, teaching, and private study purposes. Any substantial or systematic
reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any
form to anyone is expressly forbidden. Terms & Conditions of access and use can be found at http://
www.tandfonline.com/page/terms-and-conditions
 Resource Allocation and Scheduling in Cloud Computing:
Policy and Algorithm
Tinghuai Ma1, Ya Chu2, Licheng Zhao3 and Otgonbayar Ankhbayar2
1
Jiangsu Engineering Centre of Network Monitoring, Nanjing University of Information Science & Technology, Nanjing, 210044, China,
2
School of Computer and Software, Nanjing University of Information Science & Technology University, Nanjing, 210044, China, 3National
Information Meteorological Center, China Meteorology Administration, Beijing, China

ABSTRACT
Cloud computing is a new distributed commercial computing model that aims at providing computational
resources or services to users over a network in a low-cost manner. Resource allocation and scheduling (RAS)
is the key focus of cloud computing, and its policy and algorithm have a direct effect on cloud performance and
cost. This paper presents five major topics in cloud computing, namely locality-aware task scheduling; reliability-
aware scheduling; energy-aware RAS; Software as a Service (SaaS) layer RAS; and workflow scheduling. These
five topics are then classified into three parts: performance-based RAS; cost-based RAS; and performance- and
cost-based RAS. A number of existing RAS policies and algorithms are discussed in detail accordingly with
regard to their given parameters. In addition, a comparative analysis of five identified problems with their repre-
sentative algorithms is performed. Finally, some future research directions of cloud RAS are pointed out.
Downloaded by [78.56.45.90] at 00:21 18 June 2014

Keywords:
Cloud computing, Data locality, Reliability, Energy efficiency, Resource allocation and scheduling, Software as
a service, Workflow scheduling.

1. INTRODUCTION RAS from different viewpoints. For example, to improve

the task’s execution efficiency and strengthen the reli-
As a new distributed commercial computing model, ability of the cloud system, data locality [4
14] and reli-
cloud computing aims at providing on-demand, ability-aware [15
18] scheduling have been extensively
highly reliable and infinitely scalable computational researched. To minimize operating costs and maximize
resources or services to the remote users over a net- service provider’s profits while meeting QoS require-
work in a low-cost manner. In order to achieve this ment of users, some works focus on energy consump-
goal, data centre management, virtualization, resource tion [19
30] and the Software as a Service (SaaS) layer
allocation and scheduling (RAS), quality of service [31
34] RAS. To satisfy objective functions imposed by
(QoS) guarantee, security and privacy have become users, how to allocate resources to workflow tasks
research hotspots, which are very challenging in cloud [35
43] becomes a hot topic in cloud RAS. The study
computing [1]. In particular, to improve resource utili- directions in this work are boiled down to five aspects:
zation and execution efficiency, save energy, increase (1) locality-aware task scheduling; (2) reliability-aware
the cloud provider’s profits and satisfy the QoS scheduling; (3) energy-aware RAS; (4) SaaS layer RAS;
requirements of users, RAS plays an important role in and (5) workflow scheduling. RAS policies and algo-
cloud computing, and its policy and algorithm have a rithms will be introduced in detail in Section 3.
direct effect on cloud performance and cost.
RAS usually comprises two types: static and dynamic
In the cloud computing environment, computational RAS [44]. The static RAS schedules known resources in
resources or services are provided by cloud providers prescient circumstances through adopting pre-schedul-
(e.g. Google, Amazon and Microsoft), and are accessed ing method, whereas dynamic RAS depends on both
as general utilities that can be leased and re-leased by the foregone resources and the current system state to
users [2]. Cloud providers are responsible for resource make RAS schemes. The advantages of static RAS are
management, storage and guarantee of reliability, while simple implementation, low running load and high effi-
users only need to pay for their use in executing tasks ciency. Although dynamic RAS has many uncertain fac-
[3]. To satisfy user requirements and maximize profit, tors and a relatively high scheduling overload, it can
cloud providers have to take into account cloud perfor- adapt to a varied environment and assign resources
mance and cost during RAS in cloud computing. Hence, flexibly during program execution. Furthermore, in
many researchers have been attracted to study cloud

4 IETE TECHNICAL REVIEW | VOL 31 | NO 1 | JAN
FEB 2014

 Ma T, et al.: Resource Allocation and Scheduling in Cloud Computing: Policy and Algorithm

some algorithms, both are often used simultaneously. challenge. For example, MapReduce [46] can automati-
Therefore, in our discussion, we will not distinguish cally handle failures: if a node crashes, MapReduce can
whether the algorithm is static or dynamic. re-run its tasks on another node. There are some
research works [15
18] that consider reliability.
The rest of the paper is organized as follows. In Sec-
tion 2, we present five hot topics of cloud RAS and Problem 3: Energy-aware RAS problem
how to
describe them in detail. In Section 3, we provide a reduce energy consumption of data centres to mini-
detailed classified discussion on the various RAS poli- mize operating costs of cloud providers.
cies and algorithms according to the five problems.
Comparative analysis of the RAS policies and algo- With the demand for high-performance computing
rithms of the five problems is introduced in Section 4. infrastructures growing dramatically, the energy con-
Finally, we draw a conclusion and discuss future sumption of data centres is increasing and has therefore
research directions in Section 5. become an important research issue. High energy con-
sumption not only contributes to high operational costs,
2. PROBLEMS AND DESCRIPTION which will reduce cloud providers’ profit, but also
emits substantial carbon dioxide (CO2) which is
In solving the problem of RAS in cloud computing, unfriendly to the environment. Therefore, there is an
researchers have performed significant numbers of urgent need to research the problem of energy-aware
studies from various aspects. We can sum up five hot RAS to reduce energy consumption, while maintaining
topics of cloud RAS from these researches: (1) locality- high SLAs between users and cloud platform providers.
Downloaded by [78.56.45.90] at 00:21 18 June 2014

aware task scheduling; (2) reliability-aware schedul- Many studies [19
30] have been performed to find
ing; (3) energy-aware RAS; (4) SaaS layer RAS; and (5) energy-efficient RAS approaches to save energy and
workflow scheduling. In this section, we discuss the reduce the usage of power in cloud computing.
five topics in detail.
Problem 4: SaaS layer RAS problem
how to opti-
Problem 1: Locality-aware task scheduling problem
mize RAS in SaaS layer to maximize profits of SaaS
how to exploit data locality in RAS of cloud to improve providers.
execution efficiency and save network bandwidth.
Services in the cloud can be categorized as: Infrastruc-
In cloud computing, one of Hadoop’s basic principles ture as a Service (IaaS), Platform as a Service (PaaS)
is “moving computation is cheaper than moving data” and SaaS [47]. SaaS is a software delivery model pro-
[45], which indicates that migrating computation viding software service to customers at a low cost
closer to the data source is better than moving data to across the Internet independently without supplying
where the application is run. Furthermore, network the underlying hosting infrastructure [48]. SaaS pro-
bandwidth is a scarce resource, so in order to minimize viders obtain profits from the margin between the
network congestion and increase the overall through- operational cost of infrastructure and the revenue gen-
put of systems, it is necessary to research the problem erated from customers. Therefore, researching the
of locality-aware task scheduling to enhance data problem of SaaS layer RAS to minimize the overall
locality between jobs. In general, good data locality infrastructure cost without adverse influence to the
means that a computation task is located near the data customer for SaaS providers has become an urgent
source. There are many researchers [4
14] attempting problem to be solved in the cloud computing environ-
to achieve good data locality as much as possible. ment. There are some other research works [31
34]
that focus on maximizing the profits of SaaS providers.
Problem 2: Reliability-aware scheduling problem

how to reduce failure rate of tasks in RAS of cloud, to Problem 5: Workflow scheduling problem
how to
improve the reliability and execution efficiency of optimize workflow scheduling to trade off time and
cloud system. cost.

In the cloud environment with thousands of data Workflows constitute a common model for describing
nodes, resource failures are inevitable, which may lead a wide range of applications in distributed systems
to execution abort, data corruption and loss, perfor- [49]. Their scheduling methods try to minimize the
mance degradation, service level agreement (SLA) execution time (makespan) of the workflows in distrib-
violation, and consequently devastating loss of uted systems like Grids. However, in clouds, there is
customers. Therefore, researching the problem of another important parameter than execution time, that
reliability-aware scheduling to improve the reliability is, economic cost [50]. In general, the faster resources
of a cloud computing environment is a crucial are, the more they cost. Therefore, researching the

IETE TECHNICAL REVIEW | VOL 31 | NO 1 | JAN
FEB 2014 5

 Ma T, et al.: Resource Allocation and Scheduling in Cloud Computing: Policy and Algorithm

RAS, cost-based RAS and performance- and cost-based

RAS. In addition, the locality-aware task scheduling
and the reliability-aware scheduling are attributed to
performance-based RAS because they try to improve
tasks’ execution efficiency and the reliability of the
cloud system, respectively. The energy-aware RAS
and SaaS-layer RAS are attributed to cost-based RAS
because they attempt to minimize operating costs and
maximize service provider’s profits while meeting the
QoS requirement of users. The workflow scheduling
Figure 1: Example of workflow. belongs to performance- and cost-based RAS because
it attempts to optimize cost and time simultaneously.

problem of workflow scheduling to select time
cost 3.1 Performance-based RAS

tradeoffs is a significant project.
Performance-based RAS policies and algorithms are
A workflow application is commonly denoted by a classified into locality-aware task scheduling and
directed acyclic graph, G ¼ (T, E), where T is a set of n reliability-aware RAS. Most researches on locality-
tasks (t1, t2, . . . , tn), and E is a set of dependencies aware task scheduling focus on solving the Problem 1,
between tasks. Each dependency, ei,j ¼ (ti, tj) 2 E, and researches about reliability-aware RAS focus on
Downloaded by [78.56.45.90] at 00:21 18 June 2014

where ti is the parent task of tj and tj is the child task of solving the Problem 2.
ti, represents a precedence constraint, which indicates
that a child task cannot be executed until all of its par- 3.1.1 Locality-aware Task Scheduling
ent tasks have been completed. Labels on tasks repre-
sent computation costs, for example, number of Mass data processing platform (e.g. MapReduce [46],
instructions; labels on dependencies represent commu- Dryad [51], Hadoop [52], etc.) of cloud computing
nication cost, for example, bytes to transmit. An exam- requires the concurrent execution of data-intensive
ple of workflow is shown in Figure 1, and the main parallel jobs (each job consists of several subtasks). In
objective of the workflow scheduler is to decide where general, jobs will compete for computing resources
each task component of the directed acyclic graph will and network bandwidth, so in order to improve execu-
execute. tion efficiency through decreasing network transmi-
ssion during the execution of jobs, some researchers
have indicated that tasks should be assigned to the
3. RAS POLICIES AND ALGORITHMS node with data as far as possible to improve data local-
ity [4
14].
In this section, according to different optimization
goals, RAS algorithms and policies can be classified There are many improved methods, from improving
into three parts (see Figure 2): performance-based data locality to a tradeoff between data locality and

Figure 2: A taxonomy of RAS algorithms and policies.

6 IETE TECHNICAL REVIEW | VOL 31 | NO 1 | JAN
FEB 2014

 Ma T, et al.: Resource Allocation and Scheduling in Cloud Computing: Policy and Algorithm

fairness, proposed based on the Hadoop schedulers, network environment is poor and the cluster is over-
thus, locality-aware task scheduling policies and algo- loaded, Balance-Reduce tries to enhance data locality
rithms are described from three perspectives: the and decreases data locality to increase optimization. In
Hadoop scheduler, improving data locality and com- addition, taking into account the job character and
bining data locality and fairness. data locality while killing tasks, which is lacking in
Fair Scheduler, Tao et al. [6] proposed an improved
3.1.1.1 Perspective of Hadoop Scheduler. Hadoop has Fair Scheduler, which adopts different policies for
several job schedulers, such as the FIFO Scheduler I/O-bound and CPU-bound jobs based on data locality
[53], the Hadoop on Demand Scheduler [54], the Fair to decrease the data movement and speed up the exe-
Scheduler [55, 56] and the Capacity Scheduler [57]. cution of jobs. Furthermore, Seo et al. [7], Zhang et al.
The First In First Out (FIFO) Scheduler, the Hadoop [8, 9] and Hammoud et al. [10] tried to improve
default scheduler, greedily searches for a data-local MapReduce performance through enhancing the data
task in the head-of-line job and allocates it to an idle locality of map tasks and reduce tasks respectively.
server. Although the policy is easy to implement, it Seo et al. [7] built the High Performance MapReduce
does not consider global optimization, so it cannot Engine (HPMR), which suggests inspecting input
guarantee higher data locality and should increase splits in the map phase and predicts which key-value
completion time of small jobs. The Hadoop on pairs to be partitioned in the reduce phase. HPMR
Demand Scheduler solves the drawback of poor assigns the expected data to map tasks near the future
response times for short jobs in the presence of large reducers, which leads to the time of job execution not
jobs in the FIFO scheduler by letting users share a com- changing with network bandwidth. Zhang et al. [8]
Downloaded by [78.56.45.90] at 00:21 18 June 2014

mon file system while owning private MapReduce proposed a next-k-node scheduling (NKS) method
clusters. However, it has two problems: poor locality based on a homogeneous computing environment to
and poor utilization. To solve these problems, Fair improve the data locality of map tasks. Afterwards,
Scheduler gives each user the illusion of owning a pri- Zhang et al. [9] introduced an original data locality
vate cluster, and dynamically redistributes capacity aware scheduling method based on the NKS method
unused by some users to other users. However, for MapReduce framework in heterogeneous environ-
although trying to minimize wasted computation, it ments. Contrary to HPMR, Locality-Aware Reduce
does not consider the job character and data locality, Task Scheduler, proposed by Hammoud et al. [10]
which increases the network traffic and data transfer attempts to collocate reduce tasks with the maximum
while rescheduling the killed Map/Reduce tasks. The required data computed after recognizing input data
Capacity Scheduler, developed by Yahoo, is similar to network locations and sizes to seek a good data
the Fair Scheduler. It is defined for large clusters by locality.
maintaining multiple job queues sorted by job arrival
time, and it provides a minimum capacity guarantee 3.1.1.3 Perspective of Combining Data Locality and
and the sharing of excess capacity among users. Fairness. Workload in cloud computing
includes two kinds of job: real-time and long-term
3.1.1.2 Perspective of Improving Data Locality. To fur- jobs. A real-time job has the characteristics of few sub-
ther achieve data locality, some researchers proposed tasks, short execution time and sensitive response
improved methods and algorithms [4
10] based on time. In contrast, a long-term job’s characteristics are
the existing Hadoop scheduling algorithms. To reach multiple subtasks and long execution time. The Fair-
global optimization and assign tasks efficiently in ness Scheduling algorithm can allocate resources for
Hadoop, Fischer et al. [4] introduced an idealized real-time jobs in time to achieve a fast response. How-
Hadoop model called Hadoop Task Assignment ever, data locality and fairness are in opposition to
(HTA) problem to find the assignment that minimized each other, and it is difficult to satisfy them simulta-
the job completion time, and proved it to be NP- neously in a scheduling policy. Therefore, to further
complete. To solve the problem, they designed a flow- improve job execution efficiency, some researchers
based algorithm called MaxCover-BalAssign, which focus on how to combine data locality and fairness in
works iteratively to produce a sequence of assign- scheduling policies and algorithms. Zaharia et al. [11]
ments and output the best one. Although the solution proposed a delay scheduling algorithm that has been
has been shown to be near optimal, it takes a long time used in the Fair Scheduler and is based on max
min
to deal with a large problem instance owing to its high fairness [58]. The algorithm requires the job scheduled
time complexity. Therefore, Jin et al. [5] proposed a next according to fairness to wait for a small amount
heuristic task scheduling algorithm called Balance- of time; if it is unable to launch a data-local task, it
Reduce, which schedules tasks by taking a global view allows other jobs to execute tasks instead. It is worth
and adjusts task data locality dynamically according to waiting for a local task if assuming servers become
network state and cluster workload. When the idle quickly enough. However, this assumption is too

IETE TECHNICAL REVIEW | VOL 31 | NO 1 | JAN
FEB 2014 7

 Ma T, et al.: Resource Allocation and Scheduling in Cloud Computing: Policy and Algorithm

Table 1: The overview of locality-aware task scheduling policies and algorithms

Hadoop on Fischer et al. [4]/ Zaharia et al. [11]/
Demand [54]/Fair [55,56]/ Jin et al. [5]/ Seo et al. [7]/ Quincy [12]/GCC [13]/
FIFO [53] capacity [57] Tao et al. [6] Zhang et al. [8, 9] Hammoud et al. [10] Jin et al. [14]
p p p p p
Data locality
p p
Fairness
p p p p
Execution efficiency
p p p
Utilization
p
Global optimization
p
Map tasks
p
Reduce tasks

strict, and based on static waiting time threshold, it novel scheduling algorithm based on a kind of trust
cannot adapt to dynamic load of a data centre, so mechanism. These are described in more detail below.
delay scheduling does not work well when servers
slowly free up. Moreover, there are two works similar The performance of Hadoop Scheduler is seriously
to delay scheduling, Quincy [12] and Good Cache limited owing to its use of assumptions of homoge-
Compute (GCC) [13]. Quincy, which can be used for neous cluster nodes and linear progress made by tasks
Downloaded by [78.56.45.90] at 00:21 18 June 2014

Mapreduce and Dryad, maps the scheduling problem to speculate when to re-execute tasks that appear to be
to a graph data structure according to a global cost stragglers. However, in practice, the assumptions can-
model, and solves the problem by a well-known mini- not hold well in a heterogeneous environment charac-
mum-cost flow algorithm. It can achieve better fair- terized by equipment that varies greatly with respect
ness, but it has a negligible effect on improving data to the capacities of computation and communication,
locality. GCC sets a utility threshold that is the upper architectures, memory and power. Therefore, Zaharia
bound of the number of idle servers, and it can skip et al. [15] studied a discovery mechanism of abnormal
servers when the idle server number is below the util- tasks under a heterogeneous environment, and pro-
ity threshold. To address the issue of the delay sched- posed the Longest Approximate Time to End (LATE)
uling, Jin et al. [14] proposed an adaptive waiting time scheduling algorithm based upon the speculative exe-
threshold model and designed an adaptive delay cution scheduler in Hadoop. The LATE algorithm uses
scheduling algorithm (ADS). ADS adjusts jobs’ wait- a Slow Task Threshold, meaning that it is slow enough
ing time threshold dynamically to reduce the job to prevent unnecessary speculation, ranks tasks by
response time, according to the information on idle estimated time remaining and starts a copy of the high-
servers’ arrival intensity, available network band- est ranked task that has a progress rate lower than the
width and job running status. Slow Task Threshold. The advantage of the LATE
algorithm is robustness to node heterogeneity, since
A comparison of the policies and algorithms discussed only some of the slowest (not all) speculative tasks
above has been performed from the aspects of data are restarted. However, the method does not compute
locality, fairness, execution efficiency, utilization, the remaining time for tasks correctly, and cannot find
global optimization and so on. Table 1 shows the over- really slow tasks in the end. To this end, Chen et al.
view of locality-aware task scheduling policies and [16] proposed a Self-Adaptive MapReduce (SAMR)
algorithms. scheduling algorithm, which guarantees robustness
and significantly improves execution efficiency and
resource utilization of MapReduce. Although launch-
3.1.2 Reliability-aware Scheduling
ing backup tasks for slow tasks, Hadoop and LATE
Reliability refers to the probability of failure not cannot find the appropriate tasks that really delay the
appearing during run-time (program execution). In the overall execute time, because the two schedulers
cloud environment with thousands of data nodes, always use a static way of finding slow tasks. In con-
resource is inevitably unreliable, which has a great trast, SAMR incorporates historical information
effect on task execution and scheduling. Therefore, recorded on each node to tune parameters and dynam-
some studies have been performed on the reliability of ically find slow tasks that really need backup tasks. In
RAS. For example, the scheduling algorithm of Chen addition, Kumar et al. [17] also utilized task historical
et al. [16] is relative to Zaharia et al. [15], and has com- values such as past success rate, failure rate of task
mon characteristics, such as historical information, and previous execution time in the proposed algo-
with Kumar et al. [17]. Wang et al. [18] proposed a rithm, which adopts modified linear programming

8 IETE TECHNICAL REVIEW | VOL 31 | NO 1 | JAN
FEB 2014

 Ma T, et al.: Resource Allocation and Scheduling in Cloud Computing: Policy and Algorithm

Table 2: The overview of reliability-aware scheduling policies and algorithms

Zaharia et al. [15] Chen et al. [16] Kumar et al. [17] Wang et al. [18]
p p p p
Reliability
p p p p
Execution efficiency
p
Resource utilization
p
Static
p p p
Dynamic
p p
Historical information

problem transportation-based RAS for decentralized 3.2.1.1 Perspective of Server. The existing techniques
dynamic cloud computing. for energy savings at a server farm [21] can roughly be
divided into two categories: dynamic voltage/fre-
Furthermore, by evaluating the trustworthiness of quency scaling (DVFS) [61] inside a server and
machines in a cloud environment, Wang et al. [18] pro- dynamic power management [62], turning on/off serv-
posed a trust dynamic level scheduling algorithm ers on demand. The former depends on the setting of
named Cloud-DLS, which is based on a kind of trust hardware components to perform scaling tasks [19, 20]
mechanism and integrates the existing DLS algorithm and the latter ensures near-zero electricity consump-
[59, 60] to decrease the failure probability of the task tion through turning off servers [21, 22]. Mezmaz et al.
assignments. [19] and Huang et al. [20] adopted DVFS technique for
Downloaded by [78.56.45.90] at 00:21 18 June 2014

the scheduling of parallel applications to reduce

A comparison of these policies and algorithms dis- energy consumption and minimize makespan. Mez-
cussed above has been made from the aspects of reli- maz et al. [19] proposed a parallel bi-objective hybrid
ability, execution efficiency, resource utilization, static, genetic algorithm that combines a multi-objective par-
dynamic and historical information. Table 2 shows the allel genetic algorithm and energy-conscious schedul-
overview of reliability-aware scheduling policies and ing heuristic (ECS). It can acquire exploration power,
algorithms. intensification capability, a cooperative approach and
parallelism from the genetic algorithm, ECS, the island
model, and the multistart model respectively. Huang
3.2 Cost-based RAS
et al. [20] presented an Enhanced Energy-efficient
Cost-based RAS policies and algorithms are classified Scheduling algorithm, which prioritizes the tasks in
into energy-aware RAS and SaaS-layer RAS. Research order and assigns them to the servers in the mapping
in energy-aware RAS focuses on solving Problem 3, phase, extending the mapping schedule to meet the
and research on SaaS-layer RAS focuses on solving deadline in the slacking phase. Duy et al. [21] and
Problem 4. Mazzucco et al. [22] introduced allocation policies that
turn off unused servers and restart them to minimize
3.2.1 Energy-aware RAS the number of running servers to save energy in the
cloud computing. In order to deal with user demand,
With the rapid growth of computing applications and which is time-varying and difficult to predict with
data, larger servers and disks are needed to process absolute accuracy, the policy of Duy et al. [21] integra-
them fast enough within the required time period, and tes a neural network predictor to predict future load
energy consumption is becoming the prime cost of demand based on historical demand, and the policy of
data centres. Therefore, how to lower the energy usage Mazzucco et al. [22] aims to find the parameters
of data centres has been a challenging and complex leading to the highest revenues according to the
issue [24]. There are many researchers studying energy forecasting.
consumption through RAS of cloud computing, which
can not only help reduce operation costs of data
centres, but also reduce CO2 emissions. 3.2.1.2 Perspective of Virtualization Technique. The key
current technology for energy-efficient operation of
There are some solutions trying to allocate servers servers in data centres is virtualization [63], which
directly, some adopting virtualization techniques, and reduces energy consumption and improves the utiliza-
some achieving energy efficiency across multiple data tion of resources through consolidating multiple vir-
centre locations. Therefore, energy-aware RAS policies tual machine (VMs) instances into a physical server
and algorithms are discussed from three aspects: and guarantees the SLAs between users and cloud
server, virtualization technique and multiple data platform providers by resolving QoS crosstalk between
centres. services and applications. According to management

IETE TECHNICAL REVIEW | VOL 31 | NO 1 | JAN
FEB 2014 9

 Ma T, et al.: Resource Allocation and Scheduling in Cloud Computing: Policy and Algorithm

decisions, VMs that encapsulate virtualized services simple optimization process to maximize the
can be created, copied, moved and deleted. provider’s profit based on an economic model merging
several factors, such as power consumption, SLA, out-
Younge et al. [23] described a Green Cloud Framework sourcing capabilities, virtualization overhead and het-
to optimize energy consumption in a cloud computing erogeneity management, and then VM placement were
environment and considered relative aspects of VM decided based on the optimization result.
scheduling using a new greed-based algorithm to mini-
mize power consumption and VM management using Furthermore, the Credit Scheduler [64], which is the
VM imaging and live migration to reduce the number default VM scheduler of Xen, is a kind of fair-share
of active hosts. However, this neglects SLA violation scheduler that distributes minimum processor time to
penalties when reducing energy consumption. Belogla- a VM according to the proportion of selected VM
zov et al. [24
26] defined an architectural framework credit value to the sum of all VM credit values. How-
and principles for energy-efficient cloud computing ever, processor time is not a satisfactory criterion for
and divided energy optimization approaches into two estimating energy consumption accurately by a VM
sub-phases: “VM Selection” and “VM Placement”. and there are some in-processor events that affect pro-
Based on this architecture, Beloglazov and Buyya [24] cessor energy consumption more significantly. There-
proposed three stages of VM placement optimization: fore, Kim et al. [28, 29] suggested an estimating energy
optimizing the utilization of resources to minimize the consumption model that calculates the amount of con-
number of physical nodes in use, optimizing the sumption of a VM via monitoring of processor perfor-
network to minimize the overhead of data transfer over mance counters, and then, by modifying the Credit
Downloaded by [78.56.45.90] at 00:21 18 June 2014

network and optimizing temperature to minimize cool- Scheduler, they proposed an Energy-Credit Scheduler
ing system operation, and then Beloglazov et al. [25] that schedules VMs according to user-defined budgets
introduced four heuristics that can effectively handle instead of the processor time credits to limit the energy
strict SLA requirements for the “VM selection” sub- consumption rates of VMs.
phase: a Single Threshold policy and three double-
threshold policies like the Minimization of Migrations 3.2.1.3 Perspective of Multiple Data Centers. The solu-
policy, the Highest Potential Growth policy and the tions mentioned above target saving energy within a
Random Choice policy, and the Modified Best-Fit single server or a single data centre (with many serv-
Decreasing algorithm for the “VM Placement” sub- ers) in a single location. Garg et al. [30] proposed some
phase. However, fixed values for the thresholds are not simple, yet effective generic scheduling policies based
suitable for an environment with dynamic and unpre- on various contributing factors such as energy cost,
dictable workloads, and thus Beloglazov and Buyya carbon emission rate, workload and CPU power effi-
[26] proposed an approach for dynamic consolidation ciency to achieve energy efficiency across multiple
of VMs based on adaptive utilization thresholds, which data centre locations.
further reduces SLA violation and VM migrations.
A comparison of these policies and algorithms dis-
Moreover, Goiri et al. [27] adopted an economic cussed above has been made from the aspects of
approach to solve the energy-efficient and multiface- energy-aware RAS, server, virtualization, multiple data
ted resource management problem. They performed a centres, application, SLA and QoS and so on. Table 3

Table 3: The overview of energy-aware RAS policies and algorithms

Mezmaz Huang Duy Mazzucco Younge Beloglazov Goiri Kim Garg
et al. [19] et al. [20] et al. [21] et al. [22] et al. [23] et al. [24
26] et al. [27] et al. [28, 29] et al. [30]
p p p p p p p p p
Energy-aware RAS
p p p p p p p p p
Server
p p p p
Virtualization
p
Multiple data centres
p p
Application
p p p p
SLA and QoS
p p p
Processor
p
Network
p
Cooling
p
Outsourcing
p p
Economic modelling
p
CO2 emission

10 IETE TECHNICAL REVIEW | VOL 31 | NO 1 | JAN
FEB 2014

 Ma T, et al.: Resource Allocation and Scheduling in Cloud Computing: Policy and Algorithm

shows the overview of energy-aware RAS policies and Table 4: The overview of SaaS-layer RAS policies and
algorithms. algorithms
Wu et al. Li et al. Espadas et al.
3.2.2 SaaS Layer RAS [31, 32] [33] [34]
p p p
There are several studies that focus on the SaaS layer, Cost-effective
p p p
where the SaaS provider (such as Saleforce.com) leases Profit of SaaS provider
p
Profit of cloud provider
resources from cloud providers (such as Amazon EC2) p
QoS of users
and provides services to SaaS users with the goal of min- p
Tenant-based
imizing the payment and maximizing the profit while p
Improved workload distribution
guaranteeing meeting QoS requirement. However, less
research has been performed compared with the other
four, which implies that there is a large space for further
study. Detailed descriptions of SaaS-layer RAS policies about workflow scheduling focuses on solving the
and algorithms have been made as follows. Problem 5.

Wu et al. [31, 32] proposed three cost-effective admis- 3.3.1 Workflow Scheduling
sion control and scheduling algorithms: minimizing
the number of VMs, rescheduling and exploiting the Workflow scheduling is a kind of global task schedul-
penalty delay, for SaaS providers to maximize profit ing, because it focuses on mapping each task to a suit-
able resource and ordering the tasks on each resource
Downloaded by [78.56.45.90] at 00:21 18 June 2014

through minimizing infrastructure cost and maximiz-

ing customer satisfaction level. They, however, focused to satisfy some performance criterion. Workflow
mainly on maximizing the profit of the SaaS provider, scheduling strategy can be classified into two forms:
without considering optimization of Cloud providers best-effort-based scheduling strategy and QoS-
and SaaS users. Li et al. [33] presented an optimal cloud constraint-based scheduling strategy [65]. The best-
resource provisioning algorithm, which is beneficial for effort-based workflow scheduling strategy, which is
the SaaS users, SaaS provider and cloud provider. The used by cluster and grid whose resources are free to
algorithm includes a resource level optimization algo- access and share among community members, only
rithm for the SaaS provider and cloud provider and attempts to minimize the makespan (overall comple-
application level optimization algorithm for the SaaS tion time) of workflow applications without consider-
provider and SaaS users, which maximizes the profit of ing monetary cost and QoS requirements of users. In
the SaaS provider and cloud provider and satisfies the contrast, the QoS-constraint-based workflow schedul-
SaaS user’s QoS needs. In addition, SaaS providers may ing strategy, which is used by cloud with a market-
be charged for global resource usage, including unused oriented business model, attempts to satisfy the user-
idle processes and resources without considering specified QoS requirements (mainly on deadline, cost
resources used by each tenant. Therefore, Espadas et al. and reliability) as well as guarantee the best perfor-
[34] established a formal measure for over- and under- mance of a workflow scheduling system (e.g. satisfy-
utilization of virtualized resources in cloud infrastruc- ing user’s deadline-constrained requirement as well as
tures and proposed a tenant-based resource allocation minimizing the cost of workflow execution and satisfy-
model comprising three approaches
tenant isolation, ing the user’s budget-constrained requirement as well
VM instance allocation and load balancing
to as minimizing the time of workflow execution).
improve workload distribution when deploying SaaS
applications over cloud computing platforms by con- According to different QoS constraints, such as deadline-
sidering their multitenancy. constraint, budget-constraint and multiple QoS-con-
straints, workflow scheduling policies and algorithms
A comparison of these policies and algorithms has can be divided into three parts for discussion. Regard-
been made from the aspects of cost-effectiveness, profit ing the deadline-constrained requirement, algorithms
of SaaS provider, profit of cloud provider, QoS of are presented from the methods of heuristic and meta-
users, tenant-based and improved workload distribu- heuristic respectively.
tion. Table 4 shows the overview of SaaS-layer RAS
policies and algorithms. 3.3.1.1 Perspective of Deadline-constrained Requirement.
There is some research focusing on the problem of
3.3 Performance and Cost-based RAS satisfying the deadline-constrained requirement of the
user as well as minimizing the cost of workflow
Performance and cost-based RAS policies and algo- execution. In addition, since task scheduling is a well-
rithms mainly relate to workflow scheduling. Research known NP-complete problem [66], many heuristic or

IETE TECHNICAL REVIEW | VOL 31 | NO 1 | JAN
FEB 2014 11

 Ma T, et al.: Resource Allocation and Scheduling in Cloud Computing: Policy and Algorithm

meta-heuristic methods have been proposed for cloud deal with the optimization of the Task-to-VM
workflow scheduling. assignments.

The heuristic-based approach is to develop a schedul- 3.3.1.2 Perspective of Budget-constrained Requirement.

ing algorithm that fits only a particular type of prob- There is only one study focusing on satisfying the
lem. Bittencourt et al. [35] proposed a Hybrid Cloud user’s budget-constrained requirement as well as mini-
Optimized Cost scheduling strategy to optimize the mizing the time of workflow execution. Zeng et al. [41]
monetary execution costs through deciding which addressed the problem of scheduling many-task work-
resources should be leased from the public cloud and flow on clouds and presented a budget-conscious
aggregated to the private cloud to provide sufficient scheduling algorithm used by both cloud service pro-
processing power to execute a workflow within a user- viders and users to improve the overall makespan and
defined deadline. Abrishami et al. [36] extended their resource utilization of many-task workflow applica-
previous algorithm: Partial Critical Paths [67], for tions in cloud.
deadline-constrained workflow scheduling in a utility
Grid environment, to design a SaaS Cloud Partial Criti- 3.3.1.3 Perspective of Multiple QoS-constrained
cal Path algorithm for workflow scheduling in SaaS Requirements. There are some studies that con-
clouds. The algorithm schedules the partial critical centrate on satisfying the user’s multiple QoS require-
paths recursively, ending at previously scheduled ments as well as guaranteeing the best performance of
tasks. Furthermore, for the IaaS cloud environment, scientific workflow scheduling systems. Oliveira et al.
Abrishami et al. [37] proposed two new algorithms: [42] introduced an adaptive scheduling heuristic,
Downloaded by [78.56.45.90] at 00:21 18 June 2014

IaaS Cloud Partial Critical Paths and IaaS Cloud Partial which is based on a three-objective weighted cost
Critical Paths with Deadline Distribution, which con- model considering execution time, reliability and
sider several main features of the current commercial financial cost, for parallel execution of scientific work-
Clouds, such as on-demand resource provisioning, flows in the cloud to meet the deadline and fit the bud-
homogeneous networks and the pay-per-use pricing get provided by scientists. Varalakshmi et al. [43]
model. Furthermore, Genez et al. [38] presented an adopted the resource discovery algorithm to reduce
integer linear program (ILP) to solve the workflow flooding of information first, and then proposed the
scheduling problem in SaaS or PaaS clouds with two optimal workflow scheduling algorithm to schedule
levels of SLA, and then presented two heuristics, task-based clusters to the resources within users’ mul-
namely BMT and BMEMT, to find feasible integer sol- tiple QoS parameters such as execution time, reliability
utions that minimize the global SaaS infrastructure and monetary cost. This achieves a significant
monetary costs to reach maximum profit and SLA vio- improvement in CPU utilization by reducing the slow-
lations over the relaxed runs of the proposed ILP. down rates and the waiting times.

The meta-heuristic-based approach is to develop a A comparison of these policies and algorithms dis-
scheduling algorithm based on a meta-heuristic cussed above has been made from the aspects of
method (e.g. Genetic Algorithm (GA), Particle Swarm deadline-constraint, budget-constraint, multiple QoS-
Optimization (PSO) and Ant Colony Optimization constraints, heuristic, meta-heuristic, optimized cost
(ACO)), which provides a general solution method for and optimized makespan. Table 5 shows an overview
developing a specific heuristic to satisfy a particular of workflow scheduling policies and algorithms.
kind of problem. Pandey et al. [39] presented a sched-
uling heuristic based on PSO to minimize the overall 4. COMPARATIVE ANALYSIS
cost of executing application workflows, including the
communication cost between resources and the execu- In this section, a comparative analysis of the five prob-
tion cost of compute resources. Wu et al. [40] pro- lems discussed above is presented in Table 6 in terms
posed a market-oriented hierarchical scheduling of several criteria, that is, execution efficiency, cost-
approach, including a service-level scheduling stage effective, reliability, resource utilization and QoS. In
and a task-level scheduling stage to satisfy the QoS addition, two examples of RAS policies and algorithms
constraints for each individual task while minimizing are given for each problem during the comparison.
the whole cost of cloud workflow system. The service- Problem 1 focuses on locality-aware task scheduling
level scheduling adopts the package-based random to optimize execution efficiency, and takes the
scheduling algorithm to deal with the Task-to-Service MaxCover-BalAssign algorithm proposed by Fischer
assignment, and the task-level scheduling adopts the et al. [4] and the ADS algorithm proposed by Jin et al.
ACO-based scheduling algorithm, which can be [14] as comparative examples. Problem 2 focuses on
replaced by a GA and PSO-based scheduling algo- RAS to improve the reliability of cloud computing and
rithm according to the requirement of performance, to improve execution efficiency, and takes the SAMR

12 IETE TECHNICAL REVIEW | VOL 31 | NO 1 | JAN
FEB 2014

 Ma T, et al.: Resource Allocation and Scheduling in Cloud Computing: Policy and Algorithm

Table 5: Overview of workflow scheduling policies and algorithms

Bittencourt Abrishami Genez Pandey Wu Zeng Oliveira Varalakshmi
et al. [35] et al. [36, 37] et al. [38] et al. [39] et al. [40] et al. [41] et al. [42] et al. [43]
p p p p p
Deadline-constraint
p
Budget-constraint
p p
Multiple QoS-constraints
p p p p p p
Heuristic
p p
Meta-heuristic
p p p p p p p
Optimized cost
p p p
Optimized makespan
p p
Scalability
p p
Reliability
p p
Resource utilization
p p
SaaS/IaaS

scheduling algorithm proposed by Chen et al. [16] and saving operational cost and satisfying QoS of users,
Transportation Problem based scheduling algorithm and Problem 5 not only improves execution efficiency,
proposed by Kumar et al. [17] as comparative exam- but also maximizes profits without violating the QoS
ples. Problem 3 focuses on energy-aware scheduling to requirements of users.
Downloaded by [78.56.45.90] at 00:21 18 June 2014

optimize energy consumption and reduce operation

cost, and takes the Modified Best-Fit Decreasing algo- 5. CONCLUSION AND FUTURE
rithm proposed by Beloglazov et al. [24] and an eco-
RESEARCH DIRECTIONS
nomic approach, which is to solve the energy-efficient
resource management problem, proposed by Goiri In this paper, we present five hot topics of RAS in cloud
et al. [27], as comparative examples. Problem 4 focuses computing, and perform a detailed analysis and discus-
on SaaS-layer RAS to maximize the profit of the SaaS sion of various existing RAS policies and algorithms of
provider, and takes an optimal cloud resource provi- the presented problems in terms of different parame-
sioning algorithm, which is beneficial for the SaaS ters. In addition, we make a comparative analysis of
users, SaaS providers and cloud providers, proposed five problems with their representative algorithms.
by Li et al. [33], and a tenant-based resource allocation
model, which improves workload distribution, pro- Future research directions in the field of cloud RAS
posed by Espadas et al. [34], as comparative examples. could be summed up in the following points. First,
Problem 5 focuses on workflow scheduling to tradeoff except for considering data locality in the task schedul-
time and cost, and takes the Hybrid Cloud Optimized ing, we can also take into account code locality [56],
Cost scheduling strategy proposed by Bittencourt et al. which turns writing code into a sequence of separate
[35] and an adaptive scheduling heuristic, which is tasks where one task can be completed before the next
based on a three-objective weighted cost model consid- is undertaken. The problem of maximizing “code local-
ering execution time, reliability and financial cost, pro- ity” is like affinity scheduling in multiprocessor oper-
posed by Oliveira et al. [42], as comparative examples. ating systems, where we attempt to pin a thread to a
Table 6 shows a comparative analysis of RAS policies processor for as long as possible. Second, load balanc-
and algorithms of five problems. We find that Prob- ing [68, 69] is also a significant parameter to improve
lems 1 and 2 aim at improving execution efficiency performance in RAS of cloud computing. Considering
and resource utilization, Problems 3 and 4 aim at load balancing helps to avoid hotspots and improve

Table 6: Comparative analysis of RAS policies and algorithms of five problems

Problem 1 Problem 2 Problem 3 Problem 4 Problem 5

Fischer Jin Chen Kumar Beloglazov Goiri Li Espadas Bittencourt Oliveira

et al. [4] et al. [14] et al. [16] et al. [17] et al. [25] et al. [27] et al. [33] et al. [34] et al. [35] et al. [42]
p p p p p p
Execution efficiency
p p p p p p
Cost-effective
p p p
Reliability
p p p
Resource utilization
p p p p p
QoS

IETE TECHNICAL REVIEW | VOL 31 | NO 1 | JAN
FEB 2014 13

 Ma T, et al.: Resource Allocation and Scheduling in Cloud Computing: Policy and Algorithm

resource utility. Third, RAS of the hybrid cloud [70, 71] European Conference on Computer systems, New York, Apr.
will arouse researchers’ attention with the requirement 2010, pp. 265
78.
of deciding to allocate applications to the public cloud 12. M. Isard, V. Prabhakaran, J. Currey, U. Wieder, K. Talwar, and
A. Goldberg, “Quincy: Fair scheduling for distributed comput-
or to private cloud. This should optimize the total cost ing clusters”, in Proceedings of ACM SIGOPS 22nd symposium
for running all applications with the different charac- on Operating Systems Principles, New York, 2009, pp. 261
76.
teristics between the public cloud and the private 13. I. Raicu, I. T. Foster, Y. Zhao, P. Little, et al., “The quest for scal-
cloud. Fourth, multi-objective RAS has been a popular able support of data-intensive workloads in distributed sys-
tems”, in Proceedings of the 18th ACM International
topic in the field of academic research. The more Symposium on High Performance Distributed Computing, New
parameters are taken into account, the better the per- York, Jun. 2009, pp. 207
16.
formance and lower the cost will have. 14. J. H. Jin, J. Z. Luo, A. B. Song, and F. Dong, “Adaptive delay
scheduling algorithm based on data center load analysis”, J.
Commun., Vol. 32, no. 7, pp. 47
56, Jul. 2011.
Funding 15. M. Zaharia, A. Konwinski, A. D. Joseph, R. Katz, and I. Stoica,
“Improving mapreduce performance in heterogeneous envi-
This work was supported in part by Special Public Sector ronments”, in 8th USENIX Symposium on Operating Systems
Research Program of China (no. GYHY201206030), Qing Lan Design and Implementation (OSDI), San Diego, CA, Dec. 2008,
Project of Jiangsu Province and was also supported by the pp. 29
42.
Priority Academic Program Development (PAPD) of Jiangsu 16. Q. Chen, D. Q. Zhang, M. Y. Guo, Q. N. Deng, and S. Guo,
Higher Education Institutions. “SAMR: A self-adaptive MapReduce scheduling algorithm in
heterogeneous environment”, in 2010 IEEE 10th International
Conference on Computer and Information Technology (CIT),
REFERENCES Bradford, June 2010, pp. 2736
43.
S. K. S. Kumar and P. Balasubramanie, “Dynamic scheduling
Downloaded by [78.56.45.90] at 00:21 18 June 2014

17.
1. J. Z. Luo, J. H. Jin, A. B. Song, and F. Dong, “Cloud comput- for cloud reliability using transportation problem”, J. Computer
ing: Architecture and key technologies”, J. Commun., Vol. 32, Sci., Vol. 8, no. 10, pp. 1615
26, Oct. 2012.
no. 7, pp. 3
21, Jul. 2011. 18. W. Wang, G. S. Zeng, D. Z. Tang, and J. Yao, “Cloud-DLS:
2. Q. Zhang, L. Cheng, and R. Boutaba, “Cloud computing: Dynamic trusted scheduling for Cloud computing”, Expert Sys-
State-of-the-art and research challenges”, J. Internet Services tems Applic. Int. J., Vol. 39, no. 3, pp. 2321
9, Feb. 2012.
Applic., Vol. 1, no. 1, pp. 7
18, May 2010. 19. M. Mezmaz, N. Melab, Y. Kessaci, Y.C. Lee, et al., “A parallel bi-
3. Y. J. Wang, W. D. Sun, S. Zhou, and X. Y. Li, “Key technologies objective hybrid metaheuristic for energy-aware scheduling for
of distributed storage for cloud computing”, J. Software, Vol. cloud computing systems”, J. Parallel Distributed Comput.,
23, no. 4, pp. 962
986, Feb. 2012. Vol. 71, no. 11, pp. 1497
508, Nov. 2011.
4. M. J. Fischer, X. Y. Su, and Y. T. Yin, “Assigning tasks for effi- 20. Q. J. Huang, S. Su, J. Li, P. Xu, K. Shuang, and X. Huang,
ciency in Hadoop: Extended abstract”, in 22nd ACM Sympo- “Enhanced energy-efficient scheduling for parallel applications
sium on Parallelism in Algorithms and Architectures, New York, in Cloud”, in 12th IEEE/ACM International Symposium on Clus-
2010, pp. 30
9. ter, Cloud and Grid Computing (CCGrid), Ottawa, May 2012,
5. J. H. Jin, J. Z. Luo, A. B. Song, F. Dong, and R. Q. Xiong, “Bar: pp. 781
6.
An efficient data locality driven task scheduling algorithm for 21. T. V. T. Duy, Y. Sato, and Y. Inoguchi, “Performance evaluation
cloud computing”, in 11th IEEE/ACM International Symposium of a green scheduling algorithm for energy savings in cloud
on Cluster, Cloud and Grid Computing (CCGrid), Newport computing”, in 2010 IEEE International Symposium on Parallel
Beach, USA, May 2011, pp. 295
304. and Distributed Processing, Workshops and Phd Forum
6. Y. C. Tao, Q. Zhang, L. Shi, and P. H. Chen, “Job scheduling opti- (IPDPSW), Atlanta, GA, Apr. 2010, pp. 1
8.
mization for multi-user MapReduce Clusters”, in 2011 Fourth 22. M. Mazzucco and D. Dyachuk, “Optimizing cloud providers
International Symposium on Parallel Architectures, Algorithms revenues via energy efficient server allocation”, Sustainable
and Programming, Tianjin, China, Dec. 2011, pp. 213
7. Comput.: Informat. Systems, Vol. 2, no. 1, pp. 1
12, Mar. 2012.
7. S. Seo, I. Jang, K. Woo, I. Kim, J. Kim, and S. Maeng, “HPMR: 23. A. J. Younge, G. von Laszewski, L. Z. Wang, S. Lopez-Alarcon,
Prefetching and pre-shuffling in shared MapReduce computa- and W. Carithers. “Efficient resource management for cloud
tion environment”, in IEEE International Conference on Cluster computing environments”, in 2010 International Green Com-
Computing and Workshops (CLUSTER), New Orleans, LA, puting Conference, Aug. 2010, pp. 357
64.
Sep. 2009, pp. 1
8. 24. A. Beloglazov and R. Buyya, “Energy efficient resource man-
8. X. H. Zhang, Z. Y. Zhong, B. Tu, S. Z. Feng, and J. P. Fan, agement in virtualized cloud data centers”, in 2010 10th IEEE/
“Improving data locality of mapreduce by scheduling in homo- ACM International Conference on Cluster, Cloud and Grid Com-
geneous computing environments”, in Proceedings of IEEE puting (CCGrid), Melbourne, May 2010, pp. 826
31.
9th International Symposium on Parallel and Distributed Proc- 25. A. Beloglazov, J. Abawajy, and R. Buyya, “Energy-aware
essing with Applications, Busan, Korea, May 2011, pp. 120
6. resource allocation heuristics for efficient management of data
9. X. H. Zhang, Y. H. Feng, S. Z. Feng, J. P. Fan, and Z. Ming, “An centers for Cloud computing”, Future Generation Comput.
effective data locality aware task scheduling method for Map- Syst., Vol. 28, no. 5, pp. 755
68, May 2012.
Reduce framework in heterogeneous environments”, in 2011 26. A. Beloglazov and R. Buyya, “Adaptive threshold-based
International Conference on Cloud and Service Computing approach for energy-efficient consolidation of virtual machines
(CSC), Hong Kong, Dec. 2011, pp. 234
42. in cloud data centers”, in Proceedings of the 8th International
10. M. Hammoud, and M. F. Sakr, “Locality-aware reduce task Workshop on Middleware for Grids, Clouds and e-Science,
scheduling for MapReduce”, in 2011 IEEE Third International Nov. 2010, pp. 1
6.
Conference on Cloud Computing Technology and Science 27. I. Goiri, J. Ll. Berral, J. O. Fito , R. Nou, J. Guitart, R.
 , F. Julia
(CloudCom), Athens, Nov. 2011, pp. 570
6. Gavalda , and J. Torres, “Energy-efficient and multifaceted
11. M. Zaharia, D. Borthakur, J. S. Sarma, K. Elmeleegy, et al., resource management for profit-driven virtualized data cen-
“Delay scheduling: A simple technique for achieving locality ters”, Future Generation Comput. Syst., Vol. 28, no. 5,
and fairness in cluster scheduling”, in Proceedings of 5th pp. 718
31, May 2012.

14 IETE TECHNICAL REVIEW | VOL 31 | NO 1 | JAN
FEB 2014

 Ma T, et al.: Resource Allocation and Scheduling in Cloud Computing: Policy and Algorithm

28. N. Kim, J. Cho, E. Seo, “Energy-based accounting and sched- in 2012 IEEE International Systems Conference (SysCon), Van-
uling of virtual machines in a cloud system”, in 2011 IEEE/ACM couver, Mar. 2012, pp. 1
6.
International Conference on Green Computing and Communi- 45. Moving computation is cheaper than moving data. Available:
cations (GreenCom), Sichuan, China, Aug. 2011, pp. 176
81. http://hadoop.apache.org/docs/r0.20.2/hdfs_design.html
29. N. Kim, J. Cho, and E. Seo, “Energy-credit scheduler: An 46. J. Dean, and S. Ghemawat, “MapReduce: Simplified data
energy-aware virtual machine scheduler for cloud systems”, processing on large clusters”, Commun. ACM, Vol. 51, no. 1,
Future Generation Comput. Syst., Vol. 32, pp. 128
37, Mar. pp. 107
13, Jan. 2008.
2014.
47. C. N. Ho€ fer, G. Karagiannis, “Cloud computing services: taxon-
30. S. K. Garg, C. S. Yeo, A. Anandasivam, and R. Buyya, “Envi- omy and comparison”, J. Internet Services Applic., Vol. 2, no.
ronment-conscious scheduling of HPC applications on distrib- 2, pp. 81
94, Sept. 2011.
uted Cloud-oriented data centers”, J. Parallel Distributed
Comput., Vol. 71, no. 6, pp. 732
49, June 2011. 48. R. Buyya, C. S. Yeo, S. Venugopal, J. Broberg, and I. Bran-
dic, “Cloud computing and emerging IT platforms: Vision,
31. L. L. Wu, S. K. Garg, and R. Buyya, “SLA-based resource allo- hype, and reality for delivering computing as the 5th utility”,
cation for software as a service provider (SaaS) in cloud com- Future Generation Comput. Syst., Vol. 25, no. 6, pp.
puting environments”, in 11th IEEE/ACM International 599
616, June 2009.
Symposium on Cluster, Cloud and Grid Computing (CCGrid),
Newport Beach, USA, May 2011, pp. 195
204. 49. E. Deelman, D. Gannon, M. Shields, and I. Taylor, “Workflows
and e-science: An overview of workflow system features and
32. L. L. Wu, S. K. Garg, and R. Buyya, “SLA-based admission con- capabilities”, Future Generation Comput. Syst., Vol. 25, no. 5,
trol for a Software-as-a-Service provider in Cloud computing envi- pp. 528
40, May 2009.
ronments”, J. Comput. Syst. Sci., Vol. 78, no. 5, pp. 1280
99,
Sept. 2012. 50. S. Abrishami, M. Naghibzadeh, and D. H. J. Epema, “Deadline-
constrained workflow scheduling algorithms for infrastructure
33. C. L. Li, L. Y. Li “Optimal resource provisioning for cloud com- as a service clouds”, Future Generation Comput. Syst., Vol. 29,
puting environment”, The J. Supercomputing, Vol. 62, no. 2, no. 1, pp. 158
69, Jan. 2013.
pp. 989
1022, Apr. 2012.
51. M. Isard, M. Budiu, Y. Yu, A. Birrell, and D. Fetterly, “Dryad: Dis-
nez, M. Molina, R. Ramírez, and
Downloaded by [78.56.45.90] at 00:21 18 June 2014

34. J. Espadas, A. Molina, G. Jime tributed data-parallel programs from sequential building
D. Concha, “A tenant-based resource allocation model for scal- blocks”, in Proceedings of the 2007 European Conference
ing Software-as-a-Service applications over cloud computing Computer Systems, Lisbon, pp. 59
72, Mar. 2007.
infrastructures”, Future Generation Comput. Syst., Vol. 29, no.
1, pp. 273
86, Jan. 2013. 52. Hadoop.Available: http://hadoop.apache.org/
35. L. F. Bittencourt and E. R. M. Madeira, “HCOC: A cost optimiza- 53. T. White, Hadoop: The Definitive Guide, 1st edn. O’Reilly
tion algorithm for workflow scheduling in hybrid clouds”, J. Media, 2009.
Internet Serv. Applic., Vol. 2, no.3, pp. 207
27, 2011. 54. Hadoop on Demand Scheduler. Available: http://hadoop.
36. S. Abrishami, M. Naghibzadeh, “Deadline-constrained work- apache.org/docs/r1.1.1/hod_scheduler.html
flow scheduling in software as a service Cloud”, Scientia Ira- 55. Fair Scheduler. Available: http://hadoop.apache.org/docs/
nica, Vol. 19, no. 3, pp. 680
9, June 2012. r1.1.1/fair_scheduler.html
37. S. Abrishami, M. Naghibzadeh, and D. H. J. Epema, “Deadline- 56. M. Zaharia, D. Borthakur, J. Sarma, et al., “Job scheduling for
constrained workflow scheduling algorithms for infrastructure multi-user mapreduce clusters”, EECS Department, University
as a service clouds”, Future Generation Comput. Syst., Vol. 29, of California, Berkeley, Tech. Rep. UCB/EECS-2009-2055,
no. 1, pp. 158
69, Jan. 2013. 2009.
38. T. A. L. Genez, L. F. Bittencourt, E. R. M. Madeira, “Workflow 57. Capacity Scheduler. Available: http://hadoop.apache.org/
scheduling for SaaS/PaaS cloud providers considering two docs/r1.1.1/capacity_scheduler.html
SLA Levels”, in IEEE Network Operations and Management 58. Max
min fairness. Available: http://en.wikipedia.org/wiki/Max-
Symposium (NOMS), Maui, HI, Apr. 2012, pp. 906
12. minfairness
39. S. Pandey, L. Wu, S. M. Guru, and R. Buyya, “A particle swarm 59. A. Dogan and F. Ozg€ u €ner, “Reliable matching and scheduling
optimization-based heuristic for scheduling workflow applica- of precedence-constrained tasks in heterogeneous distributed
tions in cloud computing environments”, in 24th IEEE Interna- computing”, in Proceedings of 29th International Conference
tional Conference on Advanced Information Networking and on Parallel Processing, Toronto, Jun. 2000, pp. 307
14.
Applications (AINA), Perth, Apr. 2010, pp. 400
7. € u
60. A. Dogan and F. Ozg € ner, “Matching and scheduling algo-
40. Z. J. Wu, X. Liu, Z. W. Ni, D. Yuan, and Y. Yang, “A market-ori- rithms for minimizing execution time and failure probability of
ented hierarchical scheduling strategy in cloud workflow sys- applications in heterogeneous computing”, IEEE Trans. Parallel
tems”, J. Supercomputing, Vol. 63, no. 1, pp. 256
93, Jan. Distributed Syst., Vol. 13, no. 3, pp. 308
23, Mar. 2002.
2013.
61. N. B. Rizvandi, J. Taheri, and A. Y. Zomaya, “Some observa-
41. L. F. Zeng, B. Veeravalli, and X. L. Li, “ScaleStar: Budget con- tions on optimal frequency selection in DVFS-based energy
scious scheduling precedence-constrained many-task work- consumption minimization”, J. Parallel Distributed Comput.,
flow applications in cloud”, in 2012 IEEE 26th International Vol. 71, no. 8, pp. 1154
64, Aug. 2011.
Conference on Advanced Information Networking and Applica-
tions (AINA), Fukuoka, Japan, Mar. 2012, pp. 534
41. 62. L. Benini, A. Bogliolo, and G. D. Micheli, “A survey of design
techniques for system-level dynamic power management”,
42. D. Oliveira, K. A. C. S. Ocan ~ a, F. Baia
~o, and M. Mattoso, “A IEEE Trans. Very Large Scale Integration (VLSI) Syst., Vol. 8, no.
Provenance-based Adaptive Scheduling Heuristic for Parallel 3, pp. 299
316, Jun. 2000.
Scientific Workflows in Clouds”, J. grid Computing, Vol. 10, no.
3, pp. 521
52, Sept. 2012. 63. A. Berl, E. Gelenbe, M. D. Girolamo, G. Giuliani, H. D. Meer, M.
Q. Dang, and K. Pentikousis, “Energy-efficient cloud comput-
43. P. Varalakshmi, A. Ramaswamy, A. Balasubramanian, and P. ing”, Comput. J., Vol. 53, no. 7, pp. 1045
51, Jul. 2010.
Vijaykumar, “An optimal workflow based scheduling and
resource allocation in cloud”, in Proceedings of First Interna- 64. E. Ackaouy, “The Xen Credit CPU scheduler”, in Proceedings
tional Conference on Advances in Computing and Communi- of 2006 Fall Xen Summit, Sep. 2006.
cations (ACC), Kochi, India, Jul. 2011, pp. 411
20. 65. J. Yu, R. Buyya, and K. Ramamohanarao, “Workflow schedul-
44. E. M. Mocanu, M. Florea, M. I. Andreica, and N. Tapus, “Cloud ing algorithms for grid computing”, Metaheuristics Schedul.
computing
Task scheduling based on genetic algorithms”, Distributed Comput. Environ. Stud. Comput. Intell., Vol. 146,
pp. 173
214, 2008.

IETE TECHNICAL REVIEW | VOL 31 | NO 1 | JAN
FEB 2014 15

 Ma T, et al.: Resource Allocation and Scheduling in Cloud Computing: Policy and Algorithm

66. M. R. Garey, and D. S. Johnson, Computers and Intractability: a of the 2010 international conference on Web information sys-
Guide to the Theory of NP-Completeness. Philadelphia, PA: W. tems and mining (WISM), Sanya, China, Oct. 2010, pp. 271
7.
H. Freeman, 1979. 70. R. V. D. Bosschce, K. Vanmechelen, and J. Broeckhove,
67. S. Abrishami, M. Naghibzadeh, and D. H. J. Epema, “Cost- “Online cost-efficient scheduling of deadline-constrained work-
driven scheduling of Grid workflows using partial critical loads on hybrid clouds”, Future Generation Comput. Syst., Vol.
paths”, in 2010 11th IEEE/ACM International Conference on 29, no. 4, pp. 973
85, June 2013.
Grid Computing (GRID), Brussels, Oct. 2010, pp. 81
8. 71. R. N. Calheiros, C. Vecchiola, D. Karunamoorthy, and R.
68. D. Babu and P. V. Krishna, “Honey bee behavior inspired load Buyya, “The Aneka platform and QoS-driven resource provi-
balancing of tasks in cloud computing environments”, Appl. sioning for elastic applications on hybrid Clouds”, Future Gen-
Soft Comput., Vol. 13, no. 5, pp. 2292
303, May 2013. eration Comput. Syst., Vol. 28, no. 6, pp. 861
70, June 2012.
69. Y. Q. Fang, F. Wang, and J. W. Ge, “A task scheduling algorithm
based on load balancing in cloud computing”, in Proceedings

Authors
Tinghuai Ma is an associate professor in com- Licheng Zhao is the director-general of the
puter sciences at Nanjing University of Infor- National Meteorological Information Center of
mation Science & Technology, China. He the China Meteorological Administration
received his Bachelor’s (HUST, China, 1997), (CMA), and is a senior engineer on research.
master’s (HUST, China, 2000), PhD (Chinese He received his Bachelor’s degree in com-
Downloaded by [78.56.45.90] at 00:21 18 June 2014

Academy of Science, 2003) and was post-doc- puter science at the Wuhan University, China
toral associate (AJOU University, 2004). From in 1983. He has also minored in MS courses in
November 2007 to July 2008, he visited the meteorology of the Lanzhou University. From
Chinese Meteorology Administration. From July 1983 to July 2006, he worked in the
February 2009 to August 2009, he was a visiting professor in Ubiqui- National Satellite Meteorological Center, later served as the deputy
tous Computing Lab, Kyung Hee University. His research interests director general of the center, deputy chief designer for construction
are in the areas of data mining, privacy protected in ubiquitous sys- projects on the meteorological satellite ground application system,
tem, grid computing, ubiquitous computing, privacy preserving etc. etc. Since October 2009 when assuming his current position, he has
He has published more than 70 journal and conference papers. He is been engaging in system design, construction, organization, opera-
principle investigator of several NSF projects and is a member of tion and management of CMA’s information networking system. He
IEEE. has won a First National Award for Scientific Advance and a CMA Sci-
ence Progress Award. He is also the director of the Basic Meteorologi-
E-mail: [email protected] cal Information Technology Standardization Committee, deputy
Ya Chu received her Bachelor’s degree in director of the Technical Committee for Standardization of Meteoro-
Software engineering from Nanjing University logical Satellites and Space Weather, executive member of the China
Meteorological Society, associate editor-in-chief of the Journal of
of Information Science & Technology, China in
2011. Currently, she is a candidate for a Mas- Applied Meteorology, and member of the editorial board of the Jour-
ter’s degree in meteorological information nal of Remote Sensing.
technology and security in Nanjing University E-mail: [email protected]
of Information Science & Technology. Her
research interest is cloud computing. Otgonbayar Ankhbayar received his Bache-
lor’s degree from School of Mathematics and
E-mail: [email protected] Computer Science, National University of
Mongolia in 2011. Currently, he is a candidate
for a Master in computer science and engi-
neering in Nanjing University of Information
Science & Technology. His research interest is
privacy preservation and data mining.
E-mail: [email protected]

DOI: 10.1080/02564602.2014.890837; Copyright © 2014 by the IETE

16 IETE TECHNICAL REVIEW | VOL 31 | NO 1 | JAN
FEB 2014