You are on page 1of 10

Adaptive Data Replication with Load Balancing in Data Grid

Niyoyita Jean Paul, Dong Shou Bin

Adaptive Data Replication with Load Balancing in Data Grid


1
Niyoyita Jean Paul, 2Dong Shou Bin
1,
School of Computer Science and Engineering, South China University of Technology,
niyoyitajp@gmail.com
2,
School of Computer Science and Engineering, South China University of Technology,
sbdong@scut.edu.cn

Abstract
A distributed computing environment such as a grid computing involves large scale resource
sharing among collaborators located in geographically dispersed areas. Dynamically choosing when,
where and how to replicate data and coordinating the replica access accordingly is an open problem
and is becoming more relevant with the advances of cloud computing systems. In this paper we present
an Adaptive Data Replication for Load Balancing in Data Grid (ADRLB). The method solves access
concurrency on a replica by analyzing the replica access frequency in an adaptive way. It dynamically
creates additional replicas and re-balance other data in the cluster in case of a sudden high demand
for a particular file. Adapted to an environment with constrained storage space with variable
availability, ADRLB strategy gives better performance compared to the other algorithms while
guarantying the low network usage and balanced load.

Keywords: Data grid, Data availability, Data replacement

1. Introduction
In a distributed system, data objects are shared by different users from different locations. File
replication is an effective technique for reducing data access overhead and for assuring data availability.
Maintaining an optimal distribution of replicas implies that the system must be able to modify the
geographic location of data files.
There are various ways of replicating data objects [1][2] and the choice of which type depends on
the goal of replication. Data replication also has different goals such as data sharing, high availability,
fault tolerance, hot-site backup, or a combination of some of these goals. On the other hand it is a
challenge to keep data copies with consistency and integrity. Furthermore, to fully take advantage of
the processing power of all replicas, adaptive load-balancing schemes are needed. Dynamically
choosing when, where and how to replicate data and coordinating the replica access accordingly is an
open problem and is becoming more relevant with the advances of cloud computing systems.
The existing replication architectures are classified as centralized, group based architecture, and
distributed approaches. In a centralized mechanism, a dedicated entity has a global knowledge of the
resources state in the whole network and it is in charge of replica placement and load balancing [3].
The centralized option is unsuitable for geographically distributed systems, such as the cloud or more
in general massively distributed systems, since no entity can have global and updated information
about all system resources. Indeed, communication overhead required to share the resource state
information is not negligible and the delay to achieve state information from remote nodes could lead a
centralized resource manager to a very inaccurate decisions due to dynamic changes in system
conditions, as well as resource consumption fluctuations or unexpected events [4]. As for the group
level, the whole system is divided into different groups and then the replica schedulers are distributed
among each group. This method is also used in load balancing [5]. Due to the decreased
communication cost of a load balancing decision, such a replica scheduling algorithm has a small
impact on the whole system performance, thereby yielding better performance. However pursuing
global optimization may need special mechanisms to balance the load among groups. Hence, the
replication process delay in the group decision maker may still be obvious.
Distributed replication management policies have been proposed to govern efficiently
geographically distributed systems that cannot implement centralized decisions and support strong
interactions among the remote nodes. The distributed resource management is very challenging since
one node’s decisions may inadvertently degrade the performance of the overall system.

International Journal of Digital Content Technology and its Applications(JDCTA) 94


Volume8, Number3, June 2014
Adaptive Data Replication with Load Balancing in Data Grid
Niyoyita Jean Paul, Dong Shou Bin

A typical distributed system involves a large number of geographically distributed storage nodes
which can be interconnected and effectively utilized in order to achieve performances not ordinarily
attainable on a single node. In this study, we consider a distributed system comprising N heterogeneous
nodes (in terms of storage capacity, load, and bandwidth), grouped in different sites. The N nodes of
the system are connected by a communication network. A link between two nodes i and j has a positive
integer c(i, j) associated with it; giving the communication cost for transferring a data unit between
node i and j. If the two nodes are not directly connected by a communication link then the above cost is
given by the sum of the costs of all the links in a chosen path from node i to node j. We adopt a general
assumption as depicted in [5][6], that c(i, j)=c(j, i). Let B be a set of objects, each identifiable by a
unique name dk and size in simple data units sk where 1=k=B. The problem addressed in this paper is
concerned with minimizing the data access delay, i.e., maximizing the probability of serving all the
requests in the required time and data availability.

Contribution:
This work presents a novel technique of data replication and load balancing: Adaptive Data
Replication for Load Balancing in Data Grid (ADRLB). In ADRLB, node load and data availability
are the main factors of replication decision. ADRLB has three main contributions: Initially, the aim is
to balance concurrency on data objects by dynamically creating additional replicas in case the file
cannot be retrieved in required time caused by a sudden high demand for a particular replica. Secondly,
we introduce a new method for selecting target nodes that is in respect to the availability, processing
capacity and based on their relative storage capacity. Therefore we avoid the undesired situation where
some under-loaded peers suddenly become heavily loaded due to too many data (tasks) migrated from
outside. Thirdly, we do meticulous simulations to evaluate the new strategy with the existing ones.

The rest of this paper is as follows: in section 2 we present research works related to this paper. The
new distributed replica placement model is presented in section 3. Simulation results are analyzed in
Section 4 and we conclude the paper and also point out the future work in section 5.

2. Related works
With the diffusion of large data centers, data replication techniques have received a great interest
both within the industry and academic sectors. Different authors like Lei et al. [7], and Schintke and
Reinefeld [8] addressed the data replication for availability in the face of unreliable components. Jiang
and Zhang [9] proposed a technique in Data Grids to measure how soon a file is to be re-accessed
before being evicted compared with other files. They consider both the consumed disk space and disk
cache size. Their approach can accurately rank the value of each file to be cached. For our case we are
not dealing with caching file on disk but replicating file on a storage node. Tang et al. [10] presented a
dynamic replication algorithm for multitier Data Grids. Their system has two dynamic replica sub
algorithms: Single Bottom Up and Aggregate Bottom Up. Performance results showed both
algorithms reduce the average response time of data access compared to a static replication strategy in
a multitier Data Grid. However the system seems to be complex and therefore hard to implement.
Chang and Chang [11] introduced a dynamic data replication mechanism called Latest Access Largest
Weight (LALW). LALW associates a different weight to each historical data access record: a more
recent data access has a larger weight. Though LALW seems to give a precise metric to determine a
popular file for replication, it is not adaptive to sudden file access.
Park et al. [12] proposed a dynamic replication strategy, called BHR, which benefits from “network-
level locality” to reduce data access time by avoiding networking congestion in a Data Grid. In BHR,
sites are divided into several regions, where network bandwidth between the regions is lower than the
bandwidth within the regions. However BHR performs better only when the storage capacity is small.
Later Sashi [13] et al. extended BHR to Modified BHR (MBHR). MBHR strategy replicates a file that
has been accessed most and it may also be used in near future. Mansouri and Dastghaibyfard [14]
presented a Dynamic Hierarchical Replication (DHR) strategy that stores replica in suitable sites where
a particular file has been accessed most, instead of storing file in many sites. It also decreases access
latency by selecting the best replica when different sites hold replicas. Zhang et al. [15] have applied
data replication for military strategy. Their algorithm is based on greedy replication for intra-group

95
Adaptive Data Replication with Load Balancing in Data Grid
Niyoyita Jean Paul, Dong Shou Bin

mobile nodes. Even if their method is efficient for data availability, it can not work well in a system
where replicas are updated frequently.
Data replication problem becomes more complicated under heterogeneous network environments. The
problem of determining the optimal replication solution in a heterogeneous network is similar to the
facility location problem and the K-median problem, and both are proved to be NP-hard. An
approximation algorithm for the data replication problem with uniform-size data items in
heterogeneous networks has been proposed in [16]. Later, Tang [17] designed a polynomial-time
centralized replication algorithm which can achieve a 4-approximation solution, and a localized
distributed algorithm in heterogeneous ad hoc networks. Their works focused on network topology and
bandwidth thus differs from ours which is focusing on data availability and load balancing. In [18] the
authors proposed a load balancing replication strategy, Fair-Share Replication (FSR) that takes into
account both the access load and storage load of the replica servers before placing a replica. The
system is based on hierarchical order of storage node and a centralized replicas management whereas
for us we use a distributed structure. Another closely related work was performed done by Nukarapu
[19] in his strategy for replica placement. This greedy algorithm, at each step, places one data file into
the storage space of one site such that the reduction of total access cost in the data grid is maximized at
that step until all storage spaces of the sites have been replicated with data files or the total access cost
cannot be reduced. In this algorithm, the client site counts hops for each site that hold replicas, and the
model selects the one with the least number of hops from the requesting client; but it does not consider
current network bandwidth.

3. Proposed replica placement strategy


Once a data file is needed for processing within a grid site, it is then copied in the local storage.
The ADRLB strategy is described in three main parts: replicas creation, selection of destination nodes,
and storage cleanup.

3.1. Replica creation


The replica creation may be triggered by one of the two situations: (1) when the file is not found in
the grid site yet it is needed for processing within the same site. (2) When the data file already exists in
cluster but its availability is not achieved because the data file is extremely accessed or a low upload
speed on the host node. In both cases the aim is to guarantee the data availability and therefore, the
number of replicas to be created and the host nodes are taken into account.

3.1.1. Calculating the number of replicas

Based on the retrieval time required for a data object, the system determines the number of replicas
to create for that data object. Here the objective is to allow computing resource to retrieve data from
multiple data hosts or edge servers in parallel to minimize the total transfer time [20] [21]. Since data is
transferred in segments, the transfer process is carried out in parallel when using multiple data sources.
To determine the number of data replicas that will achieve a specific performance requirement, we
need to consider that in a computational grid, the retrieval time of the requested data object depends on
the performance of each individual node hosting a copy of that file, the node failure rate, and also the
network failure rate. At least, we must have one replica to be available. Also, the number of replicas
should not exceed the number of storage resources.
Let n be the total number of data replicas,  be the service rate of node i, which is measured as the
number of requests that can be replied within a specific period of time without interruption. Then, the
service time distribution function  (), referring to the probability that the retrieval time Ti is less than
t, which conforms to an exponential distribution, can be represented as

 () = Pr( ≥ ) = ∫ Λ      =    (1)
The probability that the data file can be retrieved by time t, which is the cumulative distribution
function of the exponential distribution, is
  () = 1 −  () (2)

96
Adaptive Data Replication with Load Balancing in Data Grid
Niyoyita Jean Paul, Dong Shou Bin

Then, the probability, Pnode(t), that at least one node will reply the request by time t is

 ( ,  , … ,  ) ≥ ) = 1 −   ∑  
 () = 1 − Pr(  (3)

If we evaluate the mean of the service rates  ,  = ∑  (4)

in each storage node within the grid site, we are able to estimate a proper number of file replicas, n.
Assume we want at least one request reply at time t with probability , then, we need to have at least n
replicas in grid site, which is obtained in equation (5)
 ()
= . (5)
Λ

3.2. Evaluating the destination node


3.2.1 Node load
Replicating data files is useful only if it cannot overwhelm the destination node. Hence, before
replicating a file, the load of target node should be evaluated. In storage system, users independently
request to read a data object dik of size sik, hosted at i. Since the requests are independent to each other
and the number of requests is large, the arrival of the requests on dik hosted at i can be modeled as a
single Poisson process with mean arrival rate λi. As the accesses to different data objects are
independent to each other, the arrival Poisson processes of the requests can be combined into a single
Poisson process with
 = ∑∀  |ℎ   ℎ   (6)
For simplicity of analysis, we do not address the load difference of update and read requests.
Therefore we use the term “request” for any update or read request. For each interval t, we assume
there are wi independent requests on i which the total load of node i can be calculated as
| |
Ω = ∑    > 0 (7)
However since we are in a dynamic environment, requests might come at any time. Therefore we also
consider the expected coming request in time t+1 represented by ∆  . To avoid overloading and to
keep the node efficient, the amount of load beyond its processing capacity  . In equation (8) we
formulate the expected load status of i at the epoch t.
()
 = Ω + ∆  (8)
Then, considering the load factor, the storage node i is eligible to hold the replica if
()
 < Λ (9)

We can get the value of ∆ based on the historical records of requestes submitted to the node i
within specific periods. We opt for Autoregressive Moving Average (ARMA) prediction method for
its performance over other prediction algorithms [22]. Note that the prediction methods are beyond
this paper; our main effort is to evaluate the coming requests such that the capacity of serving requests
does not degrade. In ARMA model, the notation ARMA (p,q) refers to the model with p autoregressive
terms and q moving average terms. The ARMA (p,q) is written as
 
( + ) =  + ( + ) + ∑  ( +  − 1) + ∑  ( +  − ) (10)
Where c is a constant, ( + ), …, ( +  − ) are white noise error terms,  …,  are
autoregressive parameters (coefficients), and,  , …,  are moving average parameters. Note that
( + ) is generally assumed to be independent and identically distributed. In other words,
( + )~ (0,   ), with the variance   .

3.2.2 Node availability


A storage node becomes unavailable if it fails to respond positively to requests sent to it in required
time. Nodes can become unavailable for a large number of reasons and the majority of such
unavailability events are transient and do not result in permanent data loss. For the sake of simplicity,
we will not analyze the details of node failure i.e. we just consider how the node has responded to the
data request. With atd,  , and  , the availability threshold, the total number of requests to the node i,
and the number of requests not replied by the node i respectively, we define the availability of node i as:
  
 =   (11)


97
Adaptive Data Replication with Load Balancing in Data Grid
Niyoyita Jean Paul, Dong Shou Bin

From the node we can calculate the overall cluster availability which is the average availability, , of
all N nodes which constitute the cluster.

 = ∑  (12)
||

The ideal situation is  =1 and the less is  , the poor is the node performance. To ensure system
availability, in case of load balancing, data files are only copied to qualified nodes. Therefore, if two
nodes are under-loaded with the same processing capacity, the priority is given to the node with
optimal availability. The unqualified nodes may participate in data requisition only when the storage of
other peers is full (this might happen in case a node is removed from the system and therefore its data
has to be distributed to the remaining storage peers).

 = 1       


    
  ≤  <        ℎ   (13)
 <      
  ,  > 0
3.3 File replacement strategy
Due to limited storage capacity, the system only keeps the important data files and the other less
necessary replicas are replaced with more needed data. At first files which have not been used for long
time are removed. If the storage is not enough, we start removing those files which have been recently
accessed. In this process, it is wiser to do not delete valuable files which will cost more when needed
again. Therefore before evicting a file from the storage, two factors are taken into account: the file
retrieval cost which is the distance to the second nearest replica site (where it can be found in case it is
deleted) and the history of access frequency to the data file. We opt for the Least Cost Beneficial based
on K backward references (LBC-K) [23] after modifying it by including the node availability while
calculating the retrieval cost. The LBC-K is compatible with our policy by the fact that it considers the
file access hit and retrieval cost in case it is deleted and, in addition, LBC-K has low computational
complexity comparatively to others.
After evicting files which have not been accessed, we evaluate the utility function Uf(t) of
remaining data objects. An object f of size sf has a retrieval cost cf,j(t) from storage node j. The retrieval
costs vary with time depending on when and where the data file is fetched. In the environment like the
data grid there could be replicas of the same file in different sites. We will denote the cost simply by
cf(t) with the understanding that this is the minimum cost over all replica storage nodes with acceptable
availability, as specified in section 3.2.2 (node availability). At each instant in time t, when we need to
acquire space of size sf for a file k, we order all files which are not currently used in non-decreasing
order of their utility functions and evict the first m files with the lowest values of Uf(t),as calculated in
equation (14) and whose sizes sum up to or just exceed sk.

 ()  ()
 () = ∗

(14)
()
Where  () is the number of the most recent accesses retained within the time interval  −  , and
 is the time of the  () backward accesses to file f;  () is the cumulative count of references to
the file over the active period of access.
3.4 Data replication algorithm
The proposed algorithm consists of replica placement based on the load of hosting node and number
of replica accesses which will determine the number of replicas needed. In ADRLB, the destination
node should not be overloaded because if the storage node is serving many requests the incoming
requests will have to wait and this will slow down the replica access [24][25]. Also to ensure the data
availability, recipient nodes are selecting according to their reliability. We summarize the ADRLB in
following steps:

98
Adaptive Data Replication with Load Balancing in Data Grid
Niyoyita Jean Paul, Dong Shou Bin

INPUT: New file f, file size Sf, required retrieval time t, nodes in grid site N
BEGIN
//Initially, we get the number of replicas to satisfy the availability of the new file f using
equation (5).
k=NumberReplica(f,t,Sf, N)
//Get the qualified storage nodes M from N
M= Available storage nodes (obtained using equation 13) in descending order of load
(calculated in equation 9)1.
For each replica of f (from 1 to k)
Select a distinct storage node i in M
If the available storage space of i is not enough
Free the storage space in i as specified in section 3.4
END
END
END
END

4. Experiments
To evaluate the performance of ADRLB algorithm we have done experiments using OptorSim
simulation tool [28]. The simulation environment was intended to represent a typical data grid system.
The environment we used to evaluate and to compare the data replication had 27 sites with bandwidth
between 100 and 10000Mbs. The time is divided into sessions and each session has a random set of
requests issued by clients in the system. In total, 100 files with size varying from 700 to 10000
Megabytes have been used in the simulation. During the simulation we varied the number of jobs from
400 to 2000. As for the storage size of nodes in the grid site, we have considered a case when the
storage space is very small (50 Gigabytes for storage node) and a case the storage space was large (500
Gigabytes for storage node). Two metrics are considered: mean job execution time and network usage
as evaluation metrics. Mean job execution time is a good measure for effectiveness of the replication
strategies. Jobs in the data grid request a number of files. If the file is at a local site, response time is
assumed to be zero; otherwise the file must be transferred from the nearest replication site. Thus, job
execution time incorporates the response time required to transport a file. The evaluation metric
Effective Network Usage (ENU) in OptorSim is used to estimate the efficiency of network resource
usage and is defined as follow:
    
 = (15)
 

with “Remote accesses” standing for the number of accesses that computing
element reads a file from a remote site, “Replications” is the total number of
file replication occurs, and “Total accesses” is the number of times that
processor reads a file from a remote site or reads a file locally. A good
replication strategy should minimize both the mean job execution time and the
network usage rate.

To evaluate its performance, ADRLB has been compared with four strategies in different types of
access patterns. These replication algorithms (currently new) used for evaluation are: MBHR, DHR,
MLALW, and DDCA algorithm presented in [19].
• The Modified Bandwidth Hierarchy (MBHR) strategy replicates the file within the region in a site
where that file has the highest access.
• Dynamic Hierarchical Replication (DHR) algorithm places replicas in appropriate sites that has the
highest number of access for that particular replica and minimizes access latency by selecting the
best replica and by considering the replica requests that are waiting in the storage and data transfer
time.

1
The sort approach is used for solving file assignment problems and has yield good results [26][27].

99
Adaptive Data Replication with Load Balancing in Data Grid
Niyoyita Jean Paul, Dong Shou Bin

• Modified Latest-Access-Largest-Weight (MLALW), as DHR, places replicas in appropriate sites that


have the highest number of access for that particular replica and deletes replicas by considering three
factors: least frequently used replicas, least recently used replicas, and size of replica.
• Distributed Data Caching Algorithm (DDCA)

4.1 Job execution time


The mean job execution time is defined as the total time required to execute all the jobs, divided by the
number of jobs completed. Figure 1 and figure 2 illustrate the job execution time of five replication
algorithms in case of sufficiency and lack of storage capacity respectively. Though, with a small
number of jobs, the execution time is in the same range for all algorithms, it is clear that as the number
of jobs increases ADRLB performs better than others. The main reason is that within a site, we
replicate files to nodes which are not overloaded and we do not remove files which will be needed in
coming time; hence the access time will reduce.
MBHR

MLALW
DDCA
DHR

2500
Mean Job Execution time in seconds

MBHR

MLALW
DHR

DDCA
2000

ADRLB
ADRLB

MBHR
1500

MLALW
DDCA
DHR

ADRLB
1000

500

0
Sequential access Random access Random walk
Access pattern
Unitary

Figure 1: Job execution time when storage nodes have higher storage capacity

0.5
Network usage (out of 1)

0.4

0.3

0.2

0.1

0
MBHR DHR MLALW DDCA ADRLB
Algorithm

Figure 2: Job execution time when storage nodes have smaller storage capacity

100
Adaptive Data Replication with Load Balancing in Data Grid
Niyoyita Jean Paul, Dong Shou Bin

60000
50000

Time in seconds
MBHR
40000
30000 DHR

20000 MLALW
10000 DDCA
0 ADRLB
400 1000 1500 2000
Number of jobs

Figure 3: Performance of replication strategies according the varying number of jobs

4.2 Effective network usage


We have also evaluated ADRLB algorithm in terms of network usage to promote the efficiency use of
network bandwidth. As illustrated in figure 4 and figure 5, ADRLB performs better in both situations
of storage size. It is worth noting that the file size and storage size used are not for real grid storage but
for simulation.
0.5
Network usage (out of 1)

0.4

0.3

0.2

0.1

0
MBHR DHR MLALW DDCA ADRLB
Replication algorithm

Figure 3: Network usage for storage nodes with big capacity in sequential file access

1
Network usage (out of 1)

0.8
0.6
0.4
0.2
0
MBHR DHR MLALW DDCA ADRLB
Replication algorithm

Figure 4: Network usage for storage nodes with small capacity in sequential file access

5. Conclusion
We have proposed a data replication strategy based on load balancing. The method solves access
concurrency on a replica by adaptively analyzing the replica access frequency, dynamically creates
additional replicas and rebalances other data in the cluster in case of a sudden high demand for a
particular file. Experimental results show the effectiveness of our algorithm.
We hope to improve ADRLB algorithm by finding an adequate file access reference. In fact for
algorithms that use access history to decide replication, there is a time interval for that period; and so

101
Adaptive Data Replication with Load Balancing in Data Grid
Niyoyita Jean Paul, Dong Shou Bin

far that interval is arbitral. However, in a real situation, that observation time can not be the same for
all the times and for all sites. Hence each site should dynamically decide the observation interval
depending on the traffic it is observing.

References
[1] Hara, T.; Madria, S.K., "Data Replication for Improving Data Accessibility in Ad Hoc Networks,"
IEEE Transactions on Mobile Computing, vol.5, no.11, pp.1515-1532, 2006
[2] L. Yin and G. Cao. “Balancing the tradeoffs between data accessibility and query delay in ad hoc
networks”. IEEE International Symposium on Reliable Distributed Systems, pages 289—298,
2004
[3] S. Ghemawat, H. Gobioff, and S. Leung, "The Google file system," in ACM Symposium on
Operating Systems Principles pp. 29 - 43, 2003.
[4] Ardagna, Casolari, Colajanni, and Panicucci, “Dual time-scale distributed capacity allocation and
load redirect algorithms for cloud systems”, Journal of Parallel and Distributed Computing Vol.
72, pp. 796–808 2012
[5] W. Liu, M. Wu, X. Ou, W. Zheng, M. Shen, “Design of an I/O balancing file system on web
server clusters”, in: Proceedings of the 2000 International Workshop on Parallel Processing,
ICPP’00, pp. 119–126, 2000
[6] Xindong You; Guiran Chang; Xueyao Chen; Cuihua Tian; Chuan Zhu, "Utility-Based Replication
Strategies in Data Grids," Fifth International Conference on Grid and Cooperative Computing,
pp.500,507, 2006
[7] M. Lei, S.V. Vrbsky, and X. Hong, “An Online Replication Strategy to Increase Availability in
Data Grids,” Future Generation Computer Systems, vol. 24, pp. 85-98, 2008
[8] F. Schintke and A. Reinefeld, “Modeling Replica Availability in Large Data Grids,” Journal of
Grid Computing, vol. 2, no. 1, pp. 219-227, 2000
[9] Song Jiang; Xiaodong Zhang, "Efficient distributed disk caching in data grid management," In
Proceedings of IEEE International Conference on Cluster Computing, pp.446,451, 2003
[10] M. Tang, B.-S. Lee, C.-K. Yeo, and X. Tang, “Dynamic Replication Algorithms for the Multi-Tier
Data Grid,” Future Generation Computer Systems, vol. 21, pp. 775-790, 2005
[11] R.-S. Chang and H.-P. Chang, “A Dynamic Data Replication Strategy Using Access-Weight in
Data Grids,” Journal of Supercomputing, vol. 45, pp. 277-295, 2008
[12] S.M. Park, J.H. Kim, Y.B. Lo, and W.S. Yoon, “Dynamic Data Grid Replication Strategy Based
on Internet Hierarchy,” In Proceedings of Second International Workshop of Grid and Cooperative
Computing (GCC), 2003
[13] Sashi, K. and A.S. Thanamani, “Dynamic Replication in a Data Grid Using a Modified BHR
Region based Algorithm”, In proceedings of the Future Generation Computer Systems, 27(2): pp.
202-210, 2011
[14] Najeme Mansouri and Gholam Hosein Dastghaibyfard, “A Dynamic Replica Management
Strategy in Data Grid”, Journal of Network and Computer Applications, Vol.35, no.4, pp.1297-
1303, 2012
[15] Zhang Y. et al., “Data Replication in Mobile Tactical Networks”, Military Communications
Conference pp.797,803,2011
[16] I. Baev, R. Rajaraman, and C. Swamy, “Approximation Algorithms for Data Placement
Problems,” SIAM Journal of Computing, vol. 38, no. 4, pp. 1411-1429, 2008
[17] B. Tang, H. Gupta, and S. Das, “Benefit-based Data Caching in Ad Hoc Networks,” IEEE TMC,
vol. 7, no. 3, pp. 289–304 (2008)
[18] Taheri, Javid; Zomaya, Albert Y.; , "A Pareto Frontier for Optimizing Data Transfer and Job
Execution in Grids," Parallel and Distributed Processing Symposium Workshops & PhD Forum,
IEEE 26th International, pp.2130-2139, 2012
[19] Nukarapu, D.T.; Bin Tang; Liqiang Wang; Shiyong Lu, "Data Replication in Data Intensive
Scientific Applications with Performance Guarantee," IEEE Transactions on Parallel and
Distributed Systems, vol.22, no.8, pp.1299-1306, 2011
[20] Jun Feng; Humphrey, M., "Eliminating replica selection - using multiple replicas to accelerate data
transfer on grids," Tenth International Conference on Parallel and Distributed Systems,
pp.356,366, 2004

102
Adaptive Data Replication with Load Balancing in Data Grid
Niyoyita Jean Paul, Dong Shou Bin

[21] X. Zhou, E. Kim, J. W. Kim, and H. Y. Yeom, “ReCon: A Fast and Reliable Replica Retrieval
Service for the Data Grid,” In Proceedings of the Sixth IEEE International Symposium on Cluster
Computing and the Grid. pp. 446–453, 2006
[22] P.A. Dinda and D. O'Hallaron, “An Evaluation of Linear Models for Host Load Prediction” in
Proceedings of 8th IEEE International. Symposium on High- Performance Distributed Computing
(HPDC-8), pp. 87-96, 1999.
[23] Otoo, E.; Olken, F.; Shoshani, A., "Disk Cache Replacement Algorithm for Storage Resource
Managers in Data Grids," In Proceedings of ACM/IEEE 2002 Conference on Supercomputing,
pp.12,12, 2002
[24] Phongsathon Fongta and Yuthapong Somchit, “Object request reduction in home nodes and load
balancing of object request in Hybrid decentralized web caching”, International Conference on
Information Communication and Management (ICICM 2012), 2012
[25] Bin Dong, Xiuqiao Li, Qimeng Wu, Limin Xiao, Li Ruan, “A dynamic and adaptive load
balancing strategy for parallel file system with large-scale I/O servers” Journal of Parallel
Distributed Computing Vol. 72, pp.1254–1268, 2012
[26] Lin-Wen Lee; Scheuermann, P.; Vingralek, R., "File assignment in parallel I/O systems with
minimal variance of service time," IEEE Transactions on Computers, vol.49, no.2, pp.127,140,
2000
[27] Yu, Y.; Zhu, Y.; Ng, W.; Samsudin, J.; Li, Z., "A File Assignment Strategy Towards Minimized
Response Time for Parallel Storage Systems,"IEEE Transactions on Magnetics, vol.49, no.6,
pp.2459,2465, June 2013
[28] D. G. Cameron, A. P. Millar, C. Nicholson, R. Carvajal-Schiaffino, K. Stockinger, and F. Zini.
“Analysis of scheduling and replica optimization strategies for data grids using optorsim”. Journal
of Grid Computing, vol.2, no.1 pp.57–69, 2004

103

You might also like