A NOVEL JOB SCHEDULING MODEL TO ENHANCE
EFFICIENCY AND OVERALL USER FAIRNESS
OF CLOUD COMPUTING ENVIRONMENT
Wenjuan Li
1, 2
, Xuezeng Pan
1
1
College of Computer Science and Technology, Zhejiang University, Hangzhou, Zhejiang, China
2
Hangzhou Normal University, Hangzhou, Zhejiang, China
Qifei Zhang, Lingdi Ping
College of Computer Science and Technology, Zhejiang University, Hangzhou, Zhejiang, China
Keywords: Cloud computing, Job scheduling model, User fairness, System throughput.
Abstract: Job scheduling and resource allocation are the key factors which affect the performance of parallel and
distributed systems. This paper first analyzed several common used job scheduling models in large and
distributed environment. Since the main purpose of cloud computing is to provide “low-cost” and “on-
demand” services, this paper introduced a novel two-level based cloud scheduling model. In our model, job
scheduling in the cloud is divided into User Scheduling and Task Scheduling which were designed to
implement different scheduling strategies. In the User Scheduling level, it applies the distribution justice
theory in the sociology area to achieve the highest average user fairness. While in Task Scheduling level, it
applies adaptive strategies to achieve the binding of micro tasks and the virtual machines according to
different objectives of tasks. The new model redefined the messaging mechanism between main cloud
entities and a new advertising mode was introduced for cloud providers to promote their services. The
results of emulation experiments show that compared to traditional algorithms the proposed model can
guarantee the user expectations and the relative fairness of the services on the basis of high overall
efficiency of cloud systems.
1 INTRODUCTION
1
Cloud computing is an emerging business
computing model for sharing information
infrastructure. It successfully packages all kinds of
resources belonged to different cloud providers in
different abstract levels as virtualized services over
the network and provides end-users low-cost and on-
demand services. (Foster I. and Kesselman C.,
1999). Obviously, cloud will become one of the
most important service types in the next generation
network market. Currently, however, cloud
computing is faced with many issues needed to be
solved urgently, such as security and trust,
standardization, uniform architecture, programmable
parallel computing and virtualization, etc. (Mladen
1
This project is supported by Chinese Zhejiang Provincial
Research Project (2010C31003).
A. Vouk., 2008).
Resource allocation and task scheduling are the
key factors affecting the efficiency of paralleling
computing. (Guoliang Chen, Hong An and Lin Chen,
2004). They are also the key of cloud researches.
Although scholars have designed many excellent job
scheduling models for traditional distributed and
parallel environment, there still lacks a perfect and
widely-accepted cloud scheduling model. (Fuxi Zhu
and Yanxiang He, 2003). What’s more, in cloud all
kinds of resources are abstracted into the form of
virtual machines. (Foster I., Yong Zhao, Ioan Raicu
and Shiyong Lu, 2008). This specific point
determines that cloud resource allocation and
scheduling mechanisms are quite different from
traditional distributed systems.
This paper proposed a novel two-level based
cloud scheduling model to guarantee the cloud
users’ satisfactory and the total throughput capacity
152
Li W., Pan X., Zhang Q. and Ping L..
A NOVEL JOB SCHEDULING MODEL TO ENHANCE EFFICIENCY AND OVERALL USER FAIRNESS OF CLOUD COMPUTING ENVIRONMENT.
DOI: 10.5220/0003381101520157
In Proceedings of the 1st International Conference on Cloud Computing and Services Science (CLOSER-2011), pages 152-157
ISBN: 978-989-8425-52-2
Copyright
c
 2011 SCITEPRESS (Science and Technology Publications, Lda.)

of cloud systems. It divided cloud scheduling into
two levels: User Scheduling and Task Scheduling
and designed different strategies for them. User
Scheduling can be seen as resource allocation which
is the overall scheduling as the unit of cloud user.
And the target is to achieve the maximum agreement
of user expectation to the quality of service. Task
Scheduling is the scheduling within one user’s tasks.
It realizes the binding of tasks to virtual machines.
And its goal changes with the real requirement of
tasks. Different scheduling strategies are designed
for different scheduling levels. Through message
mechanism, cloud providers can advertise their
services to customers.
This paper was constructed as follows: part 2
describes the related concepts of scheduling and
scheduling model and Part 3 analyzes and compares
several existing task scheduling models. Part 4
introduces the new two-level based scheduling
model. Part 5 is the design, results and analysis of
the simulation experiments and the last part is
conclusion and future work.
2 RELATED CONCEPTS
2.1 Scheduling in Cloud
The goal of cloud scheduling is to select an
appropriate distribution strategy dispatching
paralleling tasks to resource nodes under the premise
of priority constraints and certain performance
indicators and to achieve the minimal total execution
time.
Since all kinds of cloud resources at different
abstract layers are virtualized, scheduling will
undoubtedly involve two aspects: one is the binding
of cloud providers’ actual resources to the virtual
resources and the other is the binding of users’ tasks
to virtual resources.
2.2 Distribution Justice
Distributive justice concerns the fair, just or
equitable distribution of benefits and burdens. It is
widely regarded as an important concept and
influential force in philosophy and the social
sciences. The socialists have found three distinct
principles of justice: the Need Principle, the
Efficiency Principle and the Accountability
Principle. And also justice is context-based.
(Guillermina Jasso, 1989).
In cloud computing environment whose highest
objective is to provide “low-cost” and “on-demand”
services, how to realize the true sense of distribution
justice of users and how to achieve the goal of users’
willing burden in line with their gained services is a
most important issue that cloud scheduling model
has to concern about.
2.3 Berger’s Theory
Berger’s theory comes from the sociology field and
it is about the fair distribution of the society. The
theoretical basis of Berger’s model is “Expectation
States Theory” which makes the model more
reasonable and accurate. Berger’s theory believes
that “balance” is gained through the comparison of
reference structure and local structure. (Berger, J., B.
P. Cohen and M. Zelditch., 1966). It is easy to
extend Berger’s theory into computer field. We
believe that when majority network users receive
considerable quality of service with their expectation,
network balance or distribution justice establishes.
3 SCHEDULING MODELS IN
TRADITIONAL DISTRIBUTED
SYSTEMS [22-24]
3.1 Min-Min Algorithm
Min-Min (Minimum-Minimum Completion Time)
algorithm and Max-Min (Maximum-Minimum
Completion Time) algorithm are both classical task
scheduling algorithms and the research basis of
many of today’s parallel scheduling. The main idea
of Min-Min algorithm is constantly searching for
and scheduling the task which is smallest in the
Minimum Completion Time when deployed in all
cloud resources. Min-Min algorithm has the
advantage of fast execution. But since short tasks are
always prior scheduled, it may easily lead to uneven
host load and low utilization rate. Max-Min has the
opposite characteristics. It solves the problem of
uneven load while with the slower execution speed.
3.2 Ant Colony Optimization
Algorithm
ACO (Ant Colony Optimization) algorithm is one of
the most common used heuristic scheduling
algorithms in distributed systems. (Dorigo M,
Maniezzo V. and Colorni A., 1996). Ant colony
algorithm draws on the principles of ants’ finding
food in nature, uses the history pheromone of each
execution experience to achieve positive feedback
effect and through iteration keeps closer to the
A NOVEL JOB SCHEDULING MODEL TO ENHANCE EFFICIENCY AND OVERALL USER FAIRNESS OF
CLOUD COMPUTING ENVIRONMENT
153

optimal solution. But ACO algorithm is a kind of
bomimetic algorithm of which the implementation
process has somewhat randomness. In addition, due
to the lack of pheromone in the early execution
stage, it takes longer time to find the solution of the
overall problem. What’s more, just like other
evolutionary algorithms, ACO algorithm may be
premature and easy to fall into local optimum.
3.3 Berger’s Model based Algorithm
Chunyan Zhao in her Master’s thesis proposed a
novel cloud scheduling algorithm based on Berger’s
theory. (Chunyan Zhao, 2009). Berger’s Model
based task scheduling algorithm treats the justice
distribution of user tasks as its’ primary goal.
Through the binding of specific user’s tasks to
specific resources, it attempts to achieve the goal of
meeting the greatest consistency of user expectations
and the actual allocation of resources.
3.4 Brief Analysis
Berger’s model based algorithm only takes into
account the distribution relationship between user's
tasks and virtual machines without involving the
resource side. Although all types of resources in
cloud are provided in the form of virtual machines
which appears that job scheduling process is
simplified. But a complete scheduling model needs
consider not only the binding of virtual machines to
users’ tasks but also real resources to virtual
machines. ACO and Min-Min algorithm runs fast
which can guarantee high throughput of cloud
system. However, they ignore user's expectations
and their actual payment capability which may lead
to the consequences that some high expectations,
strong payment capability users obtain low-quality
services, while some users access to expensive and
high quality services beyond their capabilities and
expectations which make them to bear additional
cost.
4 PROPOSED SCHEDULING
MODEL
We proposed a novel scheduling model for cross-
clouds environment. Following is the detail of our
model.
4.1 Realization Mechanism
Figure 1 shows the basic realization framework of
the new model.
The model divides cloud scheduling into two
stages: User Scheduling and Task Scheduling. User
Scheduling refers to the scheduling basing on the
unit of cloud customers. It realizes the binding of
specific customers to specific providers and their
real resources according to the overall QoS(quality
of services) requirements and the current situation of
cloud market. Task Scheduling achieves the binding
of tasks within one cloud user and the virtual
resources that are allocated to the user. Obviously,
User Scheduling stands a higher level than Task
Scheduling. The model sets User Center (UC) and
Cloud Information Service (CIS) Center. Cloud
users register themselves when first login. And then
they login cloud system by the registered account
and submit all of their requirements including
resource requirements, time and cost constraints to
UC each time. Cloud providers register themselves
to CIS center through which they can obtain
customers’ information. Using the user list,
providers can publish resource advertisements to
users.
Figure 1: The Basic Realization Framework of the New
Model.
In the implementation process, a cloud user can
freely choose a provider or a certain set of providers.
Also it can use the more reasonable allocation
solution recommended by User Scheduler which is
obtained according to user expectations and the
current situations of the cloud resources. When the
user is bound with the providers, the providers will
assign a number of virtual resources to complete the
overall work of the user. Next task scheduler will be
on work. In our model, task scheduler is
configurable and its scheduling goal and algorithm
changes with different user requirements. It cares
about which is sensitive to the user tasks completed
time, cost or the other constraints. It will choose a
most suitable allocation strategy binding tasks to
Resource
Node
Resource
Node
Resource
Node
Cloud
Provider
Domain
CIS
Center
Cloud
Provider
User
Center
Register
Register
User
List
Information
Exchange
Register
Agent
Advertise
Service
Request
Transaction
Cloud
Customer
Information
Cooperate
Open Cloud Environment
CLOSER 2011 - International Conference on Cloud Computing and Services Science
154

virtual machines and tasks’ execution order to
achieve the specific user target.
4.2 User Scheduling Algorithm
In the stage of User Scheduling, the new model
references the Berger’s model to deploy its
scheduling strategy to ensure the overall user
fairness. The general process of User Scheduling is
as follows: first of all, to calculate the
comprehensive expectations of the tasks belonged to
one cloud user. And then it clusters cloud resources
of different cloud providers according to the user’s
requirement and calculates different resource set’s
comprehensive capability. At last, it choose the most
suitable resource set to the user according to the
Euclidean Distance from user’s comprehensive
expectation to resource set’s comprehensive
capability.
Algorithm to bind a Certain User to a Certain Set
of Resources belonged to One Provider.
AllocateResourcesForUser(CloudProvider cp,
CloudUser user) {
int requiredCapability=0;
List<Resource> rList=new
ArrayList<Resource>();
int resourceNum=cp.getResourceList().size();
//obtain the number of resources of this
provider
Resource r=null;
rList.clear();
requiredResource=user.
computeRequireCapability ();
//compute the user’s total resource
requirements.
While(requiredResourceCapability>0&&resource
Num>0){
r=findResource(requiredResourceCapability,cp
.getResourceList())
//find the resource with the Minimum
Euclidean Distance to user’s requirement.
rList.add(r);
//add the chosen resource into the resource
list.
requiredResource-=r.capability;
//required capability minus obtained
capability from the chosen resource
resourceNum--;
}
Return rList;
}
4.3 Task Scheduling Algorithm
In the new model, Task Scheduling strategy is
configurable. The general process of Task
Scheduling is as follows: first of all, to judge the
type of the task whether it is “time-sensitive” or
“cost-sensitive”. And then it chooses an appropriate
strategy on the basis of the task’s category.
Algorithm to bind a Certain Task to a Certain
Virtual Machine.
allocateVMForTask(List<VirtualMachine>vmList,
List<Cloudlet> cloudletlist){
while(cloudletList.iterator().hasNext()){
Cloudlet c1 = (Cloudlet) iter.next();
int tasktype=c1.getClassType();
//obtain the type of the task
int virtualmachineID=0;
switch(tasktype){
// choose appropriate virtual machine in vmList
to the task according to its type.
case 1:
vmID=this.findVMforTimeCaredTask(vmList);
c1.setVmId(vmID);
break;
case 2:
vmID=this. findVMforCostCaredTask (vmList);
c1.setVmId(vmID);
break;
case 3:
…
}
}
5 SIMULATION EXPERIMENTS
5.1 The Design of Simulation
Experiments
This paper designed simulation experiments to test
the performances of the new scheduling model, Min-
Min and Berger’s Model based scheduling
algorithm. Our simulation experiments only took
into account “time-sensitive” tasks, that is, cloud
users cared only about their tasks’ completion time.
The simulation experiments set up three evaluation
factors: task sets’ final completion time, total
implementation Cost and the overall customer
satisfaction evaluation. Implementation cost means
the execution cost to complete users’ tasks. It is
calculated based on pre-specified unit costs of
resources and the consumption time. Overall
customer satisfaction evaluation means the average
fairness of all tasks. Since our experiments only
considered “computation-sensitive” tasks, the
overall customer satisfaction is evaluated by tasks’
actual CPU occupation time and their expectation
CPU time. Formula 1 below shows the calculation
method.
log( )
exp
actualCPUTime
Jvalue
ectationCPUTime
=
(1)
The simulation experiments simulated cross-clouds
system that contains one CIS center, one UC(user
center), two cloud providers and a certain number of
cloud users. Each provider contains random number
of resource nodes. Each customer should complete a
A NOVEL JOB SCHEDULING MODEL TO ENHANCE EFFICIENCY AND OVERALL USER FAIRNESS OF
CLOUD COMPUTING ENVIRONMENT
155

certain number of cloud tasks on resource nodes.
According to the specified task number, a function
generates cloud task and randomly deploy each
task’s parameters. In the initial time, providers
register themselves to CIS center and customers
register themselves to UC. CIS center and UC can
often exchange data. When cloud customers want to
finish their tasks, they will first send a message to
UC and UC will help user to find an appropriate
provider set. When transaction relationship is
successfully established, user scheduler will help to
choose the appropriate resource list for the customer.
And task scheduler will work to bind the specific
task to the specific virtual machine.
5.2 Experiments’ Results and Analysis
The following three figures (Figure2-Figure4) show
the results of the simulation experiments.
Figure 2: Comparison of task set’s final completion time.
Figure 3: Comparison of task set’s total implementation
cost.
The results show that when the total number of
tasks is not significant, the Berger’s Model has some
advantages in the customer satisfaction evaluation
for the reason that its original purpose is to ensure
the overall fairness of tasks. However, the pre-
processing cost spent on generating scheduling
sequence is very large which leads to the rapid
increase on execution time when the number of tasks
increases. And the proposed model can ensure high
transaction throughput and relatively low
implementation cost on the basis of high overall
customer fairness compared to traditional scheduling
models in cross-domain environment.
Figure 4: Comparison of task set’s overall customer
satisfaction evaluation.
6 CONCLUSIONS
AND FUTURE WORK
This paper introduces a novel two-level based
scheduling model that can be used in cross-clouds
environment. We divided cloud scheduling into two
different stages: User Scheduling and Task
Scheduling. We set up a User Center (UC) to
manage cloud customers’ information. In the stage
of User Scheduling, the model tries to achieve the
target of highest user fairness and satisfaction which
meets the goal of cloud computing. In the stage of
Task Scheduling, its strategies can be justified
according to actual task requirement which can
increase the total throughput of cloud system.
Simulation experiments show the proposed model
can establish transaction relationship between cloud
customer and provider and gain relatively good
performance.
In future, there are still a lot of issues to be
studied. For example, how the two scheduling levels
interact, how to cluster resources, etc. We will
establish a cross-clouds prototype system and
implement and perfect the proposed model in the
test-bed.
CLOSER 2011 - International Conference on Cloud Computing and Services Science
156

REFERENCES
Foster, I., Kesselman, C., 1999. The Grid: Blueprint for a
New Computing Infrastructure. Morgan Kaufmann,
San Francisco.
Mladen A. Vouk., 2008. Cloud Computing-Issues
Research and Implementations, In ITI 2008, the 30th
International Conference on Information Technology
Interfaces. IEEE.
Guoliang Chen, Hong An, Lin Chen, 2004. Practice of
Parallel Algorithm. Higher Education Press. Beijing.
1st edition.
Fuxi Zhu, Yanxiang He, 2003. Parallel Scheduling
Algorithm in Distributed Computing Theory and
Design. Wuhang University Press. Wuhang. 1st
edition.
Foster I., Yong Zhao, Ioan Raicu, Shiyong Lu, 2008.
Cloud Computing and Grid Computing 360-Degree
Compared. In GCE '08 (2008), the Grid Computing
Environments Workshop. IEEE.
Guillermina Jasso, 1989. The Theory of the Distributive-
Justice Force in Human Affairs: Analyzing the Three
Central Questions. Sociological Theories in Progress:
New Formulation. Sage Pub. Newbury Park, Calif.
Berger, J., B. P. Cohen, M. Zelditch, 1966. Status
Characteristics and Expectation States. Sociological
Theories in Progress (Volume One). Houghton-Miffin.
Boston.
Dorigo M, Maniezzo V. and Colorni A., 1996. Ant
system: Optimization by a colony of cooperating
agent. Man and Cybemetics-Part B: Cybemetics,
26(1). IEEE Trans on Systems.
Chunyan Zhao, 2009. Research and Realization of Job
Scheduling Algorithm in Cloud Environment. Beijing
Jiaotong University Master’s Thesis, Beijing.
A NOVEL JOB SCHEDULING MODEL TO ENHANCE EFFICIENCY AND OVERALL USER FAIRNESS OF
CLOUD COMPUTING ENVIRONMENT
157