Fuzzy Logic–Based Scheme for Load Balancing in Grid Services

doi:10.4236/jsea.2012.512B029

Paper Menu >>

Journal Menu >>

A Journal of Software Engineering and Applications, 2012, 5, 149-156

doi:10.4236/jsea.2012.512b029 Published Online December 2012 (http://www.scirp.org/journal/jsea)

Fuzzy Logic–Based Scheme for Load Balancing in Grid

Services

Tarek Helmy*, Hamdi Al-Jamimi, Bashar Ahmed, Hamzah Loqman

Information and Computer Science Department, Col lege of Computer Scien ce and Engin eering, King Fahd Uni versit y of Pet roleum

and Minerals, Dhahran 31216, Mail Box, 413, Saudi Arabia; “On leave from College of Engineering, Department of Computers

Engineering & Automatic Control, Tanta University, Egypt”.

Email: helmy@kfupm.edu.sa

Received 2012

ABSTRACT

Load balancing is essential for efficient utilization of resources and enhancing the responsiveness of a computational

grid, especially that hosts of services most fr eq uentl y us ed , i.e . food , health a nd nutritio n. Var io us te c hniques ha ve b ee n

developed and applied; each has its own limitations d ue to the dyna mic natur e of the grid. Efficient load balancing can

be achieved by an effective measure of the node’s/cluster’s utilization. In this paper, as a part of an NSTIP project #

10-INF1381-04 and in order to assess of FAQIH fra me work ability to support the load balance in a co mputational grid

that hosts of food, health and nutrition inquire services. We detail the design and implementation of a proposed

fuzzy-logic-based scheme for dyna mi c lo a d ba la nci ng in gr i d co mputi n g services. The proposed scheme works b y using

a fuzzy lo gic inference system which uses some metrics to capture the variability of loads and specifies the state of each

node per a cluster. Then, based on the overall nodes’ states, the state of the corresponding cluster will be d efined in o r-

der to assign the newly arrived inquires such that load balancing among different clusters and nodes is accomplished.

Many experiments are conducted to investigate the effectiveness of the proposed fuzzy-logi c-based scheme to support

the load balance where the results show that the proposed scheme achieves really satisfactory and consistently load

balance than of other randomize approaches in grid computing services.

Keywords: Fuzzy Logic; Load Balancing; Grid Co mputing.

1. Introduction

Grid computing can be considered as a type of parallel

and distributed systems that enables the selection,

distr ibutio n, and aggr egati on of r esour ces d ynamica lly a t

run time depending on their availability, capability, and

performance. It focuses on large-scale resource sharing

and distributed system integration for the sake of effective

utilization and high-performance orientation. Indeed, at

the service level of the grid, software infrastructure, the

wor klo ad and r e so ur c e mana ge me nt a re e ss e ntial p ro vide d

functions. In turn, distributed systems have become a

natural setting in various environments either for business

or academia [1,14,17-20]. In like these systems settings,

tasks arrive to the different no des of the cluster randomly.

This random fashion leads to a non-uniform distribution

of the workload among different nodes of the cluster.

Loading imbalance is harmful to the whole system

performance in terms of the mean response time of tasks

and resources utilization. Therefore, load balancing tries

to increase system throughput by making all processors

as busy as possible. The objective can be achieved by

two ways: either by avoid sending more tasks to the

overloaded nodes and keep checking the nodes’ states

periodically, or taking a migration decision to migrate

some tasks fr om the overloaded n odes t o the light ly loa ded

nodes for the sake of improving the s ystem per formance

as a whole. The load balancing dilemma in a distributed

computing environment has been addressed extensively

by many studies and researches. Based on that, many

algorithms have been introduced to tackle the load

balancing pro ble m. All kinds o f these algorith ms fall into

one of the follo wing two ca tegories: dynamic or static [ 2,

18]. Dynamic load balancing algorithms exploit the

information descr ibing the state o f the system in ord er to

improve th e accuracy o f load balanci ng decision. Therefore,

the dynamic algorithms are appropriate to be utilized in

practical situations. However, this kind of algorithms

should have the ability to collect the information about

*Corresponding author.

Fuzzy Logic–Based Scheme for Load Balancing in Grid Services

150

the current state of the system, to store the collected

information and to analyze them. Thereby, the dynamic

algorithms may cause additional overhead computations

for the system. On the other hand, the static algorithms

don’t exploit such information. In static algorithms, the

decisions are taken based on a priori knowledge of the

system. Thus, the sta tic algorithms don’t have the abilit y

to deal with the dynamic changes of such environments.

An efficient load balancing scheme is needed for grid

computing environment. The effective scheme needs to

know the global system’s sta te such as the distrib ution of

the workload. Ho wever, the lo ad model of t he distrib uted

systems tends to change dynamically, and hence, it is

very complicated to model the system accurately. Thus,

precise and fast load balancing scheme is an essential

factor in increasing the efficiency and performance of

grid computing components. However, many factors lead

to ambiguous information about the cl us ter s/nodes states,

i.e. lack of shared memory among independent clusters.

This ambiguity leads to uncertainty in decision for load

balancing. Another important aspect is that the states of

the clusters/nodes can be changed rapidly. Therefore

there is always some amount of uncertainty in the state

information used for making a decision. Hence it is

necessary that the decision making process deals with

these uncertainties. Fuzzy logic is one of the methods

that deal with the uncertain information [4,5,13,15,16].

In this paper, in order to tackle the problem of load

balancing in the grid computing environment where

uncertainty is unavoidable, we utilize a Fuzzy Inference

System (FIS) -based approach to model those contributed

factors to determine the nodes’ states and thereby the

clusters’ states. In some previous studies, both length of

the ready queue and CPU utilization have been used to

provide a good load indicator. In the case of multiplicity

of resources, a good measure can be considered like a

linear combination of the length of all resource queues.

In this paper we use four parameters as input variables

for the sake of more accurate fuzz y-logic based approach.

We consider a continuous stream of independent jobs

which arrive to the system and are stored into a single

job queue. We assume that all jobs are not preemptable,

and each job is described by its ID, submission time,

deadline time, an estimation of its execution time and the

number of resources it requires. The computing grid is

composed of many independent clusters of homogenous

processing machi nes. Because of the lack o f a mechanism

in the computing platform to notify configuration changes

to the system, we assume to have a static number of

resource instances available to cluster’s processing node.

The rest of the paper is organized as follows: Section 2

describes technical background about fuzzy logic concepts.

Section 3 introduces the related work. The proposed

fuzzy-logic based scheme is demonstrated in Section 4.

The implementation details of the proposed scheme are

described in Section 5, while Section 6 introduces the

simulation environment and discusses the obtained

results. Section 7 concludes the paper and introduces

some trends for future work.

2. Technical Background

In this section, we introduce a brief ba ckground abo ut the

concept s used in this paper. Fuzzy logic methodolo gy can

be utilized in solving the problems tha t are complex to be

analyzed quantitatively or in the case of natural phenom-

ena that are not easy to be modeled mathematically [6].

IF-THEN rules are used in the FIS where the output is

concluded from the fired rules based on the given inputs

[3,4]. The system parameters are modeled as linguistic

variables based on expert knowledge; also the corre-

sponding membership functions are designed for each

parameter. Therefore, fuzzy logic theory can be employed

even for the nonlinear systems suffe ring from uncer tainty

and complexity. That complex uncertain system can be

modeled effectively based on fuzzy logic without need

for complicated mathematical models [5]. A general FIS

engine mainly involves the following components as

shown in Figure 1:

• The fuzzifier receives the inputs which are the node’s

information in the proposed model case. The fuzzifier

maps t he given numerical inputs to fuzzy sets and lin-

guistic v ariables. Where th e def ined lin gui stic variables

can be matched with the fuzzy rules premises that are

predefined in the rule base of the specified application.

• The rule base includes a set of lingui stic rules designed

in the form (i f-then rule s). These rule s define the con-

sequent of the model in terms of the given linguistic

variables such as low, moderate and high. In addition, it

specifies the type and number of the used membership

functions of input and output parameters.

• The fuz zy infer ence engine is consi dered the central part

of the fuzzy system. In this stage, in order to produce

the intende d output, the inference engine is applied to a

set of rules included in the fuzzy rule base. This pro-

ced ure invo lves ma n y step s as follo ws: first, it matc hes

the linguistic variables of the input with the rules’

premises. Second, it activates the matched rules in order

to deduce the resultant of each fired rule, and finally it

combines all consequents by using fuzzy set union in

order to generate the final output which is represented

as fuzzy set output.

• The defuzzifier: as described in the previous phase, the

output is produced as a linguistic variable, which is

fuzzy and can be interpreted in different ways. There-

fore, the fuzzy set ou tput in th is stage is co nverted to a

Fuzzy Logic–Based Scheme for Load Balancing in Grid Services

151

crisp output, which is a numerical value.

3. Related Work

The issues related to load balancing in the distributed

systems have been addressed broadly by many studies

and researches. In the literature, many studies have

proposed the using of fuzzy logic the ory to solve the l oad

balancing problem. Since the load balancing problem can

be considered as NP-hard [6], in this case heuristics are

sought to tac kle th is problem.

Nejad et al. [8] proposed a fuzzy logic-base load

balanci ng in ce ntraliz ed di stributed sys tem. T he effic iency

of proposed load balancing is studied in OPNET

simulation environment. The lengths of sent packet and

service rate in each node have been considered non-

const ant. The current load of t he syste m and waiting time

for the last processed task in each node are considered as

the fuz zy co ntr oll er inp uts . Ba sed on the weig ht as signe d

for each node, a percentage of existing tasks is assigned

to that nod e . The result of the experi ments in the states of

constant and variable nodes, and constant and variable

tasks, indicates that this algorithm performs much better

than both static and dynamic algorithms in terms of

thro ughput, dr op rate and response time.

Bey et al. [9] introduced a model that can be used for

pre dic tin g the f utur e CP U lo ad in a d yna mic en viro n ment.

The proposed model is for single-step-ahead CPU load

prediction. The multiple local Adaptive Networks-based

Fuzzy Inference Systems (ANFIS) predictors were used

to build the introduced prediction model. In turn, the

predictors controlled via the Naïve Bayesian Network

(NBN) inference between clusters states of CPU load

time points obtained b y C-me ans clustering process. The

ANFIS models were used to carry out short-term

accurate and mid-term reliable prediction. The selection

was based on the basis of divide-and-conquer principle

that divides a CPU load time series into several clusters.

Afterwards, those clusters can be used separately. I ndeed,

in that case, the ANFIS system behaved accurately and

gave a reasonable performance.

Figure 1. The general architecture of the fuzzy logic inference

engine.

Revar et al. [10] focused on analyzing load balancing

requirements in a grid environment and proposed design

and implementations of innovative load balancing system

for grid environment using machine learning. Their

technique balances the load dynamically. The technique

uses initial load information sto red in the database at t he

initial level. When load imbalance takes place, the

current load information is collected and stored as raw

data. Afterwards, various machine learning algorithms

have been used to process and analysis the recorded data.

As a final step, the rules automatically generated b y data

mining techniques and used for migrating jobs for load

balancing.

K. Ming and C. Hsun [11] designed and implemented

a load-balancing system based on fuzzy logic control in

loosely coupled distributed system. The run-q ueue le ng t h

and CPU utilization were used as input variables. They

used si x wo r ksta tions running different UNIX OS to bu ild

a heterogeneous computing environment. These works-

tations are located on different workstations and connected

by communication network. Based on their experimental

results, the fuzzy logic model for load balancing decreased

effectively the amount of communication traffic of the

network and provided substantial enhancement in the

overall performance.

Rantonen et al. [12] designed and implemented a

fuzzy expert system for load balancing in a symmetric

multiprocessor environment. The novelty of their

algorithm is the use of the algorithm for load balancing

only on demand, instead of using the algorithm

periodically. Thereby the computations overhead can be

reduced that makes the algorithm more fair and fast. Two

parameters are considered as input to the proposed model:

the number of threads per processor and the sum load of

ready queues. They claimed that, the evaluation results

proved the best load balance using different techniques.

To the best of our knowledge, none of the previous

proposed fuzzy logic-based techniques has used more

than two parameters to evaluate the workload. In this

paper, to evaluate the node’s workload, we use four

parameters (Node’s ready queue length, burst time, CPU

utilization, and the available resources needed to

accomplish the assigned tasks), for the sake of more

accuracy. Moreover, two fuzzy-logic based models are

used i n di fferent level s whi ch ar e node -level and clu ster-

level.

4. The Proposed Fuzzy-Logic-Based Load

Balancing Scheme

In the proposed scheme, the load balancing task is de-

signed to be performed at two levels: i) at the cluster

level by global manager, and ii) at the node level by local

manager. The grid consists of many clusters, where sev-

Fuzzy Logic–Based Scheme for Load Balancing in Grid Services

152

eral nodes constitute one cluster. In each cluster, there is

a local manger to evaluate the states of the nodes be-

longing to its c luster and to per form local load balancing

in its own cluster. The nodes of each cluster only com-

municate with the local manager. Assuming that, the task

allocated to each cluster/node will be processed in that

cluster/node and will not be relocated. Global manager

communicates with the local manager of each cluster.

The states of nodes in each cluster are sent to the global

manager, to specify the state of the corresponding cluster.

The struct ure of the fuzzy lo gic-based scheme is d e mon-

strated in Figure 2. I n this scheme, the workload state for

the nodes and the clusters are evaluated. This evaluation

is carried out in the grid and the cluster levels by the

glob al and the local manager s respectively. Based on the

generated states, a newly arrived task will be assigned to

the lightly loaded cluster. Then, based on the states of the

nodes, the arrived tasks will be assigned to the lightly

loaded node/nodes. Deciding the state of the nodes is

made by fuzzy logic-based model, and thereby the clus-

ters states are evaluated. Four parameters as mentioned

above are collected from the current information tasks

and used as inputs to the fuzzy model. The process of

load balancing is handled in different stages as follow-

ings:

4.1 Data Collection Stage

Fuzz y logic -based scheme need s data about each node to

determine the state of that node. The parameters collector

collects the data about each node from the table contains

the information about the current tasks. This collected

information is then passed to the FI S that performs the

needed analysis and takes a decision about the node’s

state. All the nodes sates are sent from time to time to the

local manager that carries out the local balancing. To

exchange the data, each of the nodes notifies its manager

about its load state periodically. That is, to specify the

current state of the grid, clusters and nodes, the balancing

data in the system, is updated without any request. Each

local manager does the task of collecting the node’s data,

in its group, by using the data collector. Then, based on

the data received from the nodes, the fuzzy logic scheme

evaluates the workload of the nodes, determines their

states, and sends the results to the local manager. After-

wards the local manager store s the wor kload states of the

nodes and sends a copy to the global manager. Based on

the data received from local managers, the global man-

ager calculates the workload of each cluster and decides

about its global state in general, and also sends the clus-

ters states to the grid manager.

4.2 Specifying the Workload Sate stage

In this stage, the workload state of each node is deter-

mined and the decision on load distribution is taken. We

consider four different inputs variables to the fuzzy logic

model, namely: the node’s ready queue length, the esti-

mated burst time of the tasks assigned to the node, the

CPU bound for each task, and the available resources for

each node. Three fuzzy subsets have been considered for

each input to determine the fuzz y rules in order to get the

output.

4.3 Distribution Satge

The grid manager, based on the inferred states of differ-

ent nodes and clusters, ranks the clusters and nodes

based on their availability. In that sense, the distributor

routes the arriving tasks to the appropriate cluster and

node based o n the rank. I n this stage, the grid d istributor

utilizes the workload information about the different

clusters to decide straightforwardly which cluster is

lightly loaded. It then distributes the new arriving tasks

to the clusters/nodes based on their workload states col-

lected in advance. Finally, the local manager inside the

target cluster forwards the arrived tasks to the lightly

loaded nodes based on their inferred states.

5. Implementation of the Proposed Scheme

Using Matlab’s FIS toolbox, we implemented two fuzzy

logic based models. The first model to evaluate the

node’s workload based on four parameters explained

above while the second fuzzy model is to evaluate the

clusters workload based in the nod es state s that belo ng to

the target cluster. Following, we will detail the design

and implementation of each model by introducing its

inputs, rules, output, and accuracy measure when train-

ing and validati ng each model.

Figure 2. Th e architecture of the propos ed scheme

Fuzzy Logic–Based Scheme for Load Balancing in Grid Services

153

5.1 Fuzzy Logic Model – Node Level

Each input variable is represented by five Membership

Functions (MFs) of Gaussian type except the last para-

meter only three MFs are used. For example, when the

length of queue is 45, this value has around 0.05 degree

of membership to the function as short and about 0.75

degree of membership to the function as moderate and

also has zero degree of membership to the function is

long. Thus, we can classify the queue length of 45 as

moderate. For all the input s t he r ange s di ffer fro m o ne to

another. For instance, the range of queue length is [0

-100]. However, when the queue length is greater than

the maximum in the range which is 100, the MF is de-

fined as very long because the degree of membership to

the function very long is 1. The MFs for the rest of the

input variables operate in a similar way. In general, the

fuzzy rule base includes a set of linguistic rules. Those

rules represent the communitarians between the different

input parameters, where each parameter is represented by

linguistic variables. That is, the rules promises represent

the input parameters, where the consequent of each rule

indicates the resultant output. The rules are triggered by

mapping the input parameters to the input functions

which might be one or more. In our system, based on the

given inputs to the FIS and the degree membership of

each, 375 rules are obtained; t wo of them are shown in

Figure 3.

For example, in rule1, if the number listed in the ready

queue is low, and the CPU utilization is small, and the

available resources for the corresponding node to achieve

the assigned tasks are high, then the fuzzy output which

indicates the node state, would be very light. While in

rule 2, if the ready queue length is high and the CPU is

very heavily utilized and the resources needed are not too

much, then the node state would as heavy loaded node

and no more tasks will be send to it. The output of the

fuzzy model represents the workload value of the cor-

responding node based on the given parameters. For

trai ni ng a nd te sti ng the p ro p o sed fuz z y mo de l t o e va l uat e

the node’s workload, we generated a dataset consists of

the input parameters and has one output represents the

worklo ad base d on the give n input s. The d ataset consi sts

of six hundred data samples. To build the fuzzy logic

based model, the dataset is divided into training set and

validation set where 70 % of the samples are used for

training the model and 30% are used for validation. Fig-

ure 4 shows the actual and predicted output for both

training and testing stages. It is obvious from Figure 4

that no much difference between the actual and predicted

workload for each node. The accuracy measure for the

model is measured in terms of error measure, where the

Root Mean Square Error (RSME) for training and vali-

dation are 0.000291 and 0.00398 respectively.

5.2 Fuzzy Logic Model – Cluster Level

Similarly for this model, the input variables are

represented by five MFs of Gaussian type. Since the in-

put value is the node’s sate, the five MFs representing

the load state are ‘very light’, ‘light’, ‘moderate’, heavy’

and ‘very heavy’. The range of each parameter is be-

tween 0 and 1 representing the workload for each node

assu ming tha t we ha ve t wenty no des in e ach cl uster. T he

output of the fuzzy model represents the workload state

of the corresponding cluster based on its nodes states.

We trained and evaluated the proposed fuzzy log-

ic–based model at the cluster level by using generated

dataset which consists of 300 data points. Each data

point has 4 inputs represents the cluster’s nodes states

and one outp ut which indicate s the workload o f the clus-

ter. 70% of the samples were used for training the model,

and 30% of the samples have been used for validation.

Figure 5 shows the actual and predicted output for both

training and testing stages. It is obvious from Figure 5

that, no much difference between the actual and pre-

dicted workload for each cluster. The accuracy measure

for the model is measured in terms of error measure,

where RSME were as follows: 0. 0.00095 for training

and 0.00329 for validation re sp e c tively.

6. Simulation and Experimental Results

In the simulatio n, we consider ed that a grid composed of

four independent clusters of homogenous processing

nodes. Because of the lack of a mechanism in the

computing platform to notify configuration changes to

the system, we assume to have a static number of

resource instances available to each cluster’s processing

node. In addition, we consider that the number of nodes

equally distributed among the different clusters where

each cluster consists of twenty nodes. We consider a

continuous stream of independent tasks which arrive to

the s ystem and are st ored into a s ingle jo b que ue. All t he

arrived tasks are not preemptable, and each task is

described by five attributes: the task ID, task’s

submission time, task’s deadline time, an estimation of

its execution time, and it’s CPU-bound. Moreover, a

dataset contains information about 5000 tasks was

generated and used. In the first run of the simulation, a

Figure 3 . Rules for fuzzy i nference system

1. If (ReadyQueueLength is VeryShort) and (BurstTime is

VeryLow) and (CPUBound is VeryLow) and

(AvailableResources is Low) then (WorkloadSatae is out1mf1

)

2. If (ReadyQueueLength is VeryHigh) and(BurstTime is

VeryHigh) and (CPUBound is VeryHigh) and

(AvailableResources is High) then (WorkloadSatae is

out1mf375)

Fuzzy Logic–Based Scheme for Load Balancing in Grid Services

154

Figure 4 . The fuzzy model – Node s tate (training and testi ng

accuracy measu res)

Figure 5. The fuzzy model – Cluster state (training and

testing accuracy measures)

random algorithm has been used to assign around 800

tasks to the available clusters and nodes randomly

without asses sing the clusters and nodes workload state s.

That is, all the jobs entered the system in a unifor m way.

Figure 6 and 7 demonstrate the distribution of the

worklo a d amon g the d i ffe re nt c lust ers and var io us no de s.

It is obvious form the charts that the workload was not

distributed in a fair manner where some nodes have

workload more than other. It is noticeable, in the

different clusters the workload was not equally

distributed among the different nodes in each cluster. For

instance, in the first cluster the nodes 2 and 5 have the

heaviest workload, whereas the nodes 1 and 20 in the

same cl uster ha ve much l ess worklo ad. In the same way,

among the different clusters, the workload was not

distributed in a balanced way as shown in Figure 7 which

shows the total number of tasks assigned to each cluster.

In Figure 7, the second and third clusters have almost

equal workload. Similarly, the cluster 1 and 4 have

almost identical load. Nevertheless, the number of tasks

assigned to cluster 2 almost double the number of tasks

assigned to the first cluster. In the same way, the

workload assigned to the forth cluster equal the half of

the work assigned to the third cluster. After that, we have

considered the case with an initial workload of 800 tasks

that have been assigned randomly to different clusters

and nodes. Then the proposed algorithm was used to

evaluate the state of each node and thereby to evaluate

the clusters workload state. Based on the available in-

formation, the arriving tasks are assigned to the clusters

and nodes that have the lo wer lo ad .

Workload Distribution : First Cluster

The different Node

Thee Number of assigned Tasks

Node 1

Node 2

Node 3

Node 4

Node 5

Node 6

Node 7

Node 8

Node 9

Node 10

Node 11

Node 12

Node 13

Node 14

Node 15

Node 16

Node 17

Node 18

Node 19

Node 20

Workload Distribution : Second Cluster

The different Node

Thee Number of assigned Tasks

Node 1

Node 2

Node 3

Node 4

Node 5

Node 6

Node 7

Node 8

Node 9

Node 10

Node 11

Node 12

Node 13

Node 14

Node 15

Node 16

Node 17

Node 18

Node 19

Node 20

Workload Distribution : Third Cluster

The different Node

Thee Number of assigned Tasks

Node 1

Node 2

Node 3

Node 4

Node 5

Node 6

Node 7

Node 8

Node 9

Node 10

Node 11

Node 12

Node 13

Node 14

Node 15

Node 16

Node 17

Node 18

Node 19

Node 20

Figure 6. The workload of nodes in different clusters at a

certain amount of time using random distribution algorithm

(after assigning 800 tasks)

Fuzzy Logic–Based Scheme for Load Balancing in Grid Services

155

Figure 8 shows the system state after using the

proposed algorith m to evalua te the worklo ad state s o f the

clusters, and to distribute the new tas ks. Clearly the tasks

are distributed in a balanced way among the different

clusters; where the number of tasks assigned to each

cluster doesn’t vary much form the others. Moreover, the

distribution of workload among the different nodes

inside the same cluster is reasonable.

Even t houg h some nodes may be assigned more work,

most of the nodes have almost identical workload. For

instance, Figure 9 demonstrates the workload states of

the nodes related to the second and third cluster. It is

obvious the workload is distributed in approximately

balanced manner among the different nodes related to

this cluster. In order to evaluate the performance of the

proposed scheme, the load balancing of the overall

system was monitored during different points of time.

Figure 10 shows the states of the different cl utters during

different times. Obviously form Figure 10, during

observed intervals of time the workload was distributed

in almost balanced way. With the passing of time,

despite of the steady increase of the arriving tasks, the

Workload Distribution : Clusters 1- 4

(Randomized Manner)

100

150

200

250

300

The different Clusters

Thee Number of assigned Tasks

Cluster 1

Cluster 2

Cluster 3

Cluster 4

Figure 7. The workload distributed randomly among different

clusters at a certain amount of time (after assigning 800

tasks).

Workload Distribution : Clusters 1- 4

after distributing around 5000

200

400

600

800

1000

1200

1400

1600

The different Clusters

Thee Number of assigned Tasks

Cluster1

Cl u s ter2

Cl u s ter3

Cl u s ter4

Figure 8. The workload states of the clusters after

distributing the tasks by applying the proposed algor ithm

workload was distributed in a balanced way. That is, the

clusters have almost ide ntical workload. Co mparin g with

the random algorithm, shown in Figures 6 and 7, the

proposed fuzzy logic-based scheme achieves more

efficiency in distributing the arriving tasks based on the

workload states of the available nodes and clusters.

7. Conclusion and Future Work

Fuzzy logic systems have the ability to deal with the

imprecision and uncertainties surrounded the input

variables and their relationships. In this paper, we

presented a new fuzzy logic-based scheme for dynamic

Workload Distribution :Nodes of Cluster 3

After Submission for around 5000 Tasks

100

The different Clusters

Thee Number of assigned Tasks

Node1

Node2

Node3

Node4

Node5

Node6

Node7

Node8

Node9

Node10

Node11

Node12

Node13

Node14

Node15

Node16

Node17

Node18

Node19

Node20

Workload Distribution :Nodes of Cluster 2

After Submission for around 5000 Tasks

The different Clusters

Thee Number of assigned Tasks

No de1

No de2

No de3

No de4

No de5

No de6

No de7

No de8

No de9

No de10

No de11

No de12

No de13

No de14

No de15

No de16

No de17

No de18

No de19

No de20

Figure 9 . The workload states o f the different nodes rel ated

to clusters 2 and 3 respectively after distributing the tasks

The Workl oad of Clust er s at di f f erent t i m es

100

1000

10000

100ms

200ms

300ms

400ms

500ms

600ms

700ms

800ms

900ms

1000ms

Time Interval

The number of assigned ta

Cluster 1

Cluster 2

Cluster 3

Cluster 4

Figure 1 0. The clusters states at different times

Fuzzy Logic–Based Scheme for Load Balancing in Grid Services

156

load balancing in the gr id comput in g services. Four input

variables have been considered in the proposed scheme

to evaluate the workload state of the presented node. W e

evaluated the performance of the proposed scheme in

terms of its ability to keep all nodes and clusters of the

overall system in a balanced way. The simulation res ult s

show that the proposed scheme achieves really

satisfactory and consistently load balancing than in other

randomized approaches. As a future work, it is possible

to use clustering for the nodes in the same cluster based

on their states since enormous number of nodes may

impair the proposed scheme performance. Another goal

is to embed this sch eme into the proposed framework of

the NSTIP project # 10-INF1381-04 in order to validate

the scalability of the proposed framework and to assess

its ability to support the load balance in a grid that hosts

of food, health and nutrition inquire services.

8. Acknowledgment

We would like to thank King Fahd University of Petro-

leum and Minerals for providing the computing faci lities.

The authors would like to acknowledge the support pro-

vided by King Abdulaziz City for Science and Technol-

ogy (KACST ) thr o ug h the Sc i e nce & T e chno log y Unit at

King F ahd University of Petroleum & Minerals (KFUPM)

for funding this work through project No.10-INF1381-04

as part of the National Science, Technology and Innova-

tion Plan.

REFERENCES

[1] C. Kesselman, S. Tuecke, “The Anatomy of the Grid:

Enabling Scalable Virtual Organizations,” International

Journal of High Performance Computing Applications,

vol. 15, iss ue3, pp 200–222, 2001.

[2] S. Sharma, S. Singh, and M. Sharma, "Performance

Analysis of Load Balancing Algorithms," World

Academy of Science, Engineering and Technology, vol.

38, 2008 .

[3] L. A. Zadeh, “Fuzzy Logic, Neural Networks, and Soft

Computing, Communications of the ACM, vol-37,

Issue-3, pp: 77-84, Mar. 1994.

[4] L. A. Zadeh, “From Computing with numbers to

computing with words-from manipulation of

measurements to manipulation of perceptions”, Int. J.

Appl. Math. Computer Sci., Vol.12, No. 3, pp. 307-324,

2002.

[5] H. T. Nguyen, N. R. Prasad, C. L. Walker, and E. A.

Walker, “A First Course in Fuzzy and Neural Control”,

CRC Press, 2002.

[6] B. Kosko, Neural Networks and F u q Systems: A

Dynamical Systems Approach to Machine Intelligence,

Prent ice Hall, New Jersey, 1992.

[7] L. S. Cheung. Load Balancing in Distributed Object

Computing Systems, M. Phil. Thesis, Department of

Electrical and Electronic Engineering, the University of

Hong Kong, August 2001

[8] S. R. Moosavi-Nejad, S.S. Mortazavi, and B. V. Vahdat,

“Fuzzy based Design and tuning of distributed systems

load balancing controller,” The 5th Symposium on

Advances in Science and Technology (SASTech), 2011.

[9] K. B Bey, F. Benhammadi, Z. Gessoum and A. Mokhtari,

"CPU Load Prediction Using Neuro-Fuzzy and Bayesian

Inferences", Journal of Neurocomputing Vol.74, pp.

1606-1616, May. 2011.

[10] A. Revar, M. And hari ya, D. Sutariya, "Load Balanci ng in

Grid Environment using Machine Learning - Innovative

Approach, " International Journal of Computer

Applications (0975 – 8887), Volume 8– No.10, October

2010

[1] K. V. Yu, and Chih-Hsun Chou, “A Fuzzy-Based

Dynamic Load-Balancing

Algorithm,” h ttp://jitas.im.cpu.edu.tw/2004-2/4.pdf

[2] M. Rantonen, Tapio Frantti, and Kauko Leivisk, "Fuzzy

expert system for load balancing in symmetric

multiprocessor systems ", Expert Systems with

Applications Journal, Vol. 37, Issue 12, pp. 8711-8720,

December 2010

[3] L. Cheung and Y. Kwok, “On load balancing approaches

for distributed object computing systems,” The Journal of

Supercomputing, vol. 27, pp. 149-175, 20 04.

[4] E. Saravanakumar and Go mathy P rathima,” A novel l oad

balancing algorithm for computational grid,

“International Journal of Computational Intelligence

Techniques, ISSN: 0976–0466 & E-ISSN: 0976–0474

Volume 1, Issue 1, 2010, PP-20-26

[5] K. Lu, R. Subrata, and A.Y. Zomaya, “An Efficient Load

Balancing Algorithm for Heterogeneous Grid Systems

Considering Desirability of Grid Sites,” Proc. 25th IEEE

Int’l Performance Computing and Comm. Conf.

(IPCCC ’06), 2006.

[6] Y. Li, Yuhang Yang and Rongbo Zhu, “A Hybrid Load

balancing Strategy of Sequential Tasks for

Computational Grids,” 2009 IEEE International

Conference on Networking and Digital Society,

pp.112-117.

[7] B. Yagoubi and Y. Slimani, “Task Load Balancing

Strategy for Grid Computing,” Journal of Computer

Science 3 (3): 186 -194, 2007.

[8] L. M. Khanli1 and Behnaz Didevar, “A New Hybrid

Load Balancing Algorithm in Grid Computing

Systems, ”Journal of Computer Science Vol-2 No 5

October, 2011.

[9] D. Ramesh and A. Krishnan, “Hybrid Algorithm for

Optimal Load Sharing in Grid Computing, Journal of

Computer Science 8 (1): 175-180, 2012.

[10] T. Helmy, F. Al-Otai bi , “Dynamic Load -Bal anci ng Based

on a Coordinator and Backup Automatic Election in

Distributed Systems“, International Journal of Computing

& Information Science, pp. 37-45, Vol. 9, No. 1, April

2011.