A report on

Load Balancing
In
Cloud Computing

Done by

Names of Students Roll No

Aviral Nigam B090871CS

Snehal Chauhan B090850CS
Varsha Murali B090484CS

Guide
Vinod Pathari
(Asst. Professor)

Department Of Computer Science And Engineering

National Institute of Technology Calicut
Calicut, Kerala 673 601

Monsoon Semester
 Department Of Computer Science And Engineering
National Institute of Technology Calicut

Certificate

This is to certify that this is a bonafide record of the project presented by

the students whose names are given below during Monsoon-2012 in partial
fulfilment of the requirement of the major project.

Names of Students Roll No

Aviral Nigam B090871CS

Snehal Chauhan B090850CS
Varsha Murali B090484CS

Guide Date
Vinod Pathari
(Asst. Professor)
 Abstract

Cloud computing is the use of computing resources (hardware and software)

that are delivered as a service over a network. Load balancing is a computer
networking methodology to distribute workload across multiple computers
or a computer cluster, network links, central processing units, disk drives,
or other resources, to optimize resource utilization, maximize throughput,
minimize response time, and avoid overload. With Clouds becoming one of
the most important concept in the field of internet and resource sharing, it is
of utmost importance to optimally balance the tasks and loads in the Cloud.
The objective of this work is to study various load balancing algorithms used
in Clouds and to design a new algorithm for tackling this issue.
Contents

1 Introduction 1
1.1 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Report Organization . . . . . . . . . . . . . . . . . . . . . . . 2

2 Literature Survey 3

3 Work Done 6
3.1 Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.1.1 Biased Random Sampling . . . . . . . . . . . . . . . . 6
3.1.2 Honeybee Foraging . . . . . . . . . . . . . . . . . . . . 7
3.1.3 Proposed Hybrid Algorithm . . . . . . . . . . . . . . . 8
3.2 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.2.1 Biased Random Sampling . . . . . . . . . . . . . . . . 9
3.2.2 Honeybee Foraging . . . . . . . . . . . . . . . . . . . . 9
3.2.3 Proposed Hybrid Algorithm . . . . . . . . . . . . . . . 10
3.3 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.3.1 Biased Random Sampling . . . . . . . . . . . . . . . . 10
3.3.2 Honeybee Foraging . . . . . . . . . . . . . . . . . . . . 11
3.3.3 Proposed Hybrid Algorithm . . . . . . . . . . . . . . . 12

4 Future Work 14

References 16

i
Chapter 1

Introduction

The Cloud Computing model enables users to access supercomputer-level

computing power elastically on an on-demand basis, freeing the users from
the expense of acquiring and maintaining the underlying hardware and soft-
ware infrastructure and components. “A Cloud is a type of parallel and dis-
tributed system consisting of a collection of interconnected and virtualized
computers that are dynamically provisioned and presented as one or more
unified computing resources based on service-level agreements established
through negotiation between the service provider and consumers”. Cloud
computing, as a current commercial offering, started to become apparent
in late 2007. It was intended to enable computing across widespread and
diverse resources, rather than on local machines or at remote server farms.
Where these clusters supply instances of on-demand Cloud computing; pro-
vision may be comprised of software (e.g. Software as a Service, SaaS) or of
the physical resources (e.g. Platform as a Service, PaaS).

1.1 Problem Statement

• A study of the existing load balancing algorithms in Cloud comput-
ing and implement a new hybrid load balancing algorithm. This will
include an analysis of these above algorithms and to draw conclusions
based on their execution time.

• Comparison of this implementation with a game-theoretic approach to

load balancing

1
1.2 Report Organization
At the outset, the report contains a brief review on the existing works done
in the field of Cloud computing focussing on the area of load balancing.
A detailed study on two of the popular load balancing algorithms and the
design of a new algorithm which is a combination of the above follows. A
comparative analysis of the implenmentation of these algorithms has been
explained. The report concludes with a plan on the future action to be done
in the following semester.

2
Chapter 2

Literature Survey

Cloud computing is a model for enabling ubiquitous, convenient, on-demand

network access to a shared pool of configurable computing resources (e.g.,
networks, servers, storage, applications, and services) that can be rapidly
provisioned and released with minimal management effort or service provider
interaction [1].

Service Models
• Cloud Software as a Service (SaaS): The capability provided to
the consumer is to use the providers applications running on a Cloud
infrastructure.
• Cloud Platform as a Service (PaaS): The capability provided to
the consumer is to deploy onto the Cloud infrastructure consumer-
created or acquired applications created using programming languages
and tools supported by the provider.
• Cloud Infrastructure as a Service (IaaS): The capability provided
to the consumer is to provision processing, storage, networks, and other
fundamental computing resources where the consumer is able to deploy
and run arbitrary software, which can include operating systems and
applications.
Load Balancing is a method to distribute workload across one or more
servers, network interfaces, hard drives, or other computing resources. Load
balancing is used to make sure that none of the existing resources are idle
while others are being utilized. To balance load distribution, migration of
the load from the source nodes (which have surplus workload) to the com-
paratively lightly loaded destination nodes can be done [3].

3
 Goals Of Load Balancing

• To improve the performance substantially.

• To have a backup plan in case the system fails even partially.

• To maintain the system stability.

• To accommodate future modification in the system.

A distributed solution is required, as it is not practical or cost efficient,

in many cases, to maintain idle service/hardware provision merely to keep
up with all identified demands. Indeed it is not possible, when dealing with
such complexity, to fully detail all system future states; it is thus necessary
to allow local reasoning, through distributed algorithms, on the current sys-
tem state. Thus efficient load balancing cannot be achieved by individually
assigning jobs to appropriate servers as the Cloud computing systems scale
up and become more complex; rather jobs must inevitably be assigned with
some uncertainty attached. This gives some immediate possible methods for
load balancing in large scale Cloud systems, which are discussed here [3].

In biased random sampling, a nodes in-degree (node capacity) is mapped

to its free resources. When a node receives a new job, it will remove one of
its incoming edges to decrease its in-degree and indicate that its available
resources are reduced. In the same way, when the node finishes a job, it will
add an edge to itself to increase its in-degree. The process of adding and
removing incoming edges is done by random sampling. Random sampling is
the process whereby the nodes in the network are randomly picked up with
equal probability. The sampling walk starts at some fixed node, and at each
step, it moves to a neighbor node chosen randomly according to an arbitrary
distribution. Then, the last node in the course of the sampling walk will be
selected for the load assignment [2].

Honeybee Foraging is one of a number of applications inspired by the

believed behaviour of a colony of honeybees foraging and harvesting food.
Forager bees are sent to search for suitable sources of food; when one is
found, they return to the hive to advertise this using a display to the hive
known as a “waggle dance”. The suitability of the food source may be derived
from the quantity or quality of nectar the bee harvested, or its distance from
the hive. This is communicated through the waggle dance display. Honey
bees then follow the forager back to the discovered food source and begin to

4
harvest it. This approach has been used for various applications in comput-
ing [2].

Experiments have been conducted to assess and analyze these procedures

for load balancing claiming that none of the load balancing methods are
mutually exclusive and it is possible that combinations could be used to
further improve the performance of the algorithms [4].

5
Chapter 3

Work Done

The algorithms that have been considered for the project are stated below :

• Biased Random Sampling

• Honey Bee Foraging

• Hybrid Algorithm

3.1 Theory
3.1.1 Biased Random Sampling
In this type of approach we need to construct a network comprising of virtual
nodes which represent all the resources present on the server to represent the
total load. The indegree of the node corresponds to the free resources of the
server such that :

• Whenever a node executes a job, it deletes an incoming edge, which

indicates reduction in the availability of free resources.

• After completion of a job, node creates an incoming edge which indi-

cates an increase in availability of the resources.

6
 The working of the algorithm can be understood by the following pseudo
code :

b = log n //n - network size

For any node that recieves a new process

/*Create and send token to
randomly selected neighbour*/
job.walklength = 1
/*Select a neighbour using
probability based function*/
Randomselect(node)
/*Send job to selected neighbour*/
Send(job,neighbour)

For any node that recieves token

/*Update token if needed and send it
to neighbour selected by the function*/
if job.walklength < b then
neighbour = Randomselect(node)
Send(job,neighbour)
else
/*Execution of job on node
indiacted by the token*/
Execute(job)
endif

3.1.2 Honeybee Foraging

The first job, on being sent into the network of servers acts, as a scout
providing information on the availability of resources at each server. This
information is published on an advert board which is referenced by the in-
coming jobs. Based on the fitness function, jobs attach themselves to the
server and simultaneously update the advert board.

7
 The working of the algorithm can be understood by the following pseudo
code :

/*Scout on entering network*/

TraverseNetwork(scout)
CreateAdvertBoard()

For all incoming jobs

/*Checking for the best server*/
node = fitness(job)
/*Job Execution*/
Attach(node,job)
/*Updating advert board with
current status of resources*/
UpdateAdvertBoard(node,job)

3.1.3 Proposed Hybrid Algorithm

This algorithm is a combination of biased random sampling and honeybee
foraging. The incoming jobs are scheduled on the basis of their job size.
The working of the algorithm can be understood by the following pseudo
code :

/*Jobs on entering network*/

if jobsize==1 then
BiasedRandomSampling(job);
else
HoneyBeeForaging(job);
endif

3.2 Implementation
An array of 100 jobs with random jobsize and execution time and an adja-
cency matrix of a fixed network of servers are given as inputs. The server
network being created by the adjacency matrix forms a directed graph which
allows for easy traversal between the servers.

• Nodequeue: This is a list of all the jobs being executed by the partic-
ular server including the job size, execution time and the time at which
the job was allocated to the server.

8
 • NodeArrayList: The list of all the servers in the network and their
accompanying node queue are contained here.
The increaseindegree(), common to all the algorithms, traverses the array
list associated with each job and checks if any job that was being executed
by each server has been completed. This is done by comparing the current
time with the sum of the execution time required for the job and the time
at which it was allocated to the server. On completion of the jobs, the array
list is updated accordingly and the resources available with the server are
incremented.

The timer functions are as follows: jobs are sent every one second and if
the job cannot be allocated presently to any server due to lack of resources
then it waits for two seconds before trying again.

3.2.1 Biased Random Sampling

• BiasedRandomWalk: The selectwalk() initially checks for the avail-
ability of the resources with the help of increaseindegree() and uses
the hopcount which gives the length of the walk. It randomly chooses
a server and checks if the job can be allocated to be followed by the
invocation of selectneighbour(). If no resources are available, then it
randomly chooses another server and continues the above process. The
selectneighbour() probabilistically chooses the next best server based
on the resources available in each of the neighbour. This neighbor se-
lection is done till the hopcount value is reached. The send() compares
the resources between the last two servers and updates the token with
the better value. It then updates the resource information of the server
to which the job has been allocated and finally returns the server node.
• GraphBiased: The creategraph() contains the server network (rep-
resented by an adjacency matrix) and also the information pertaining
to the availability of resources with each server. An object of the Bi-
asedRandomWalk class is created and we check for the availability of
resources in the network. If resources are unavailable, then the next
job to be allocated waits. Only when resources are made available the
process of random selection starts.

3.2.2 Honeybee Foraging

• Fitness: The class contains select() which initially calls increaseinde-
gree() to update the current resource status of the servers. A traversal

9
 of the server network is done whereby for each server we calculate
the difference in the resources currently available and the resources
requested. The job is then allocated to the server for which the dif-
ference is small. The resource available for that server is decremented
accordingly and it finally returns the best fit server.

• GraphHoneyBee: The server network is represented with an adja-

cency matrix. The creategraph() forms the network and calculates the
available resources with each server. This is called when the scout job is
sent. The fitness() is invoked which does the job allocation. If job could
not be allocated because of lack of resources, then the current job waits
and calls the increaseindegree() to get updated resource information.

3.2.3 Proposed Hybrid Algorithm

This algorithm makes use of the BiasedRandomWalk and Fitness classes from
the above two algorithms.

• Hybrid: The creategraph() contains the server network (represented

by the adjacency matrix) and also about the resources available with
each server. Objects of BiasedRandomWalk and Fitness classes are cre-
ated. The job() receives each job and then checks if any resources are
available in the network and then calls the selectjob() (after waiting,
if resources were unavailable). If the job requires only one resource,
then the biased random sampling algorithm is used while the honeybee
foraging algorithm is used otherwise. The selectjob() chooses the algo-
rithm for each particular job and then sends the job to the respective
functions in BiasedRandomWalk or Fitness.

The set of 100 jobs is executed by the algorithm and its total execution
time is recorded. This process is repeated for 100 times and an average
execution time is obtained for each algorithm.

3.3 Analysis
3.3.1 Biased Random Sampling
In the biased random sampling algorithm, for every batch of hundred jobs,
small changes in the range of seconds were obtained. The change which oc-
curred is because of the delay being caused due to unavailability of resources.
A random node is selected only initially and the walk is taken based on the

10
probabilistic function and hence the walk only goes up to the value of hop-
count (which is the same) for every allocation of job. The time taken for load
balancing of every job varies by seconds because the algorithm looks for a
server with maximum resources available and then tries to allocate the job.
So, overheads are obtained because of finding the path for every job followed
by comparing it; even though only finally the job size is checked.

3.3.2 Honeybee Foraging

In the honey bee foraging algorithm, the running time for the batch of jobs
is almost same only varying by milliseconds. This change is due to the delay
caused by the unavailability of resources since the other jobs are utilizing all
the servers in the network. The portions of consistent timing are due to the
traversal of the same advert board by every job being sent into the network.
Since the advert board is a graph traversal with a list of jobs attached with
each node, the algorithm takes almost similar time for the traversal in each

11
case since the job size does not vary drastically.

3.3.3 Proposed Hybrid Algorithm

In the proposed hybrid algorithm, the timing pattern shows a divergence in
the range of seconds with almost no perfectly same running time in a batch
of ten. This variation is due to the unavailability of resources which causes a
delay. This is because of the server being chosen, for a particular job, by both
the algorithms. While biased chooses the server with the maximum number
of resources, honey bee goes in for a more optimal selection taking a best fit
approach. If we consider a case where a biasing function has been applied
such that the probability of choosing either of the algorithm is same, then a
more inconsistent graph with larger number of peaks and dips are obtained
because of the above explained contradiction.

12
 All the three algorithms show changes only in the range of milliseconds
to seconds for a batch of hundred jobs being run hundred times. While
biased takes 10870 ms, honey bee runs in 10140 ms and hybrid in 17617ms
for a batch of 100 jobs on an average. This clearly shows that the honey
bee algorithm is the best approach of all the algorithms considered. Though
honey bee has a more consistent run time (the inconsistency occurs due to
reaching a deadlock state where no resources are available) as compared to
biased (the inconsistency being caused by the delays due to the checking of
job size only finally), the hybrid algorithm has a highly inconsistent run time
graph because of entering the deadlock state more due to the contradiction
in server selection between the two algorithms.

13
Chapter 4

Future Work

A complete study and implementation of the algorithms has been carried out
over the past three months. In the following semester, the area of load bal-
ancing in Cloud computing will be approached from a whole new perspective
of game theory. The problem, till now, has been addressed by scheduling
one job at a time to the best server and thus following the same strategy
for a group of jobs. The problem can also be viewed as scheduling a group
of jobs to the best servers at one instant of time. This approach has been
taken because it may not be necessary that the best allocation for each job
is the best for the given group of jobs i.e., it is possible to have a collec-
tive best outcome which may be better than a collection of individual best
outcomes. Hence this problem can be looked at from the technique of game
theory which fundamentally states the same.

Game theory is the study of strategic decision making. Strategies and

payoffs are calculated and a payoff matrix is formed which will contain the
values obtained on applying a utility function. This approach has been widely
used in various fields of computer science. The matrix given below can be
obtained for load balancing problem where each job can have different affinity
towards different servers.

Job/Server A B C
A 5 2 5
B 9 1 3
C 4 1 8

Such kind of representation has motivated us to give a game-theorectic

approach to this problem.

14
 Cloud computing, being a new concept, has a lot of scope for research.
So, one of the ways to tackle load balancing in Cloud computing is through a
game theoretic approach with proper strategies, payoffs and payoff matrices.
In the coming months, the objective will be to study game theory and to use
it to obtain optimal solution for load balancing in Cloud computing.

15
References

[1] Peter Mell, Timothy Grance; The NIST Definition of Cloud

Computing (Draft); Recommendations of the National Insti-
tute of Standards and Technology, http://http://csrc.nist.gov/
publications/nistpubs/800-145/SP800-145.pdf

[2] Martin Randles, David Lamb, A. Taleb-Bendiab; School of Com-

puting and Mathematical Sciences, John Moores University, UK; A
Comparative Study into Distributed Load Balancing Algorithms for
Cloud Computing, http://ieeexplore.ieee.org/stamp/stamp.jsp?
tp=&arnumber=5480636

[3] Ratan Mishra and Anant Jaiswal; International Journal of Web & Se-
mantic Technology (IJWesT) Vol.3, No.2, April 2012; Ant Colony Op-
timization: A Solution of Load Balancing in Cloud, http://airccse.
org/journal/ijwest/papers/3212ijwest03.pdf

[4] Martin Randles, David Lamb, A. Taleb-Bendiab; School of Computing

and Mathematical Sciences, John Moores University, UK; A Compara-
tive Experiment in Distributed Load Balancing, http://ieeexplore.
ieee.org/stamp/stamp.jsp?tp=&arnumber=5395118

[5] Sheeja S. Manakattu, S. D. Madhu Kumar; ICACCI ’12 Proceedings

of the International Conference on Advances in Computing, Commu-
nications and Informatics; An improved biased random sampling algo-
rithm for load balancing in cloud based systems, http://dl.acm.org/
citation.cfm?id=2345472

16