Upload
nicholas-dawson
View
6
Download
0
Embed Size (px)
DESCRIPTION
Wilkes100 2nd International Conference on Computing Sciences 15-16 November 2013
Citation preview
5/19/2018 Session 03_Paper 41
1/10
Proceedings of International Conference on Computing Sciences
WILKES100 ICCS 2013
ISBN: 978-93-5107-172-3
An intelligent middleware for multi parametric load balancing in
grid environment using AHP
Faz Mohammad1,*
, Sunita Yadav2
1M.Tech Scholar, Department of Computer Science and Engineering, AKGEC, Ghaziabad 201009, U.P, India.2Professor, Department of Computer Science and Engineering, AKGEC, Ghaziabad 201009, U.P, India.
Abstract
In Grid environment, nodes have the property to join or leave the grid at any instance of time. Due to joining and leaving
property some nodes may get overloaded and other may getunderloaded and the overall performance of the grid will degrade.
To maintain the performance a load balancing strategy is required. In this paper we are considering multi-parameters for load
balancing. Analytical Hierarchy Process (AHP)is used for decision making. In this paper we proposed an intelligent
middleware for multi-parametric load balancing(IMMLB) in grid environment using AHP.
2013 Elsevier Science. All rights reserved.
Keywords: Grid Computing, Multi-Parameters, Load Balancing, AHP, Fuzzy AHP.
1. Introduction
Grid is an implementation model of distributed computing. Grid is used to share resources like data, storage,
computing power and instrumentetc. Grid enables the Virtual Organizations (VO) to access the coordinated set of
services provided by another organization or departments of the same organization. Grid computing is a class ofHigh Performance Computing (HPC). The main motive of HPC is to enhance the computing capacity and overall
performance of a system[4, 5]. A unique feature of grid is that a node/user connected with a grid, may leave or
join the grid at any instance of time. There is no central authority to authenticate, restrict the connecting nodes
and to distribute the resources among the nodes/users. According to scale and functional grid model is of various
type like cluster, enterprises, global, compute and data grid [5, 6].
In grid model when a node leaves the grid, the application running on it will be transfer to another node [1, 11]. A
load balancing strategy is required to maintain the load among all computing nodes. Load balancing strategy is of
two types, one is static and second is dynamic load balancing [10, 12]. Making a decision for Load balancing is
very important task because it may affect the overall performance. Therefore decision is made by considering the
parameters like network, application characteristics and computing node capacity [8, 14, 15]. Load balancingdecision may automated or manual.
A dynamic load balancing strategy is best suited in Grid environment due to its dynamic nature. The overall
performance may get affected by different parameter together [13]. If a single parameter is considered for
decision making, the performance may limit because other parameters also affect the performance. E.g. only
communication delay is considered for load balancing and applications are transfer to another node, but
processing power of that node may be very low. In that condition there will be a large queue on the node and the
overall performance may limit [14]. There are various types of Grid model like cluster grid, enterprises grid,global grid, data grid, compute grid. In enterprise grid multiple projects or departments of organization can share
resources within a campus or enterprises. There is no need to any security concern and management policy
because all nodes are from same enterprises [3, 7].
* Corresponding author.
5/19/2018 Session 03_Paper 41
2/10
In this paper, we worked on Enterprise Grid Model. A dynamic environment has been created by using
GRIDSIM toolkit. With the help of GRIDSIM, four computing nodes N1, N2, N3 and N4 have generated. These
nodes may transfer their application from one node to another node. Multi-parameters namely
intercommunication delay, numbers of available resources and execution time has beenconsidered to select a bestnode for load transfer. We have generated an intelligent middleware IMMLB, which accepts the values of allthree parameters and works on dynamic load balancing strategy. IMMLB used AHP for making decision of best
node for load transfer. After transferring the load, overall system performance is analyzed by considering multi
and single parameter model.
GridSim
GridSim toolkit provides a facility to simulate grid environment. In grid environment different users,
heterogeneous resources, schedulers and brokers have been generated. It enables to create a dynamic environment
according to requirement and can apply modified algorithms. GridSim is very flexible tool and provides auction
model, data grid extension into GridSim, network extension into GridSim and can simulate an experiment with
multiple Virtual Organizations (VOs) [21, 22]. There are five layers in GridSim architecture namely grid
application, user-level middleware, core grid middleware, grid fabric software and grid fabric hardware [21]. Inour model, in place of user-level middleware, IMMLB works.
Parameters for Load Balancing
There are mainly three parameters and their sub-parameters which may participate in the decision process of load
balancing and hence may affect the performance of grid system. Network parameter has some sub-parameters
like Inter-communication delay, available bandwidth, communication link capacity and network latency. No. ofavailable processing unit, processing power and memory capacity are the characteristics of computing node.
Application can affect the performance due to its execution time, pre-emptive and non-pre-emptive
characteristics.
Automatic Decision Making: Analytical Hierarchy Process
AHP is a structured technique for dealing with complex decisions based on mathematics and psychology. AHP isdeveloped by Thomas L. Saaty in the 1970s and has been extensively studied and refined since then. The AHPprovides a comprehensive and rational framework for structuring a decision problem, for representing andquantifying its elements, for relating those elements to overall goals, and for evaluating alternative solutions [18].It is used around the world in a wide variety of decision situations, in fields such as government, business,industry, healthcare and education [17]. AHP develops priorities for different alternatives and according tocriteria judges the alternatives. Initially priorities are sets according to importance to achieve the goal, after that
priorities are derived for the performance of the alternatives on each criteria, these priorities are derived based onpair-wise assessments using judgments, or rations of measurements from a scales if one exists[16].
AHP has been used in many areas due to the ability to rank choices in the order of their effectiveness in meeting
conflicting objectives [19]. Researches has been successfully employed AHP in selection of one alternative from
many; resource allocation; forecasting; total quality management; business process re-engineering; quality
function deployment, and the balanced scorecard. AHP is a better method, where the parameters are categories
into sub parameters [9].
2. Proposed model
Intelligent middleware IMMLB is implemented on a central server located in an organization as shown in figure
1. All four nodes pass the values of all three parameter which we are considering for load balancing decision.
Middleware processed the preference value of parameters entered by user and generates the score for each node.The node with highest score will be best for load transferring. IMMLB is based on dynamic load balancing
strategy.IMMLB works in different steps as follows grid formation, parameters passing, intelligent middleware
processing and decision making.
5/19/2018 Session 03_Paper 41
3/10
An intelligent middleware for multi parametric load balancing in grid environment using AHP
Fig 1. Working of IMMLB model
2.1. Grid Formation
An Enterprise Grid Model has been created by using GRIDSIM toolkit. With the help of GRIDSIM,four computing nodes N1, N2, N3 and N4 have generated. Each node has some resources and application running
on it.These nodes may transfer their application from one node to another node.Figure 2 shows the generation of
resource on node N1 having resource ID: 5 and No. of parallel elements: 13.
Fig 2. Grid Formation
3. Parameters Passing
In this paper, as the considered parameters are changed, load balancing decision is also changed accordingly.
According to the environment, preference of parameter may be changed. In IMMLB, we are considering threeparameters namely Inter-communication delay (ICD) from network parameters, No. of available resources(NOAR) from communication node characteristics and execution time (E.T.) from application characteristics.
Our model makes decision according to the value of parameters. User can change the preference of parameters as
the values of parameters changes. The values of parameters are passing to the middleware.
After the generation of gird, middleware accepts the value of parameters namely ICD, NOAR and E.T. GridSim
provides the value of all three parameters as follows:
Communication Delay
user.body(): Sending Message_0, at time = 2.000 sec
test.body(): receive Message_0, at time = 2.002 sec
user.body(): Receives Ack for Message_0
total comm. Delay between middleware and node = 2.002-2.000 = 2ms
5/19/2018 Session 03_Paper 41
4/10
Execution Time
[TimeSlot={startTime=2.0, finishTime=2.14748364}]
total execution time = 0.147448364 sec
No. of Available ResourcesCreating a grid user entity with name = User_0, and id = 17
User_0:Creating 3 Gridlets
User_0:ReceivedResourceCharacteristics from Resource_1, with id = 9
GridletID-STATUS-ResourceID-Cost
0-SUCCES-9-27.8514588 ms
4. Intelligent Middleware Processing
Middleware accepts the values of parameter form GRIDSIM, and according to the preference it calculatesthe highest score. This calculation is done using AHP. Figure 3 shows the procedure of AHP on middleware.All nodes pass the value of E.T., NOAR and ICDto middleware. Middleware accepts the values of all three
parameters and compare these values. According to the environment and values of parameters, user enters
preference values to the middleware. After accepting the preference values, AHP processing starts. As shownin Figure 3, initially a weight is assigned to each criterion and then a best node according to each criterion isselected. Finally the node having the highest sore according to criterion and option will be selected as a bestnode for transferring the load.
Fig 3. Intelligent Middleware Processing
5. Decision Making
IMMLB makes a decision to select the best node for load transferring. Decision is made according to AHP
processing. AHP works in three steps as follows:
5.1. Develop the weight for parameters
To generate the weight for each criterion we have done a survey. The survey has done on forty two industrial
persons from different industries like Berizon Data service, CTS, Cybage S/W, Endeavour s/w Ltd, Ernst &
Young Ltd, Fuzitsu, IBM, Info Cept, M-Hance India Ltd, NCR Cooperation, Radius Infratel Ltd, Reppify, Syntel
Ltd, TCS, Tech at Hcentive, Vinsol and Wipro. In the survey industrial persons were asked to give preference
value to the three parameters. We have received pair-wise comparison values of parameters. Preference values
have scaled on three levels as shown in table 1.
5/19/2018 Session 03_Paper 41
5/10
An intelligent middleware for multi parametric load balancing in grid environment using AHP
Table 1: Preference Values
S. No. Preference Preference Value
1 Less Preferable < 1
2 Equally Preferable = 1
3 Highly Preferable > 1
After scaling the values, a comparison matrix has calculated as shown in table 2.
Table2: Comparison Matrix
E.T. NOAR ICD
E.T 1.000 0.500 1.000
NOAR 2.000 1.000 2.000
ICD 2.000 0.700 1.000
As middleware get the pair wise comparison values, AHP starts processing on it and generates the priority vectorfor each criterion. There are three steps to calculate priority vectors, which are as follows:
a. Calculate 3rd
root of product.
b. Priority Vector = value of 3rd
root of product/ sum of 3rd
root of product.
c. Add the value of each column which is called Sum.
Priority vector for each criterion calculated by first step of AHP, according to pair wise comparison is shown in
table 3.
Table 3: Develop the Weight for Criteria.
6. Develop the rating for each node for each parameter
Using step two of AHP, intelligent middleware IMMLB calculates priority vector for each node on the basis of
each parameter. Rating of each node for Execution Time is shown in table 4, Rating of each node for NOAR is
shown in table 5 and Rating of each node for ICD is shown in table 6.
Table 4: Rating of Each Node for Execution Time
E.T NOAR ICD 3rd Root of product Priority Vector
E.T 1.000 0.500 1.000 0.794 0.227
NOAR 2.000 1.000 2.000 1.588 0.453
ICD 2.000 0.700 1.000 1.119 0.320
Sum 5.000 2.200 4.000 3.501 1.000
5/19/2018 Session 03_Paper 41
6/10
N1 N2 N3 N43rd root of
productPriority Vector
N1 1.000 1.000 0.733 0.640 0.777 0.280
N2 1.000 1.000 0.734 0.874 0.862 0.310
N3 0.733 0.734 1.000 0.874 0.777 0.280
N4 0.641 0.642 0.874 1.000 0.360 0.130
Sum 3.374 3.376 3.341 3.388 2.776 1.000
Table 5: Rating of Each Node for NOAR
N1 N2 N3 N43rd root of
productPriority Vector
N1 1.000 0.923 0.769 0.846 0.844 0.185
N2 1.083 1.000 0.833 0.917 0.827 0.181
N3 1.300 1.200 1.000 1.100 1.716 0.377
N4 1.181 1.090 0.909 1.000 1.170 0.257
Sum
4.564 4.213 3.511 3.863 4.557 1.000
Table 6: Rating of Each Node for ICD
N1 N2 N3 N43rd root ofproduct
Priority Vector
N1 1.000 1.500 1.000 2.000 1.442 0.356
N2 0.667 1.000 0.667 1.333 0.593 0.146
N3 1.000 1.500 1.000 2.000 1.442 0.356
N4 0.500 0.750 0.500 1.000 0.572 0.142
Sum 3.167 4.750 3.167 6.333 4.049 1.000
7. Calculation of the weighted average rating for each node, chose the one with highest score
In the final step of AHP, IMMLB calculates the score for each node according to each parameter by givenformula. Node having the highest score is the best option for load transfer.
Final scores for each node according to each parameter is shown in table 7. The node N3 with highest score wins
according to AHP calculation and IMMLB processing. Node N1 considered to be overloaded. The application
can be transferred from node N1 to node N3.
5/19/2018 Session 03_Paper 41
7/10
An intelligent middleware for multi parametric load balancing in grid environment using AHP
Table 7: Final Scores for Each Node According to Each Parameter.
Criteria
Options
E.T
0.277
NOAR
0.453
ICD
0.320
Score
1.000
N1 0.280 0.185 0.356 0.261
N2 0.310 0.181 0.146 0.199
N3 0.280 0.377 0.356 0.348
N4 0.130 0.257 0.142 0.192
SUM 1.000 1.000 1.000 1.000
8. Results and Analysis
As IMMLB found that node N3 is the best node for load transfer, the load canbetransferred from N1 to N3 to
get higher performance. We get total execution time = 2.169 ms after transferring the load from N1 to N3 asshown in figure 4. Total execution time varies according to environment.
Figure 4.Snapshot of total execution time in IMMLG Model.
For analyzing the performance of multi-pametric load balancing, the total execution timeafter load transfer iscompared with the total execution time of single parametric models. Total execution time of ICD, NOAR and E.T
are 2.466 ms, 2.941 ms and 3.061 ms respectively. Only single parameter Inter-communication delay, Number of
available resources and Execution time are considered for load balancing decision in ICD, NOAR and E.Tmodels respectively.
As the total execution time varies according to environment and load, we have calculated the total execution time
at ten different times and averaged the results for finding the averaged total execution time.The comparison of
average total execution time of IMMLB model with the single parametric models is shown in table 8.
Table 8: Average Execution Time for IMMLB, ICD, NOAR and ET
5/19/2018 Session 03_Paper 41
8/10
Total ExecutionTime
Calculation Slot
IMMLG ICD NOAR ET
TIME 1 2.169 2.466 2.946 3.061
TIME 2 2.035 2.600 2.710 2.192
TIME 3 2.240 2.411 2,566 3.002
TIME 4 2.240 3.100 2.842 2.192
TIME 5 2.005 2.210 2.901 2.171
TIME 6 2.135 3.130 2.533 2.361
TIME 7 2.180 2.630 2.802 2.502
TIME 8 2.130 2.503 2.792 2.220
TIME 9 2.005 2.311 2.601 3.080
TIME 10 2.118 2.402 2.531 2.310AVERAGE 2.126 2.576 2.722 2.509
Table 8 shows that the average total execution time of IMLLB is less as compare to ICD, NOAR and E.T. Hence
considering multiple parameters for decision making for load balancing improves the overall performance as
compare to decision making using single parameter. The graph of total execution time of IMLLB, ICD, NOAR
and ET at ten different time slots is shown in figure 5.
Fig 5.Total execution time of IMLLB, ICD, NOAR and ET at ten different time slots
The performance graph of IMMLB is smoother than single parametric model as shown in figure 5. It depicts that
if multiple parameters are considered for decision making criterion in load balancing, the total execution time of
the grid is less and more stable. Hence the overall performance of the grid in terms of execution time will beincreased.
9. Conclusion And Future Work
The overall performance of IMMLB is better than single parametric model. We are considering three parameters
which can affect the overall performance. In IMMLB, the decision for load balancing is made by considering
overall effect of all three parameters. While in single parametric model, other parameters may restrict the
performance due to its effect.In future work, we can enhance the performance of other grid models and also by considering some other
parameters according to environment. To minimize the decision time we can use fuzzy AHP in place of
0
0.5
1
1.5
2
2.5
3
3.5
Time 1 Time 2 Time 3 Time 4 Time 5 Time 6 Time 7 Time 8 Time 9 Time 10
IMMLB
ICD
NOAR
E.T.
5/19/2018 Session 03_Paper 41
9/10
An intelligent middleware for multi parametric load balancing in grid environment using AHP
traditional AHP. Fuzzy AHP uses micro functions by which criterion can be compare with minimum difference
and the decision for load balancing may be more refined [20].
References
[1] Wen-Chung Shih, Chao-Tung Yang and Shain-Shyong Tseng, Performance-Based data distribution for data mining applicationon grid computing environment, Springer Journal Supercomputer (2010) 52: 171-198.
[2] Nimatya Roy and Sajal K. Das, enhancing availability of grid computational services to ubiquitous computing applications IEEETransaction On Parallel And Distributed Systems, Vol.20, No. 7, July 2009.
[3] William E. Johnston, The Role Of Computational And Data Grid In Large-scale Science And Engineering Lawrence BerkeleyNational Lab and NASA Ames Research Center- 2009.
[4] Dr. D. Janakiram (head of distributed computing lab- IIT Madras) A Book- Grid computing- 2011.
[5] Jens Mache and Amy Apon, Teaching grid computing : topics, exercises , and experiences, IEEE TRANSACTION ONEDUCATIONAL, VOL. 50, NO. 1, FEBRUARAY 2007.
[6] Henri Casanova, Distributed Computing Research Issues In Grid Computing Department Of Computer Science AndEngineering. University Of California At San Diego, La Jolla, Ca 92093-01114- 2007.
[7] Prabhat Kr. Srivastava, Sonu Gupta Et al, Improving Performance In Load Balancing Problem On The Grid Computing SystemInternational Journal Of Computer Application(0975-8887 Volume 16 No. 1 February 2011.
[8] Hong-Ha Nguyen, Mohan Gurusamy and Luying Zhou, Scheduling Network and Computing Resources for Sliding Demands inOptical Grids, journal of lightwave technology, vol. 27, no. 12, june 15, 2009.
[9] Thomas L. Saaty, decision making with the analytic hierarchy process Int. J. Services Sciences, Vol. 1, No.- ,2008.
[10] Belabbas Yagoubi and Yahya Slimani, Dynamic Load Balancing Strategy for Grid Computing, World Academy of Science,Engineering and Technology 19 2006.
[11] Krzysztof Kurowski Et al, scheduling jobs on grid Multicriteria approach computational methods in science and technology12(2), 123-138 (2006).
[12] Francesco Palmieri, Network Aware Scheduling For Real Time Execution Support In Data-Intensive Optical Grids FutureGeneration Computer System ELSEVIER 25(2009) 794-803.
[13] Jang UK In, Soocheol Lee, Seungmin Rho and Jong Hyuk Park, Policy-Based Scheduling and Resource Allocation forMultimedia Communication on grid Computing Environment, IEEE system journal vol. 5 no. 4, december 2011.
[14] Luiz F. Bittencourt and Edmundo R. M. Madeira, A performance-oriented adaptive scheduler for dependent tasks on grids,concurrency andcomputation: practice and experience Concurrency Computat.: Pract. Exper.2008;20:10291049 Published online
22November 2007.
[15] Brain L. Tierney et al, A Network-Aware Distributed Storage Cache for Data Intensive Environment, computer sciencedirectorate Lawrence Berkeley national laboratory university of California, Berkeley, CA,94720-2002.
[16] Palcic I. Lalic , analytical hierarchy process as a tool for selecting and evaluating projects Int j simul model 8(2009) 1, 16-26ISSN 1726-4529.
[17] Jiaqin Yang et al applying analytic hierarchy process in firms overall performance evaluation: a case study in chinainternational journal of business, 7(1), 2002 ISSN:1083-4346.
[18] Doraid Dalalah, application of the analytic hierarchy process (AHP) in multi-criteria analysis of the selection pf cranes Jordanjournal of mechanical and industrial engineering vol-4, Nov 2010 ISSN 1995 6665 pages 567 578.
[19] C.K. Kwong and H. Bai, A Fuzzy AHP approach to the determination of importance weights of customer requirments in qualityfunction deployment, Journal of Intelligent Manufacturing, 13, 367-377, 2002.
[20] Fatma Tiryaki and Beyza Ahlatcioglu, Fuzzy portfolio selection using fuzzy analytical hierarchy process, ELSEVIER,information science 179 (2009) 53-69.
[21] Rajkumar Buyya et al,A tool for modeling and simulating Data Grids: An extension of GridSim, CONCURRENCY AND
COMPUTATION: PRACTICE AND EXPERENCE, pract. Exper 0123;34:1-2011.[22] Rajkumar Buyya et al, Constructing A Grid Simulation with Differentiated Network Service Using GridSim, A review IIT
MADARS, distributed computing lab-2012.
5/19/2018 Session 03_Paper 41
10/10
Index
TTest case repository (TCR), 302303
WW-shaped framework, 298
ACO classifier and selector, 303for multi-faceted test case classification and selection, 299300
fuzzy synthesis based classifier and selector, 303multi-faceted test cases optimization, 302
objectives, 301302
optimization stopping criteria, 302parameters and stopping criteria identification, 301302
TCR, 302303