Journal of Network and Computer Applications 35 (2012) 1297–1303

Contents lists available at SciVerse ScienceDirect

Journal of Network and Computer Applications

1084-80

doi:10.1

n Corr

E-m

mansou

journal homepage: www.elsevier.com/locate/jnca

A dynamic replica management strategy in data grid

Najme Mansouri n, Gholam Hosein Dastghaibyfard

Department of Computer Science and Engineering, College of Electrical and Computer Engineering, Shiraz University, Molla Sadra Avenue, Shiraz, Iran

a r t i c l e i n f o

Article history:

Received 12 August 2011

Received in revised form

9 December 2011

Accepted 25 January 2012Available online 1 February 2012

Keywords:

Data grid

Data replication

Number of requests

Simulation

45/$ - see front matter & 2012 Elsevier Ltd. A

016/j.jnca.2012.01.014

esponding author. Tel.: þ98 915 3624299; fa

ail addresses: najme.mansouri@gmail.com,

ri@cse.shirazu.ac.ir (N. Mansouri).

a b s t r a c t

Data Grid provides scalable infrastructure for storage resource and data files management, which

supports several large scale applications. Due to limitation of available resources in grid, efficient use of

the grid resources becomes an important challenge. Replication is a technique used in data grid to

improve fault tolerance and to reduce the bandwidth consumption. This paper proposes a Dynamic

Hierarchical Replication (DHR) algorithm that places replicas in appropriate sites i.e. best site that has

the highest number of access for that particular replica. It also minimizes access latency by selecting

the best replica when various sites hold replicas. The proposed replica selection strategy selects the

best replica location for the users’ running jobs by considering the replica requests that waiting in the

storage and data transfer time. The simulated results with OptorSim, i.e. European Data Grid simulator

show that DHR strategy gives better performance compared to the other algorithms and prevents

unnecessary creation of replica which leads to efficient storage usage.

& 2012 Elsevier Ltd. All rights reserved.

1. Introduction

In recent years, applications such as bioinformatics, climatetransition, and high energy physics produce large datasets fromsimulations or experiments. Managing this huge amount of datain a centralized way is ineffective due to extensive access latencyand load on the central server. In order to solve these kinds ofproblems, Grid technologies have been proposed. Data Gridsaggregate a collection of distributed resources placed in differentparts of the world to enable users to share data and resources(Chervenak et al., 2000; Allcock et al., 2001; Foster, 2002;Worldwide Lhc Computing Grid, 2011).

Data replication is an important technique to manage largedata in a distributed manner. The general idea of replication is toplace replicas of data at various locations. Data replication hasbeen used in database systems (Wolfson and Milo, 1991), paralleland distributed systems (Bae and Bose, 1997; Loukopoulos et al.,2005; Rehn-Sonigo, 2007; Tzeng and Feng, 1996), mobile systems(Hara, 2001; Tu et al., 2006) and Data Grid systems (Abawajy,2004; David, 2003; Rahman et al., 2008; Ranganathana andFoster, 2001; Stockinger et al., 2001). There are three key issuesin all the data replication algorithms which are replica placement,replica management and replica selection. Placing the replicas inthe appropriate site reduces the bandwidth consumption andreduces the job execution time.

ll rights reserved.

x: þ98 711 6474605.

Each grid site has its own capabilities and characteristics; so,choosing appropriate site from many sites that have the requireddata is an important decision. The response time is an essentialparameter that influences the replica selection and thus the jobturnaround time. Replica management is the process of creatingor deleting replicas in Data Grid. To create a replica, we have toanswer some important questions, such as, which file should bereplicated? Where the file should be stored? and finally whenshould the replicas be created? Generally, replication algorithmsare either static or dynamic. In static approaches the createdreplica will exist in the same place till user deletes it manually orits duration is expired. On the other hand, dynamic strategiescreate and delete replicas according to the changes in gridenvironments, i.e. users’ file access pattern.

Meanwhile, even though the memory and storage size of newcomputers are ever increasing, they are still not keeping up withthe request of storing large number of data. Hence methodsneeded to create replicas that increase availability without usingunnecessary storage and bandwidth. In this work a novel datareplication strategy, Dynamic Hierarchical Replication (DHR) isproposed. DHR extends proposed algorithm in Horri et al. (2008)and selects best replicas when various sites hold replicas ofdatasets. The proposed replica selection strategy selects the bestreplica location for the users’ running jobs by considering thereplica requests that waiting in the storage and data transfer time.DHR also stores each replica in an appropriate site i.e. best site inthe requested region that has the highest number of access forthat particular replica. The simulated results of DHR with Optor-Sim, show that DHR outperforms over current strategies about37% and prevents unnecessary creation of replica which leads toefficient storage usage.

N. Mansouri, G.H. Dastghaibyfard / Journal of Network and Computer Applications 35 (2012) 1297–13031298

The rest of the paper is organized as follows. Section 2 gives abrief introduction of previous work on data replication for grids.In Section 3, Dynamic Hierarchical Replication (DHR) algorithm isproposed. Section 4 shows simulation results. Finally, conclusionsand future research works are presented in Section 5.

2. Related work

Recently, modeling Data Grid environments and simulatingdifferent data replication strategies as well as basic file replicationprotocols (Stockinger et al., 2001) has drawn researchers’ attention.Rahman et al. (2008) presented an algorithm for replica selectionusing a simple technique called the k-Nearest Neighbor (KNN). TheKNN rule chooses the best replica for a file using previous filetransfer logs. They also suggested a predictive way to estimate thetransfer time between sites. Accordingly, one site can request thereplica from a site which has minimum transfer time. They showedneural network (with back propagation, to train the network)predictive technique outperforms multi regression model.

Dogan (2009) evaluated the performances of eight dynamicreplication strategies under different Data Grid settings. The simula-tion results show that the file replication policy chosen and the fileaccess pattern have great influence on the real-time Grid perfor-mance. Fast Spread-Enhanced was the best of the eight algorithmsconsidered. Also, the peer-to-peer communication was indicated tobe very profitable in boosting the real-time performance.

Bsoul et al. (2011) proposed a dynamic replication strategy thattakes into account the number and frequency of requests, the sizeof the replica, and the last time the replica was requested. Thisalgorithm is a modified version of Fast Spread replication strategy(Ranganathana and Foster 2001) that holds valuable replicas whilethe other less important replicas are replaced with more importantreplicas. A dynamic threshold is used to determine if the requestedreplica should be stored at each node along its path to therequester. They claim their algorithm has better performancecomparing with Fast Spread with LRU and Fast Spread with LFU.

Zhong et al. (2010) have presented a dynamic replica manage-ment strategy. It consists of the creation strategy of dynamicreplica that can automatically increase replica based on thefrequency of the file access, the selection strategy of replica basedon the GridFTP and the replacement strategy of the replicacombining with the establishment time, number of access andfile size. They claim the proposed replica management strategyhas much better performance than the five built-in replicamanagement strategies in OptorSim.

Ranganathana and Foster (2001) have proposed six distinctreplica strategies (No Replica, Best Client, Cascading Replication,Plain Caching, Caching plus Cascading Replica and Fast Spread) formulti-tier data. They also introduced three types of localities,namely:

�

Temporal locality: The files accessed recently are much possi-ble to be requested again shortly. � Geographical locality: The files accessed recently by a client

are probably to be requested by adjacent clients, too.

� Spatial locality: The related files to recently accessed file are

likely to be requested in the near future.

These strategies evaluated with different data patterns: first,access pattern with no locality. Second, data access with a smalldegree of temporal locality and finally data access with a smalldegree of temporal and geographical locality. The results ofsimulations indicate that different access pattern needs differentreplica strategies. Cascading and Fast Spread performed the best

in the simulations. Also, the authors combined different schedul-ing and replication strategies.

Park et al. (2004) presented a Bandwidth Hierarchy basedReplication (BHR) which decreases the data access time bymaximizing network-level locality and avoiding network conges-tions. They divided the sites into several regions. Networkbandwidth between the regions is lower than the bandwidthwithin the regions. So, if the required file is placed in the sameregion, its fetching time will be less. BHR strategy has twodeficiencies, first if replica exists within the region it terminates,and second replicated files are placed in all the requested sites notthe appropriate sites.

Nukarapu et al. (2011) have proposed a data replicationstrategy that has a provable theoretical performance guaranteeand can be implemented in a distributed and practical manner.They also proposed a distributed caching strategy, which can beeasily adopted in a distributed system such as Data Grids. The keypoint of their distributed strategy is that when several replicas areavailable, each site keeps track of the closest replica. Theirsimulation result show distributed replication algorithm signifi-cantly outperforms popular existing replication strategy undervarious network parameters.

Shorfuzzaman et al. (2010) proposed Popularity Based ReplicaPlacement (PBRP) strategy in a hierarchical data grid which isguided by file ‘‘popularity’’. The ‘‘popularity’’ of a file is deter-mined by its access rate by the clients. It places replicas close toclients to decrease data access. The effectiveness of PBRP dependson the determination of a threshold value related to file popular-ity. They also proposed Adaptive-PBRP (APBRP) that calculatesthis threshold dynamically based on data request arrival rates.Their simulation results show that PBRP performs better thanother dynamic replication methods in terms of both job executiontime and average bandwidth consumption.

Khanli et al. (2011) proposed Predictive Hierarchical Fast Spread(PHFS), a dynamic replication algorithm based on Fast Spread inmulti-tier Data Grid. PHFS tries to predict user’s subsequencecomponent to adapt the replication configuration with the avail-able condition, to increase locality in access. The conceptual basisfor PHFS was the users who worked on the same context may berequested some files with high probability in the future. One of themain results is that the PHFS algorithm is appropriate for applica-tions in which the clients work on a context for some duration oftime and their requests are not random. The results of simulationalso show that PHFS has better performance and lower latencycomparing to common Fast Spread.

3. Proposed replication algorithm

In this section, first the 3-Level Hierarchical Algorithm ispresented, and then network structure is described, and finallya novel Dynamic Hierarchical Replication (DHR) algorithm isproposed.

3.1. 3-Level hierarchical algorithm

Horri et al. (2008) considered a hierarchical network structurethat has three levels. First level are Regions that are connectedthrough internet i.e. have low bandwidth. Second level comprisesLAN’s (local area network) within each region that have moder-ately higher bandwidth comparing to the first level. Finally thethird level are the nodes (sites) within each the LAN’s, that areconnected to each other with a high bandwidth.

3LHA first checks replica feasibility. If the requested file size isgreater than SE size, file will be accessed remotely. It among thecandidate replicas selects the one that has the highest bandwidth

N. Mansouri, G.H. Dastghaibyfard / Journal of Network and Computer Applications 35 (2012) 1297–1303 1299

to the requester node. If the available space in SE is greater orequal to requested file size, it replicates the file. If not, some filesshould be deleted. It deletes those files with minimum time fortransferring (i.e. only files that are exist in local LAN and localregion). This leads to a better performance comparing with LRUmethod. Although 3-layer replication makes improvement insome performance metrics, but the main weakness is by placingreplicas in all requested site not appropriate sites. The enhancedalgorithm for overcoming this limitation of the 3LHA stores thereplicated files in the site where the file is accessed for themaximum time. The storage cost and also the mean job executiontime are reduced further from 3LHA.

Moreover, in the previous work the data transfer time thatdepends on the network bandwidth is considered to predict theresponse time, but the transfer time alone is not sufficient.Indeed, the storage requests queue is other factor that playsmajor role in estimating the response time. Therefore theenhanced strategy considers transfer time and request waitingtime in the queue for replica selection.

3.2. Network structure

The grid topology of simulated platform is given in Fig. 1,which has three levels similar to what is given by Horri et al.(2008) i.e. regions’, LAN’s within each region and nodes withineach LAN.

3.3. Dynamic hierarchical replication algorithm

When a job is allocated to local scheduler, before job executionthe replica manager should transfer all the required files that arenot available. So, the data replication enhances the job scheduling

Fig. 1. Grid topology i

performance by decreasing job turnaround time. DHR algorithmhas three parts:

3.3.1. Replica selection

Generally when several replicas are available within the localLAN, the local region or other regions, DHR selects the site thathas least number of requests, since the bandwidth in each level ofthe network is fixed.

3.3.2. Replica decision

When a requested replica is not available in the local storage,replication should take place. According to the temporal and geo-graphical locality the replica is placed in the best site (BSE). To selectthe BSE, DHR creates a sorted list (by number of replica access) of allSE’s that requested that particular file in the region. Now the replicawill be placed in the first storage element (SE) of the above sorted list,i.e. BSE. If more than one SE is candidate one can be selectedrandomly. Therefore, replica is not placed in all the requested sites.Hence, storage cost as well as mean job execution time can decrease.Assume list 1 shows the sorted list created for replica R, now DHRselects site S7 from LAN3 which is shown in Fig. 2.

3.3.3. Replica replacement

If enough storage space exists in the local site, the selected file willbe replicated. Otherwise if the file is available in the local LAN, then itwill be accessed remotely. Now, if enough space for replication doesnot exist and requested file is not available in the same LAN, one ormore files should be deleted using the following rules:

�

n th

Generate a LRU (least recently used) sorted list of replicas thatare both available in the current site as well as the local LAN.Now start deleting files from the above list till space isavailable for replica.

e simulation.

Fig. 2. Replica placement strategies.

N. Mansouri, G.H. Dastghaibyfard / Journal of Network and Computer Applications 35 (2012) 1297–13031300

�

If space is still insufficient, then repeat previous step for eachLAN in current region, randomly. In other word, generate a LRUsorted list of replicas that are both available in the site as wellas the local region. � If space is still insufficient, generate a LRU sorted list of the

remaining files in the site and start deleting files from theabove list till space is available for replica. Fig. 3 describes DHRstrategy.

4. Experiments

4.1. Simulation tool

OptorSim code has been modified to implement DHR. Eur-opean Data Grid simulator i.e. OptorSim has the followingcomponents: Computing Elements (CEs) to which the job is sent;Storage Elements (SEs) where data can be kept. Resource Broker(RB), which submits jobs to grid sites according to some schedul-ing algorithms. OptorSim is shown in Fig. 4 (Cameron et al., 2004;OptorSim–A). The choices of scheduling policies for the ResourceBroker consist of the following:

�

Random scheduler selects a computing node to executespecific job randomly. � Shortest Queue scheduler calculates all of the queue length of

computing nodes and selects one that has the least number ofjobs waiting in the queue.

� Access Cost scheduler assigns the job to computing element

where the file has the lowest access cost (cost to get all filesneeded for executing job).

� Queue Access Cost scheduler selects computing element

where has the smallest sum of the access cost for the joband the access costs of all jobs waiting in the queue.

Replica Manager (RM) at each site controls data transferringand provides a mechanism for accessing the Replica Catalog. TheReplica Optimiser (RO) within the RM is responsible for theselection and dynamic creation and deletion of file replicas.

4.2. Simulation input

The OptorSim tool works based on several configuration files.

�

Parameter configuration file: The basic simulation parametersare set in the parameter configuration file such as totalnumber of jobs to be run, delays between each job submission,maximum queue size, the choice of replication strategies,access patterns for the job, etc. � Grid configuration file: contains the network topology, i.e., the

links between grid sites, the available network bandwidthbetween sites, and number of CEs and SEs, as well astheir sizes.

� Job configuration file: describes information about simulated

jobs, the files needed by each job, the probability each jobruns, etc.

� Bandwidth configuration file: specifies the background net-

work traffic.

As mentioned above, jobs requires to access files duringexecution. The order in which those files are requested isdetermined by the access pattern. Four important access patternsare as follow: Sequential (files are selected in the order stated inthe job configuration file), random (files are accessed using arandom distribution), random walk unitary (files are selected inone direction away from the previous file request and thedirection will be random) and random walk Gaussian (files arerequested in a Gaussian distribution). Sequential access pattern isused in this simulation.

4.3. Configuration

There are three regions in our configuration and each regionhas an average two LANs. Initially all master files are distributedto CERN. A master file consists of the original file and cannot bedeleted. The topology of our simulated platform includes 10 CEsand 11 SEs. The storage capacity of the master site is 300 GB andthe storage capacity of all other sites is 40 GB. Bandwidth in eachlevel is given in Table 1. There are 6 job types, and each job typeon average requires 16 files (each is 2 GB) for execution. Table 2

Fig. 3. DHR algorithm.

Fig. 4. OptorSim architecture.

Table 1Bandwidth configuration.

Parameter Value (Mpbs)

Inter LAN bandwidth (level 3) 1000

Intra LAN bandwidth (level 2) 100

Intra Region bandwidth (level 1) 10

Table 2General configuration of parameters.

Parameter Value

Number of jobs 2000

Number of jobs types 6

Number of file access per jobs 16

Size of single file (GB) 2

Job delay (ms) 2500

Maximum queue size 200

N. Mansouri, G.H. Dastghaibyfard / Journal of Network and Computer Applications 35 (2012) 1297–1303 1301

specifies the simulation parameters and their values used in ourstudy. To simplify the requirements, we assumed that the data isread-only.

4.4. Simulation results and discussion

We evaluate and compare the performance of DHR algorithmwith five replication algorithms; No Replication, Least FrequencyUsed (LFU), Least Recently Used (LRU), 3-Level HierarchicalAlgorithm (3LHA) and Bandwidth Hierarchy based Replicationalgorithm (BHR).

In No Replication strategy files are accessed remotely. Whenstorage is full, LRU deletes least recently accessed files and LFUdeletes least frequency accessed files. The BHR algorithm storesthe replicas in a site that has a high bandwidth and replicatesthose files that are likely to be requested soon within the region.The 3LHA considers a hierarchical network structure that hasthree levels. Bandwidth is an important factor for replica selectionand deletion.

In Fig. 5, the execution time of DHR is smaller than otherstrategies. Obviously, the No Replication strategy has the worstperformance as all the files requested by jobs have to betransferred from CERN. In this simulation LRU and LFU havealmost the same execution time. BHR improves data access timeby avoiding network congestions. DHR performs better than otherstrategies since it considers the differences between intra-LANand inter-LAN communication and improves the mean job execu-tion time by selecting the best replica location for execution jobswith considering number of requests that waiting in the storageand data transfer time. DHR will only replicate those files that are

Fig. 5. Mean job execution time for various replication algorithms.

Fig. 6. Mean job time based on varying number of jobs.

Fig. 7. Mean job time by varying the storage size.

Fig. 8. Mean job time based on varying size of files.

N. Mansouri, G.H. Dastghaibyfard / Journal of Network and Computer Applications 35 (2012) 1297–13031302

not available in the local LAN when available storage for replica-tion is not enough.

Figure 6 displays the mean job time based on changingnumber of jobs for 5 algorithms. DHR mean job execution timeis about 40% faster than BHR. It is clear that as the job numberincreases, DHR is able to process the jobs in the lowest mean timein comparison with other methods. It is similar to a real Gridenvironment where a lot of jobs should be executed.

Figure 7 illustrates the effect of size of storage element on themean job time. LFU and LRU are always replicate when a requestis made; hence the storage resource usage of them is high. BHRstrategy performs better than the previous two strategies since itkeeps at most one copy of file in the region. As the size of storagespace decrease in Grid sites, DHR outperforms other strategiesgreatly. If the available storage for replication is not enough,proposed algorithm will only replicate those files that are notavailable in the near sites and delete files that have less transfertime. Also instead of storing files in many sites, they can be placedin an appropriate site so storage space is saved. But if theavailable storage for replication is enough, all algorithms havethe same performance.

We continue performance evaluation with varying file sizes.Since the storage size is fixed, increasing in file size will decreasethe average number of copies of each file, so placing replicas inthe suitable sites significantly increases performance. In Fig. 8,

DHR outperforms the other methods as file size increases since itprevents from creating unnecessary replicas. It is expected thateven in the real data grid environments the difference will beeven more considerable as dataset sizes reach to many terabytes.

5. Conclusion and future work

Data Grid is the highlight in the development of the Gridtechnology, which can be treated as a suitable solution for highperformance and data-intensive computing applications. Improve-ment of data access efficiency is a main issue since number and sizeof storage devices available in grid are limited while large size ofdata files are produced. In order to solve the problem, it is a goodidea to create replicas of the files in appropriate locations.

In this paper, a new replication strategy named DynamicHierarchical Replication (DHR) for a 3 level hierarchical structurenetwork is proposed. The goal is to effectively reduce the fileaccess time due to the limited storage space of Grid sites. DHRdeletes those file that exist in local LAN (i.e. files with minimum

N. Mansouri, G.H. Dastghaibyfard / Journal of Network and Computer Applications 35 (2012) 1297–1303 1303

transfer time) when free space is not enough for the new replica.It stores the replicas in the best site where the file has beenaccessed most, instead of storing files in many sites. We alsopresented a replica selection strategy that selects the best replicalocation for the users’ running jobs by considering the replicarequests that waiting in the storage and the data transfer time. Toevaluate the efficiency of the proposed replication strategy, gridsimulator OptorSim is configured to represent a real world datagrid testbed. The simulation results show, it has less job executiontime in comparison with other strategies especially when thegrid sites have comparatively small storage size. Also, thatperformance increases as file size increases, since the proposedmethod prevents unnecessary replication by placing replica in thebest site.

In future work, DHR can be combined with a proper schedulingto improve performance. We also plan to investigate more replicareplacement strategies to further improve the overall systemperformance. Replica selection can also be extended by consider-ing additional parameters such as security.

Acknowledgments

The authors would like to thank Iran TelecommunicationResearch Center (www.itrc.ac.ir) for their financial support.

References

Allcock B, Bester J, Bresnahan J., Chervenak AL, Kesselman C, Meder S, Nefedova V,Quesnel D, Tuecke S, Foster I, Secure. efficient data transport and replicamanagement for high-performance data-intensive computing, in: Proceedingsof the First Eighteenth IEEE Symposium on Mass Storage Systems andTechnologies, 2001.

Abawajy, JH, Placement of file replicas in data grid environments, in: ICCS 2004, in:Lecture Notes in Computer Science, 2004, pp. 66–73.

Bae MM, Bose B. Resource placement in torus based networks. IEEE Transactionson Computers 1997;46(10):1083–92.

Bsoul M, Khasawneh A, Abdallah EE, Kilani Y. Enhanced fast spread replicationstrategy for data grid. Journal of Network and Compute Applications 2011;34:575–80.

Chervenak A, Foster I, Kesselman C, Salisbury C, Tuecke S. The data grid: towardsan architecture for the distributed management and analysis of large scientificdatasets. Journal of Network and Compute Applications 2000;23(3):187–200.

Cameron DG, Millar AP, Nicholson C, Carvajal-Schiaffino R, Zini F, Stockinger K.Optorsim: A simulation tool for scheduling and replica optimization in datagrids, in: International Conference for Computing in High Energy and NuclearPhysics (CHEP 2004), 2004.

David WB. Evaluation of an economy-based file replication strategy for a data grid.in: International Workshop on Agent based Cluster and Grid Computing2003:120–6.

Dogan A. A study on performance of dynamic file replication algorithms for real-time file access in Data Grids. Future Generation Computer System 2009;25:829–39.

Foster I. The grid: a new infrastructure for 21st century science. Physics Today2002;55(2):42–7.

Hara T. Effective replica allocation in ad hoc networks for improving dataaccessibility, in: Proceeding of IEEE Infocom Conference, 2001, pp. 1568–1576.

Horri A, Sepahvand R, Dastghaibyfard Gh. A hierarchical scheduling and replica-tion strategy. International Journal of Computer Science and Network Security2008;8.

Khanli L, Isazadeh A, Shishavan TN. PHFS: A dynamic replication method, todecrease access latency in the multi-tier data grid. Future Generation Com-puter System 2011;27:233–44.

Loukopoulos T, Lampsas P, Ahmad I. Continuous replica placement schemes indistributed systems, in: Proceedings of the 19th International Conference onSupercomputing, 2005, pp. 284–292.

Nukarapu DT, Tang B, Wang L, Lu S. Data replication in data intensive scientificapplications with performance guarantee. IEEE Transactions on Parallel andDistributed Systems 2011;22:1299–306.

OptorSim–A Replica Optimizer Simulation: /http://edg-wp2.web.cern.ch/edgwp2/optimization/optorsim.htmlS.

Park S, Kim J, Ko Y, Yoon W. Dynamic data replication strategy based on internethierarchy BHR, 3033. Heidelberg: Lecture notes in computer science publisher,Springer-Verlag; 2004 pp. 838–846.

Rehn-Sonigo, V, Optimal replica placement in tree networks with QOS andbandwidth constraints and the closest allocation policy, Technical Report,No. 6233, INRIA, 2007.

Rahman RM, Barker K, Alhaij R. Replica placement strategies in data grid. Journalof Grid Computing 2008;6(1):103–23.

Ranganathana K, Foster I. Identifying dynamic replication strategies for a highperformance data grid, in: Proceedings of the International Grid ComputingWorkshop, 2001, pp. 75–86.

Rahman RM, Alhajj R, Barker K. Replica selection strategies in data grid. Journal ofParallel and Distributed Computing 2008;68:1561–74.

Stockinger H, Samar A, Allcock B, Foster I, Holtman K, Tierney B. File and objectreplication in data grids, in: 10th IEEE Symposium on High Performance andDistributed Computing, 2001, pp. 305–314.

Shorfuzzaman M, Graham P, Eskicioglu R. Adaptive popularity-driven replicaplacement in hierarchical data grids. Journal of Supercomputers 2010;51:374–92.

Tzeng N-F, Feng G-L. Resource allocation in cube network systems based on thecovering radius. IEEE Transactions on Parallel and Distributed Systems1996;7(4):328–42.

Tu M, Li P, Xiao L, Yen I-L, Bastani FB. Replica placement algorithms for mobiletransaction systems. IEEE Transactions on Knowledge and Data Engineering2006;18(7):954–70.

Worldwide Lhc Computing Grid, /http://lcg.web.cern.ch/LCG/S, 2011.Wolfson O, Milo A. The multicast policy and its relationship to replicated data

placement. ACM Transactions on Database System 1991;16(1):181–205.Zhong H, Zhang Z, Zhang X. A dynamic replica management strategy based on Data

Grid, in: Ninth International Conference on Grid and Cloud Computing, 2010.