International Journal of Computer Applications (0975 – 8887)

Volume 104 – No 17, October 2014

Comparison of Dynamic Load Balancing Policies in Data

Centers
Sunil Kumar Manish Kumar Pandey
Department of Computer Science, Faculty of Department of Computer Science, Faculty of
Science, Banaras Hindu University, Varanasi- Science, Banaras Hindu University, Varanasi-
221005, Uttar Pradesh, India. 221005, Uttar Pradesh, India.

ABSTRACT and AppExchange. In August 2006 (Amazon EC2), May 2007

Cloud computing is an emerging advanced technology which (Face book Platform), April 2008 (Google App Engine) and
provides the computing facilities through the Internet to end October 2008 (Microsoft Azure Platform) were launched in
users by supplying an on demand basis as per usage like market. These are the most popular examples of cloud
water, electricity etc. This could be considered as 5th utility of computing. Main technologies behind cloud computing are
human need in this new cloud era. Earlier computing SOA, Virtualization and Web 3.0 which were basically
technique facilities were developed by the developer at the introduced towards the end of year 2007.
organization and purchased by users and deployed as per the Cloud Computing has attracted a lot of attention in recent
compatibility of applications in available infrastructure which times. In May 2008, Merrill Lynch [4] estimated the cost
usually is not sufficient, e.g. it was not able to handle the advantages of Cloud Computing to be three to five times for
demand at peak traffic period. Again, servers were not fully business applications and more than five times for consumer
utilized as peak traffic only happens in some period of time. applications. According to a Gartner press [4] release from
All big organizations are paying their attention to utilize their June 2008, Cloud Computing will be “no less influential than
server/infrastructures capability through cloud e-business”. The positive attitude towards the importance and
implementation. This technology can sort out the problem at influence of Cloud Computing resulted in optimistic Cloud-
both levels, i.e. at service provider level as well as at user related market forecasts. In October 2008, IDC [5][6]
level by facilitating more infrastructures and less forecasted an almost threefold growth of spending on Cloud
infrastructure respectively by providing the cloud-service- services until 2012, reaching $42 billion. The current trends
provider (CSP) on pay-as-you-go (PAYG) model. Next of cloud computing research is going towards energy efficient
problem is cost as per requirement of Virtual Machines green IT or computing [7] which will ultimately reduce the
(VMs) on Data Center and the response time. This paper overall cost for the end users defined by same firm IDC.
investigates the optimized synchronization between DCs
(Data Centers) and UBs (User Bases) for enhancing the Buyya et al. [7] have defined “pay as you use” model for
application performance and response time for the same cost cloud computing as: Cloud computing is a type of parallel
to the vendors and end users by using a tool called and distributed computing system. It is an extension of grid
CloudAnalyst. The reliability of cloud computing is and utility computing that are collection of federated networks
maintained through load balancing of VMs on a Data Center. of multiple clouds and virtualized computers which are
Load balancing technique improves the performance of VMs, dynamically provisioned and integrated computing resources
reduces overall response time, processing time and cost of based on service-level-agreements (SLAs) established
VMs. In this paper a comparative study was performed among through agreements between the service vendors and end
available service broker policies (Closest Data Center, users.
Optimize Response Time and Reconfigure Dynamically with IT resources like hardware, software and services that are
Load) and available load balancing algorithms (Round Robin, abstracted from the underlying infrastructure or fabric layer of
Equally Spread Current Execution Load, Throttled etc.). The hardware and provided “on-demand” basis and “at-scale-in or
objective of this study is to analyze how these policies help to out” in a multitenant environment based on Service level
coordinate between Data Centers to optimize the applications Agreement (SLA)– CISCO [8].
performance and the cost to the user.
Cloud computing is a platform for enabling ubiquitous,
Keywords convenient, on-demand network access to a shared group of
Internet applications, CSP, Load-balancing, Service broker, configurable computing resources (e.g., networks, servers,
Reliability storage, applications, and services) that can be swiftly
provisioned and unconstrained with negligible management
1. INTRODUCTION endeavor or service provider interaction. This cloud model is
In 1969, Leonard Kleinrock, one of the chief scientists of the a collection of five essential characteristics (such as
original Advanced Research Projects Agency Network virtualization, scalability, on demand services, abstraction and
(ARPANET) which seeded the Internet, said: “As of now, Service Level Agreements), three service models (such as
computer networks are still in their infancy, but as they grow SaaS, PaaS and IaaS), and four deployment models (such as
up and become sophisticated, I will probably see the spread of private, public, hybrid and federated) – NIST [9].
“computer utilities” which, like present electric and telephone
utilities, will service individual homes and offices across the The service scheduling algorithm establishes to handle some
country [1][2].” The cloud computing was evolved from fairness constraints. The first constraint is to categorize user
distributed computing, cluster computing, grid computing, applications by QoS preferences parameters such as execution
utility computing and software [3]. In March 2006 various time, size and cost of the expected applications on the
applications were introduced such as Amazon S3, Sales Force internet, and establish the general functions in accordance

9
 International Journal of Computer Applications (0975 – 8887)
Volume 104 – No 17, October 2014

with the categorization of tasks to contain the fairness of the of the best policy among them and finding out one optimal
resources in matching process. The second constraint is to algorithm which is most profitable to the user as well as
define fair resource providence justified with some functions provider.
of the resource allocation.
2. ARCHITECTURE OF CLOUD
There are a number of simulation tools available in the market
such as CloudSim [10], GridSim [11], CloudReports [12] and ANALYST
CloudAnalyst [13]. Cloudsim and Gridsim are programming CloudAnalyst [13] a cloud simulation tool which takes apart
based simulation tools. One can make changes in these and the simulation, experimental set up exercise from a
get outputs of simulated values such as no. of VMs, DCs, programming exercise and enables a modeler to concentrate
PEs, MIPSs, BWs, etc. but due to the programming on the simulation parameters rather than the technicalities of
complexities it is hard to implement. In CloudLabs, programming. It also provides the facility of repeated and
Melbourne, Australia GUI based simulation tools quick experiments of simulations with different parameter
CloudReports and CloudAnalyst were developed. values in a very easy way. The advantage of the CloudAnalyst
CloudAnalyst is used in the current study to simulate the is the graphical output of the simulations which can be easily
internet based incoming cloudlets and applications flow. and efficiently analyzed by the users. It also helps to detect
any problem in simulation logic with the performance and
CloudAnalyst tool have three types of service broker policies accuracy.
namely Closest Data Center, Optimize Response Time and
Configure Dynamically with Load. In the advanced CloudAnalyst is a technique which extends its environment to
configuration, one can select any one of the given three Load study the behavior of large scaled Internet applications such
Balancing Policies namely Round Robin (RR), Equally as Facebook, Google Scholar, and other social sites in a cloud
Spread Current Execution Load (ESCEL) and Throttled. The computing environment. This simulation tool can also be
present work compares the load balancing policy in different extended with some novel approach of load balancing
service broker policy on one or more Data Centers with algorithms which could be tried to improve the behavior of
various configurations. The aim of this comparative study is simulation for large scaled. The CloudAnalyst is built by
to select the optimal service broker policy robustly tested on using the framework of CloudSim with some additional
different parameters. The impact of this work is in selection features. [13]

Fig 1: CloudAnalyst built on top of CloudSim Toolkits Fig 2: Graphical Output of Simulation with Colored Regions

It produces output as Response time of the simulated  Graphical output

application, the usage patterns of the application, the time  Repeatability
taken by data centers to service a user request and the cost of  Ease of extension
operation.
3. RELATED WORKS
Bhathiya et al.[13] proposed/described the CloudAnalyst and
The CloudAnalyst is fully developed in Java language, using
developed this tool to simulate the best / optimal selection of
Java SE 1.6. Its GUI component is built using Java Swing
cloud provider in the respect of / term of overall response time,
components and CloudSim is the background used for
Data Center processing time and costs. There have been many
modeling data centers in CloudAnalyst. It is developed on the
studies using simulation techniques to calculate the
platform of CloudSim and CloudSim is developed on the
performance of large scale distributed system. These
platform of SimJava which is the core part of underlying
simulations are GridSim[13], MicroGrid[13], GangSim[13],
simulation framework of CloudSim and responsible for
SimGrid[13] and CloudSim[13] based on heavy java
managing simulation clock and handle basic discrete event
programming and provides the environment to develop the
simulation of cloud environments. So some features of
cloud, design cloud and deploy cloud on its. But CloudAnalyst
SimJava are directly used in CloudAnalyst. It is an open source
is also a simulation tool to simulate the performance and cost
cloud simulation tool developed in year 2009. Its VM
on the basis of basically six divided regions all over the world.
allocation policy works on Time-Shared basis and useful for
Because, Data Centers are located in the any of regions and
heterogeneous environments.
cost paying scheme also varies from vendor to vendor. Thus
2.1 Features of the Simulator one should decide that which of the scheme is profitable for the
application deployment and accessing the cloud. This facility is
 Ease of use
provided by CloudAnalyst. [13]
 Simulations can be run with high configurationally
flexibility

10
 International Journal of Computer Applications (0975 – 8887)
Volume 104 – No 17, October 2014

Brototi et al. [14] proposed the new load balancing technique

Maximum
Minimum
in cloud computing using Stochastic Hill Climbing approach

Average
Load Balancing Policy Across

(ms.)

(ms.)

(ms.)
and compared with the given load balancing policy in VM’s in a Single Data
CloudAnalyst. A comparative study is also done with Round Centers
Robin (RR) algorithm and First Come First Serve (FCFS) and
results are found to be encouraging.
Overall
In this paper, a comparison was done among given load Response 303.36 53.26 641.72
balancing policy and service broker policy with different Round Robin Time
configuration parameters. The analysis of the simulation result (RR) Data Center
can help to infer the best approach to calculate and select Processing 10.25 7.21 13.51
profitable vendors/ providers and in which regions one should Time
select to deploy for applications and data. Overall
Equally Spread Response 303.38 53.26 641.72
4. EXPERIMENTAL SETUP Current Time
This section presents the experiments and evaluations of three Execution Data Center
dynamic load-balancing policies for fixed set of parameters in Load (ESCEL) Processing 10.25 7.21 13.51
CloudAnalyst simulation tool. These experiments were Time
conducted on a Intel (R) Core (TM) i5-3210M CPU having Overall
configurations like : 2.50 GHz processor speed, 4 GB RAM, Response 303.33 53.26 641.72
64-bit OS Windows 7 Ultimate, JDK 1.7 and 500 GB HDD. Time
Simulation was set with One Data Center with two physical Throttled
Data Center
hardware units and six User Bases (UBs) with six different Processing 10.25 7.21 13.51
regions. VM policy is selected as TIME_SHARED with one Time
Data Center having five Virtual Machines. Simulation duration Data Transfer Total
was set to 60.0 in minutes (we can select hours and days also). Data Center VM Cost $
Cost $ Cost $
Advanced simulation configuration chosen as 100,000 users DC1 0.50 0.35 0.85
grouping factor in UBs and 100,000 request grouping factor
per Data Center. The power of DCs set as executable
Table 3. Simulation Result in Respect of Service Broker
instructions length 1000 Bytes per request. Internet
Policy: Reconfigure Dynamically with Load
characteristics configures as well as defined in the simulation
tool, there are specified Delay Matrix and Bandwidth Matrix

Maximum
Minimum
Average
Region by Region for six Regions.

(ms.)

(ms.)

(ms.)
Load Balancing Policy Across
Table 1. Simulation Result in Respect of Service Broker VM’s in a Single Data Centers
Policy: Closest Data Center
Maximum
Minimum

Overall
Average
(ms.)

(ms.)

(ms.)

Load Balancing Policy Across Response 333.80 56.26 672.62

VM’s in a Single Data Centers Round Robin Time
(RR) Data Center
Processing 40.90 8.41 136.51
Overall
Response 303.15 52.36 672.62 Time
Round Robin Time Overall
(RR) Data Center Equally Spread Response 333.30 56.26 672.62
Processing 10.23 7.21 13.26 Current Time
Time Execution Load Data Center
Overall (ESCEL) Processing 40.53 8.41 89.13
Response 303.34 52.36 672.62 Time
Equally Spread
Time Overall
Current Execution
Data Center Response 333.46 56.26 672.62
Load (ESCEL)
Processing 10.23 7.21 13.26 Time
Time Throttled
Data Center
Overall Processing 40.53 8.41 89.13
Response 303.39 52.36 672.62 Time
Time Data Transfer Total
Throttled Data Center VM Cost $
Data Center Cost $ Cost $
Processing 10.23 7.21 13.26
DC1 3.16 0.35 3.51
Time
Data Center VM Cost $
Data Transfer Total 5. RESULTS AND DISCUSSIONS
Cost $ Cost $
DC1 0.50 0.35 0.85
Case 1:
After fixing simulation configuration by choosing service
broker policy as “Closest Data Center”, the “Round Robin”
load balancing policy is producing best result. In this service
Table 2. Simulation Result in Respect of Service Broker
broker policy every load balancing policy produces the same
Policy: Optimize Response Time
result of Data Center Processing Time
(Average/Minimum/Maximum) and Costs (VM/Data
Transfer). Overall Response Time is also same in each load

11
 International Journal of Computer Applications (0975 – 8887)
Volume 104 – No 17, October 2014

balancing policy for the minimum and maximum conditions.

Only average overall response time varies in different load
balancing policy. Overall cost is also same in each load
balancing policy. VM costs and Data Transfer Costs are also
similar in each case of simulation configuration of different
load balancing policy.

After simulation, one can see that Round Robin load policy is
producing best result in average case of overall Response
Time. If no. of DCs, no. of VMs, etc. are increasing then its
overall response time, data center processing time and costs Graph 2: Data Center Processing Time
also increases.

Case 2:
When selected service broker policy is “Optimize Response
Time”, it is producing same results for overall response time
(minimum/maximum) as well as for Data Center processing
time (average/minimum/maximum). It is generating same costs
of the VM and data transfer for all load balancing policies.

The only difference which arises is of average simulation time

of overall response time in case of different load balancing
policies such as: Round Robin, ESCEL and Throttled. On Graph 3: Total Costs (VM + Data Transfer)
making changes in no. of DCs, no. of VMs and simulation
period such as minutes, hours and days different results will be
produced in same pattern as is mentioned above. 7. CONCLUSION AND FUTURE SCOPE
In this paper, a comparison of the given policies of load
balancing at the Data Center level on VMs was performed. By
Case 3: enhancing the given scheduling algorithms and then
When selected service broker policy is “Reconfigure
implementing those in cloud computing environment using
Dynamically with Load”, the Overall response time (minimum,
CloudSim toolkit, one can get better results. By visualizing the
maximum), minimum Data Center processing time and costs
cited parameters in graphs and resulted tables it can be easily
(VM, Data Transfer) produced the same result in all load
identify that which one is the best policy to produce improved
policy.
overall response time and data centre processing time. The
future work includes overcoming the problem of deadlocks and
But slight difference is between/among the overall response
server overflow. New service broker policy in the simulator
time (average) for different given load policy. Average /
can also be implemented. We will try to extend this
Maximum Data Center processing time are same for the
CloudAnalyst tool with some new added parameters and
ESCEL and Throttled.
include some novel policy of load balancing algorithms to
6. COMPARISON optimize the results.
After various rounds of simulation it can be summarized that 8. ACKNOWLEDGEMENT
the best service broker policy and the best load balancing
I [first Author] am very much thankful to Prof. Umesh Singh,
policy are Optimize Response Time and Round Robin as they
Coordinator, DST-CIMS, Faculty of Science, Banaras Hindu
are giving the best result for simulating the internet coming
University, Varanasi, India who has extended all financial as
requests/cloudlets/services/applications and user requests on
well as supportive guidance for carrying out this research work.
the social sites.
9. REFERENCES
[1] L. Kleinrock, A Vision For The Internet St Journal Of
Research, Nov, 2005, (1), Pg4-5.
[2] http://www.livinginternet.com/w/wi_online.htm.
[3] R.Buyya, C. Yeo, S.Venugopal, J.Broberg, I.Brandic,
Cloud computing and emerging it platforms: Vision, hype,
and reality for delivering computing as the 5th utility, in:
Future Generation Computer Systems, vo1.25, 2009, pp.
599–616.
[4] Jane Anderson, Assessing Cloud Computing Challenges
and Opportunities for Network Providers, strategic white
paper, May 2008.
Graph 1: Overall Response Time
[5] Frank Gens, IT Cloud Services Forecast – 2008, 2012: A
Key Driver of New Growth on October 8th, 2008.
[6] Frank Gens, Enterprise IT in the Cloud Computing Era
New IT Models for Business Growth & Innovation, IDC,
2008.

12
 International Journal of Computer Applications (0975 – 8887)
Volume 104 – No 17, October 2014

[7] R. R.Buyya, R.Ranjan, Intercloud: Utility-oriented [11] Anthony Sulistio, Uros Cibej, Srikumar Venugopal,Borut
federation of cloud computing environments for scaling of Robic and Rajkumar Buyya, A toolkit for modelling and
application services, in: ICA3PP 2010, Part I, LNCS simulating Data Grids: An extension to GridSim,
6081, 2010, pp. 13–31. Concurrency And Computation: Practice And Experience,
0123; 34:1, Version: 2002/09/19 v2.02.
[8] Kapil Bakshi, Cisco Cloud Computing -Data Center
Strategy, Architecture, and Solutions , Point of View [12] B.Wickremasinghe, R.N.Calheiros, R. Buyya,
White Paper for U.S. Public Sector 1st Edition 2009 Cisco Cloudanalyst: A cloudsim-based visual modeller for
Systems, Inc. analysing cloud computing environments and
applications, in: Proceedings of the 24th International
[9] Peter Mell, Timothy Grance, The NIST Definition of Conference on Advanced Information Networking and
Cloud Computing, NIST Special Publication, 800- 145, Applications (AINA 2010), Perth, Australia, 2010.
September 2011.
[13] Bhathiya Wickremasinghe, “CloudAnalyst: A CloudSim-
[10] R.N.Calheiros, R. Ranjan, A. Beloglazov, C. Rose, R. based Tool for Modelling and Analysis of Large Scale
Buyya, Cloudsim: A toolkit for modeling and simulation Cloud Computing Environments” MEDC Project Report,
of cloud computing environments and evaluation of 2009, 44 p.
resource provisioning algorithms, in: Software: Practice
and Experience (SPE), Volume 41, Number 1, ISSN: [14] Brototi Mondal, Kousik Dasgupta, Paramartha Dutta,
0038-0644, Wiley Press, New York, USA., 2011, pp. 23– Load Balancing in Cloud Computing using Stochastic
50. Hill Climbing-A Soft Computing Approach. Procedia
Technology, vol. 4, pp.783-789, 2012.

IJCATM : www.ijcaonline.org

13