Enhanced GridSim architecture with load balancing

J Supercomput (2011) 57:265–275DOI 10.1007/s11227-010-0402-6

Enhanced GridSim architecture with load balancing

Kalim Qureshi · Attiqa Rehman · Paul Manuel

Published online: 23 March 2010© Springer Science+Business Media, LLC 2010

Abstract Grid is a network of computational resources that may potentially spanmany continents. Maximization of the resource utilization hinges on the implemen-tation of an efficient load balancing scheme, which provides (i) minimization of idletime, (ii) minimization of overloading, and (iii) minimization of control overhead. Inthis paper, we propose a dynamic and distributed load balancing scheme for grid net-works. The distributed nature of the proposed scheme not only reduces the commu-nication overhead of grid resources but also cuts down the idle time of the resourcesduring the process of load balancing. We apply the proposed load balancing approachon Enhanced GridSim in order to gauge the effectiveness in terms of communicationoverhead and response time reduction. We show that significant savings are deliveredby the proposed technique compared to other approaches such as centralized loadbalancing and no load balancing.

Keywords Grid computing · Enhanced GridSim architecture · Load balancing ·Communication overhead · GridSim

1 Introduction

The explosion of internet, the computational power of personal computers and high-speed computer network gives birth to the idea of a grid of computational resources.Grid is a large distributed system that may span many continents. The idea behindthis huge system is to build a network of computational resources spreading around

K. Qureshi (�) · P. ManuelDepartment of Information Science, Kuwait University, Safat 13060, State of Kuwaite-mail: [email protected]

A. RehmanDepartment of Computer Science, COMSATS Institute of Information Technology, Abbottabad,Pakistan

mailto:[email protected]

266 Qureshi et al.

the globe and to utilize even the tiny computational power [1, 2]. It is a flexibleand coordinated resource sharing among dynamic collection of diverse users and re-sources [3]. It allows the large computational hungry applications like those of cos-mology, molecular biology, bioinformatics, gene decoding, etc. to use these resourcesto run [4].

A distributed system consists of policies for the use of the resources and the re-sources itself. Among the policies are load balancing, scheduling, fault tolerance, etc.Although Grid belongs to the class of distributed systems, traditional policies of thedistributed system cannot be applied to Grid. There are some differences betweenthe Grid and the traditional distributed systems. Some of these are nondedicated na-ture of resources, different administrative controls of different resources, unreliablecommunication links, large communication latencies, local load, etc. [5–8].

Grid Load Balancer [7] attempts to distribute the computational load among theavailable Grid processors with a goal of performance optimization. Performance op-timization leads to maximize the grid throughput, i.e., maximizing the number of jobscompletion and minimizing the response time of individual job [8]. A load balancingpolicy tries to maximize the resource utilization. The characteristics of an optimalload balancing policy are to minimize the idle time, minimize overloading, and min-imize control overheads [9]. The basic steps for load balancing [10, 11] are depictedin Fig. 1.

The taxonomy of Grid load balancing [11] in Table 1 classifies the schemes ontwo broad classes, type and granularity. The first category contains the static anddynamic schemes, while the second category contains application level and systemlevel scheduling schemes.

In this paper, we will discuss the design and development of an optimal load bal-ancing strategy for Grid. Use of Grid simulators is common [12] to analyze the effec-tiveness of a scheme. We choose GridSim [13], a popular Grid simulator due to itsextendibility and ease of use. We implement the tree-based Grid structure presented

Fig. 1 Basic load balancing steps

Table 1 Classification of load balancing schemes

Static Dynamic

Type Requires prior knowledge No prior knowledge required

Cannot cope with changes Can cope with changes

No overhead of information gathering Information gathering overhead

Application System

Granularity Minimizes the individual task’s response time Maximizes the resource utilization

Enhanced GridSim architecture with load balancing 267

in [10, 14–16] on GridSim as our test-bed and then implement the load balancing onenhanced GridSim (LBEGS) scheme on this test bed. The rest of the paper is orga-nized as follows. In the next section, we present the proposed load balancing scheme.In Sect. 3, we present the results and discussions. Finally, Sect. 4 concludes the paper.

2 Load balancing schemes for Grid environment

GridSim [13] is a famous Java-based grid simulator with a clear focus on Grid-basedeconomy and provides resource producer and resource consumer to bargain eachother as in a real marketplace. This simulator is based on entities: grid users, bro-kers bargaining on behalf of users and resources. These entities can have customizedcharacteristics. GridSim is generally used to study the economy-based schedulingdecisions taken by distributed brokers with competing requirements of budget anddeadlines.

2.1 Enhanced GridSim

We introduce some enhancements in the GridSim. Machine entity (ME) is treatedas dump entity object in GridSim 4.0 and is not able to participate in any decisionmaking activities. We propose that the machine entity should be active and shouldparticipate in load balancing at its level. We name this enhancement in GridSim theEnhanced GridSim (EGridSim). See Fig. 2.

Fig. 2 Hierarchical structure of Grid

268 Qureshi et al.

Each layer is explained in details here. Grid Broker is the top manager of a grid en-vironment which is responsible for maintaining the overall grid activities of schedul-ing and rescheduling. It gets the information of the work load from grid resources. Itsends the tasks to resources for optimization of load. Resource is next to grid Brokerin the hierarchy. It is responsible for maintaining the scheduling and load balancingof its machines. Also, it sends an event to grid broker if it is overloaded. Machineis a Processing Entity (PE) manager. It is responsible for task scheduling and loadbalancing of its PEs. Also, it sends an event to resource if it is overloaded.

2.2 Load balancing on Enhanced GridSim (LBEGS)

Our proposed load balancing (LBEGS) scheme works at three levels: Broker, Re-source, and Machine. When a new job arrives at a machine, it submits it to a PE,which is lightly loaded. Also, after a specific interval of time, it checks the load ofall PEs and classifies them: lightly loaded, over loaded, normal. We shall first presentour proposed scheme. Then two other schemes are presented with which we havecompared the performance of our scheme.

The proposed load balancing scheme is simulated on the enhanced GridSim. Theload balancing algorithm is presented in Algorithm 3 (Fig. 3). Algorithms 4, 5, and 7(Figs. 4, 5, 7) show the PE level load calculation, machine level load calculation, andresource load calculation, respectively. Algorithms 6 and 8 (Figs. 6, 8) show resourcelevel load balancing and broker level load balancing, respectively. In the followingalgorithms, PE level load is denoted by ∂ , machine level load is denoted by η, andthe resource level load is denoted by ρ.

Fig. 3 Algorithm 3 pseudo code for LBEGS scheme


In the load balancing on GridSim LBGS, load balancing is performed only at theBroker level, while in LBEGS, the load balancing is performed at broker, resource,and machine levels. Another improvement is that two or more load balancers can runsimultaneously.

Fig. 4 Algorithm 4 pseudo code for Processing Entity (PE) load calculation

Fig. 5 Algorithm 5 pseudo code for machine load calculation

Fig. 6 Algorithm 6 pseudo code for resource level load balancing scheme

270 Qureshi et al.

Fig. 7 Algorithm 7 pseudo code for resource load calculation

Fig. 8 Algorithm 8 pseudo code for Grid broker level load balancing

3 Results and discussion

There are three schemes: WLB (Without Load Balancing), LBGS, and LBEGS. Thealgorithms related to LBEGS are discussed in Sect. 2. The load balancing schemeof GridSim (LBGS) is similar to Algorithm 8, and hence it is not listed here. Thealgorithms related to WLB and LBGS are available in Refs. [17, 18].

All the algorithms are simulated in GridSim 4.0. The same configuration of re-sources is used in these three simulations. We use Windows 2000 Professional withSP2 on an Intel PIV 2.4 GHz, with 256 megabytes of RAM and 40 GB of hard disk.The simulator for these experiments is GridSim version 4.0 with JDK 1.5 updates 6and JRE 1.3.

The readings are taken by varying number of PE starting at 10 and ending at50 with a step of 10. The communication overheads are calculated by counting thenumber of messages over Internet, LAN, and Machine. The following thresholds are


Fig. 9 Communicationoverheads with constant gridlets

used in this experiment [19, 20]:

3.1 Communication overhead (LBGS and LBEGS) with constant gridlets

The results presented here are for the LBGS and LBEGS. Since no load balancingis done in WLB, it is not included. The readings are taken by varying number of PEstarting at 10 and ending at 50 with a step of 10, and the numbers of gridlets arekept constant at 100. The communication overheads are calculated by counting thenumber of messages over Internet, LAN, and Machine. The readings are the averageof 5 runs. The graph in Fig. 9 shows that communications overhead of LBEGS islower than that of LBGS when the gridlets are kept constant. It is because most ofload balancing decision are made locally at the Resource and Machine levels andGrid Broker is not contacted. This reduces the number of control messages to Brokerand thus minimizes the communication overheads.

3.2 Communication overhead (LBGS and LBEGS) with constant PEs

The results presented here are for the LBGS and LBEGS. Since no load balancingis done in WLB, it is not included. The readings are taken by varying number ofgridlets starting at 50 and ending at 250 with a step of 50 and the numbers of PEsare kept constant at 20.The communication overheads are calculated by counting thenumber of messages over Internet, LAN, and Machine. As it is observed in Fig. 10,communication overheads of LBEGS are lower than that of LBGS. It is due to thedistributed nature of the loading balancing. Another reason is that the communicationmedia among the Resource and Machine is bus which has a low communicationdelay.

272 Qureshi et al.

Fig. 10 No of gridlets vs.communication overhead

Fig. 11 Response time vs.number of gridlets

3.3 Response time (WLB, LBGS and LBEGS) with constant PEs

The results presented here are for the WLB, LBGS, and LBEGS. The readings aretaken by varying number of gridlets starting at 250 and ending at 50 with a step of50, and the numbers of PEs are kept constant at 20. The response time is the totalsimulation time.

The graph of Fig. 11 shows that LBEGS has a lower response time. It is because(i) the gridlets are assigned to a resource considering the current load, (ii) imbalancedloads are mostly handled locally.

WLB has a higher response time, as it uses a nonpreemptive scheduling policy andthe current workload of a resource is not considered. The response time of the LBGSlies between WLB and LBEGS. Although LBGS uses the preemptive scheduling pol-icy and considers the current workload of a resource like LBEGS, the communicationoverheads make the response time higher.

3.4 Percentage response time gain (LBGS and LBEGS) with constant PEs

The results presented here are for LBGS and LBEGS. The readings are taken byvarying number of gridlets starting at 250 and ending at 50 with a step of 50.


Fig. 12 Percentage responsetime gain

Fig. 13 Response time vs.number of PE

When seeing the graph of the Fig. 12, it is evident that the gain in response time ishigher in LBEGS, as compared to the other two.

3.5 Response time (WLB, LBGS, and LBEGS) with constant gridlets

The results presented here are for all the three schemes. The readings are taken byvarying number of PE starting at 10 and ending at 50 with a step of 10. The numbersof gridlets are kept constant at 100. The response time is the total simulation time.As it is displayed in Fig. 13, it is evident that the gain in response time is better inLBEGS as compared to the other two.

3.6 Response time gain (LBGS and LBEGS) with constant gridlets

The results presented here are for all the three schemes. The readings are taken byvarying number of PE starting at 10 and ending at 50 with a step of 10. The numbersof gridlets are kept constant at 100. The response time is the total simulation time.Figure 14 shows that the gain in response time is higher in LBEGS as comparedto LBGS. The improvement of LBGS is because of the consideration of the currentworkload of the resource.

274 Qureshi et al.

Fig. 14 Percentage responsetime gain vs. PE

4 Conclusion

In this paper, we propose a dynamic, distributed load balancing approach for a Grid,wherein machines and resources, in addition to the grid broker, also participate inthe load balancing operations. The outcome is reduced communication overhead andreduced response time, as the amount of information passed to the upper level en-tities is typically limited due to the local load balancing performed at lower levelsin the hierarchy. The idle time of processing entities also decreases, as load balanc-ing is effected much faster. The experimental results also justify the efficacy of theproposed approach over other schemes that are based on no load balancing or cen-tralized load balancing. The approach with no load balancing results in more idletime for the processing entities, while the centralized load balancing approach suf-fers from increased communication overhead and thus increased response time, dueto the constant involvement of the grid broker.

Our future research plan includes the implementation of various other schedul-ing, load balancing, and fault tolerance techniques on enhanced GridSim with a treestructure. Also, we would like to investigate the outcome of integrating the proposedload balancing scheme (LBEGS) with different middleware, such as GLOBUS orLEGION.

References

1. Nazir B, Qureshi K, Manuel P (2008) Adaptive fault tolerant job scheduling strategy for economybased grid, J Super Comput (October): 116–134

2. Qureshi K, Hussain SS (2008) A comparative study of parallelization strategies for fractal imagecompression on a cluster of workstations. Int J Comput Methods 5(3):463–482

3. Rehman A, Qureshi K, Manuel P, Rashid H (2008) Resource topology aware GridSim: a step ahead.J Comput 19(2):13–22; special issue on Grid and Cluster Computing

4. Foster I, Kesselman C, Tuecke S (2001) The anatomy of the Grid: enabling scalable virtual organiza-tion. Int J High Perform Comput Appl 15(3):200–222

5. Berman F, Fox G, Hey AJ (2003) Grid computing: making the global infrastructure a reality. Wiley,New York

6. Hussain S, Qureshi K, Rashid H (2009) Local predecimation with range index communication paral-lelization strategy for fractal image compression on a cluster of workstations. Int Arab J Inf Technol6(3):293–296


7. Arif Wani M, Arabnia HR (2003) Parallel edge-region-based segmentation algorithm targeted at re-configurable multi-ring network. J Supercomput 25(1):43–63

8. Bhandarkar SM, Arabnia HR (1995) The Hough transform on a reconfigurable multi-ring network.J Parallel Distrib Comput 24(1):107–114

9. Qureshi K, Manuel P (2007) Adaptive pre-task scheduling strategy for heterogeneous distributed ray-tracing system. Int J Comput Electr Eng 33:70–78

10. Lu K, Subrata R, Zomaya AY (2006) An efficient load balancing algorithm for heterogeneous gridsystems considering desirability of grid sites. In: Proceedings of the 25th IEEE international confer-ence on performance, computing, and communications, 10–12 April 2006

11. Fernandes de Mello R, Senger LJ, Yang LT (2006) A routing load balancing policy for grid com-puting environments. In: Proceedings of the 20th international conference on advanced informationnetworking and applications (Aina’06), vol 1, 18–20 April, 2006

12. Touzene A, Al Yahia S, Day K, Arafeh B (2005) Load balancing grid computing middleware. In:IASTED international conference on web technologies, applications, and services (WTAS 2005),Calgary, Canada, July 2005

13. Krauter K, Buyya R, Maheswaran M (2002) A taxonomy and survey of grid resource managementsystems for distributed computing. Softw Pract Exp 32:135–164

14. Yagoubi B, Slimani Y (2007) Task load balancing strategy for grid computing. J Comput Sci 3(3):186–194

15. Li Y, Lan Z (2005) A survey of load balancing in grid computing. In: Lecture notes in computerscience, vol. 3314. Springer, Berlin, Heidelberg, pp 280–285

16. Quetier B, Cappello F (2005) A survey of Grid research tools: simulators, emulators and real lifeplatforms. In: Proceeding of 17th IMACS world congress (IMACS 2005), Paris, France

17. Murshed M, Buyya R, Abramson D (2001) GridSim: A toolkit for the modeling and simulation ofglobal grids. Technical Report, Monash, CSSE

18. Yagoubi B, Slimani Y (2006) Dynamic load balancing strategy for grid computing. Trans Eng ComputTechnol 13:260–265

19. Yagoubi B, Slimani Y (2007) Load balancing strategy in grid environment. J Inf Technol Appl1(4):285–296

20. Payli RU, Yilmaz E, Ecer A, Akay HU, Chien S (2006) DLB—a dynamic load balancing tool for gridcomputing, scalable computing. Pract Exp 7(2):15–23

Enhanced GridSim architecture with load balancing

Documents