Popularity (Hit Rate) Based Replica Creation for Enhancing the Availability in Cloud Storage

Abstract

In cloud computing, the replication management system has been well adopted in cloud storage applications. To provide the availability and reliability, the replication system replicates the files and can be stored in different server. The system led some complicated issues such as high memory consumption, incurred high storage cost and to access the file is more complicated issues in recent cloud storage applications. In the existing technique, File Accessing Frequency based Ranking (FAFR) Algorithm and Dynamically Reduced Replica for Rarely Accessed files (DRRRA) algorithm work jointly and identify the rarely accessed files and retain the replica in two server other replicated files are deleted. To provide access to more request with 2 or 3-replica is a complicated issue. Thus, this paper proposes a Dynamic replica Creation for Availability enhanced Storage (DRCAES) algorithm which jointly work with FAFR algorithm to predict most frequently accessed files and automatically replicated to other server based on server memory. The aim of this proposed approach is to enhance the availability, thereby reducing the request-response delay time. Thus the proposed approach optimizes the number of replicas, occupied space, and cost.

Full text

Received: November 18, 2017 161
International Journal of Intelligent Engineering and Systems, Vol.11, No.2, 2018 DOI: 10.22266/ijies2018.0430.18
Popularity (Hit Rate) Based Replica Creation for Enhancing the Availability
in Cloud Storage
S. Annal Ezhil Selvi1* R. Anbuselvi1
1Department of Computer Science, Bishop Heber College,
Trichy, Tamilnadu, 620017, India
* Corresponding author’s Email: ezhilabel.bhc@gamil.com
Abstract: In cloud computing, the replication management system has been well adopted in cloud storage
applications. To provide the availability and reliability, the replication system replicates the files and can be stored in
different server. The system led some complicated issues such as high memory consumption, incurred high storage
cost and to access the file is more complicated issues in recent cloud storage applications. In the existing technique,
File Accessing Frequency based Ranking (FAFR) Algorithm and Dynamically Reduced Replica for Rarely Accessed
files (DRRRA) algorithm work jointly and identify the rarely accessed files and retain the replica in two server other
replicated files are deleted. To provide access to more request with 2 or 3-replica is a complicated issue. Thus, this
paper proposes a Dynamic replica Creation for Availability enhanced Storage (DRCAES) algorithm which jointly
work with FAFR algorithm to predict most frequently accessed files and automatically replicated to other server
based on server memory. The aim of this proposed approach is to enhance the availability, thereby reducing the
request-response delay time. Thus the proposed approach optimizes the number of replicas, occupied space, and cost.
Keywords: Cloud storage, Replication, Reduce replica, Dynamic replica, File popularity, File accessing frequency.
1. Introduction
Microsoft Azure, Amazon, Google Cloud
Storage (GCS) and as leading Cloud service
Providers offer different types of storage (i.e.,
sequences of files, etc.) with different prices for data
storage services. The data storage services and
accessing files are very difficult issues on
Redundancy Storage. Each cloud service provider
also provides and monitors the commands to
retrieve, store and delete data through network
services, which impose in- and out-network delay
and cost on an application [1]. In leading Cloud
service provider in network cost is free, while out-
network cost (network cost for accessing) is charged
and may be different for usage of cloud providers. In
cloud server Data transferring or data replication
among from one server to other server, this shows
significant price differences among them. The
existing problem on this diversification plays an
essential role in the optimization of data
management request response and delay in cloud
environments [2]. This proposed techniques at
optimizing this request response and delay that
consists of residential cost (i.e., storage) and
potential migration cost (i.e., network server cost).
Cloud service provider many applications are
moving towards a distributed interconnected
network environments. In this distributed
environment, the data storage and all computational
cloud resources are distributed during different and
widespread locations based on ranking.
A cloud server store the data, the data can have
a huge number of users that require having access to
huge data volumes. For example, consider a set of
documents or images or videos that needs to be read
and accessed by a number of user spread worldwide,
in a distributed way. The access to vast data
volumes by huge number of users can be access
very time consuming. As the size of the system is
increased, the tasks of providing such data service

Received: November 18, 2017 162
International Journal of Intelligent Engineering and Systems, Vol.11, No.2, 2018 DOI: 10.22266/ijies2018.0430.18
becomes more difficult since its users suffer from
long delays in data access.
Replica cloud storage is a new research field; the
dynamic replication policy is still rarely seen. In this
paper, the focal point is on early distributed storage
file system replica creation approach, combine with
the quality of cloud storage, this paper design a
dynamic replication strategy based on high
prediction. It create a replica according to the file
access history quality, so with the purpose of users
can access the nearby which bring the cloud system
to one of the most excellent status with ranking,
specifically the replica number of minimum, access
to the highest efficiency, network lifetime is
increased.
A fixed number of replicas for every file is
insufficient to give quick file read for hot files while
waste resources for storing replicas of cold files.
Random selection of replica destination require
observance all Data Centers active to ensure data
availability, which though waste power consumption.
As the random selection of replica destination does
not think purpose of bandwidth and request handling
capacity, network congestions could occur due to
capacity restriction of some links and server may
turn into overloaded by data requests.
In this environment, data replication is essential
so that the users can retrieve the most request data
from storage residing in nearby server. The
replication is based on server memory [3]. The
performance of the distributed networks is crucially
affected by the replication strategy used. The huge
majority of the known replication strategy
determines the replicas by computing an easy
ranking based on the number of requests for each
file on individual cloud server. The “most accessing”
files, the ones with the highest ranking value, are
selected for replication due to the memory based
other server replicated.
This ranking technique that it is quite possible
that, the accessing files with high recent demand
will be requested on cloud server, with even higher
hit rates. Since the computation is quite simple, the
strategy mainly focuses on the problem of select the
most suitable files for storing the replicas based on
memory [4].
The two main drawbacks of the strategies
proposed are related to the following techniques
implemented in this work. First one is optimization
process and second one is high prediction ranking
algorithm. The scheme presented in the literature
does not take into account the change that might
suggest itself in the interest of users for certain files.
Instead, they are mostly involved in one or more
factors that decide the importance of the file
themselves, like the file size, the number of requests
for an entity file or the contents of a file. The user
request response time and files are analysed
optimization process [5].
The Second process user request analysed and
which files are most hitting on individual server.
The existing replication decision algorithm satisfies
the better request response time but the replication
of hitting is complicated [5]. The hitting ratio is
calculated or monitor on high prediction ranking
algorithm. This algorithm identifies the most hitting
files and replicated to other server based on server
memory. The replication process implemented on
this server automatically reduced the delay and
request response time on server. At the end of this
approach the cloud storage system will act as a
Role-Based intelligent System (RBIS). So that, the
user can an efficient data storage on cloud
computing environment.
Some of the roles of the DRCAES approach is
listed below,
 Dynamically predict rarely accessed files and
most frequently accesses the file using FAFR
model.
 Through DRRRA dynamically reduce the
number of replicas of that rarely accessed files,
so that, the cost and occupied space will be
minimized. For reduction of the replica, it finds
minimum available Space of DC among DC’
where that file exists (Removal).
 In DRCAES, dynamically create and place the
new replica for the most accessed file if the
frequency of each replica is equally accessed
otherwise it won’t replicate. The new replica
placed in the data center that has more available
space and that file does not exist.
 Balanced storage retained during removal and
new replica placement. Because, it analyses all
aspects of Storage system like available Space,
SLA, Accessing frequency of all existing replica.
The existing work are discussed in Section 2,
overall back ground process are described in section
3, In section 4 the proposed work which is improve
the overall request time and reduced delay using
high prediction ranking algorithm are described. The
section 5 discusses the results details. And section 6
is conclusion.
2. Related works
Data replication issue effectively requires taking
a closer look at the arrangement of most common
services and applications deployed on storage clouds
to provide services to other parties. Such
applications are usually implement as multi-tier

Received: November 18, 2017 163
International Journal of Intelligent Engineering and Systems, Vol.11, No.2, 2018 DOI: 10.22266/ijies2018.0430.18
applications running on distributed software system.
The multi-tier application user giving the request the
overall response time is very high. The storage
strategies so far consider the number of requests as
the main hitting files for computing the popularity of
each file [6].
The hitting files identify the limitation of the
current research of data replication in cloud server:
they are whichever hypothetical investigation
without realistic consideration, or heuristics-based
execution without a provable performance guarantee.
The most directly related work to this replication
work is complicated process on clouds server. The
data replication and request response on cloud
server as a static optimization problem on user
access [7]. They show that this problem is NP-hard
and request delay, which means that present, is no
polynomial algorithm that provides an accurate
solution. They only reflect on static data replication
for the intention of proper analysis. The limitation of
the static approach is that the replication cannot
regulate to the dynamically shifting user access
prototype. Additionally, their centralized process of
integer programming technique cannot be simply
implementing in a distributed cloud server.
The request response and resource sharing use
an auction protocol to make the replication choice
and to trigger long-term optimization by with file
access patterns. In this propose utility-based
replication strategies on clouds server. In this
process address the data replication for availability
in the face of unreliable works, this is different from
this optimization work [8, 9].
The random collection of replica destination
neglects server heterogeneity (i.e., different Data
Centers vary in data request handling capacities and
network capacities). The write due to creating
replicas in production clusters at searching engine
application for almost half of all cross-rack traffic.
While the network within clusters is frequently
underutilized, there exist some traffic jam links
resulting from the network usage imbalance [10].
To assume the multi-facility cloud resource
allocation problem, they are mainly involved in
solutions that are agreeable to parallel
implementations. There is quite a lot of reason. First,
for a cloud resource allocation, problem (1) is
inherently an important convex resource
optimization problem, with millions of variables or
still more. A centralized process of cloud server
resource allocation solver is extremely inefficient in
solving such large-scale cloud storage problems [11].
3. Background process
Cloud services, such as search engines,
education portal, parallel application, social
networking, etc., are often deploy on a
geographically spread the infrastructure, i.e. data
centers placed in different regions and better and
reliability. A usual query is then how to direct the
workload from users along with the set of geo
distributed data centers in organize to achieve a
desired transaction between performance and delay,
since the power price exhibit an important degree of
geographical diversity [12]. This query has involved
much attention recently and is usually referred to as
geographical cloud server load balancing.
The resource scheduling based data replication
problem and focus on scheduling pathetically
parallel resource usage which are collected of a set
of independent responsibilities with very minimal or
no data synchronization. A huge number of
applications fit in to this type of resource sharing on
cloud storage. Examples consist of distributed
relational database query, search engine query,
BLAST searches, data processing, and image
processing applications such as shaft tracing. To
effect apathetically parallel resource allocation, each
of its tasks is placed on a physical server and
executed in an external server added for that task.
The completion time of this resource request is the
completion time of the last finished request and
overall request response, i.e., the make criticize of
that set of request are completed [13].
The conventional data caching/replication
problem have been considered extensively in the
framework of the Web distributed cloud databases
and multimedia systems. What be different from
Web caching is that disk memory and I/O bandwidth
are the main concerns in multimedia storage systems.
A number of algorithms are proposed to attain high
acceptance rate and resource utilization by balancing
the use of different request response resources.
Unlike Web search engine and multimedia data,
database contents are access by both read and write
operations based in optimization process and
ranking [14].
It is assumed that the frequent accessed files in
the past will be accessed more than the others in the
future. This is called as high prediction ranking on
temporal locality. With the property of sequential
locality, a most accepted data file is resolute by
analysing the number of access to the data files from
users. After discovery the best popular file, we trace
to the client that produce the most requests for the
popular data file and a new replica is placed in it.
Therefore, in this application have to collect history

Received: November 18, 2017 164
International Journal of Intelligent Engineering and Systems, Vol.11, No.2, 2018 DOI: 10.22266/ijies2018.0430.18
of records regarding the end-to-end data transfers to
decide which file should be replicated [15].
4. Proposed framework and algorithm
Dynamic replication of data is another
significant issue since access frequencies to
individual data items are likely to modify in most
cloud server environments. The aim is to make the
replication strategy rapidly and accurately adapt to
changes and achieve optimal ranking process on
long-term performance.
In [12], they establish that, to take advantage of
cached data, it is sometimes essential to procedure
individual queries using “suboptimal” plans in
arrange to reach high system performance. In data
replication is triggered as a result of changes of
request rates in cloud server.
4.1 Dynamic replica creation availability
enhanced storage framework
The proposed contribution is used an
optimization based high prediction ranking
algorithm. The Ranking algorithm (FAFR) is giving
better performance on request response time and
delay. The ‘N’ number of user agreed request and
access the files on cloud server. To find the files and
request user identification process and queuing
process all are calculated the cloud server. Fig. 1 is
architecture of optimization based ranking with
conclude of input request response. The user
requests the input and cloud analyse the user which
files are request. The server files viewed by ranking
(based on most popularity). Example the java.doc
files mostly requested and download the files from
user means, the java.doc is replicated to other server
based on server memory.
Design data sharing-aware optimization
algorithms for solving the resource request problem.
Before relating the algorithms establish few static
definition and assumption concerning the cloud
servers. The Data Centers manage by the cloud
service provider are in one of the following two
states: active (available) and replicate.
An active cloud server is a server that is
powered on and is currently considered for all
resource allocation by the algorithms. A replication
is a server that is most frequent files are request in
cloud server considered for replicated data on other
server based on memory allocation by the
algorithms. So it can denote by Fi, Si the set of most
frequent access file and number of Data Centers.
When the entire cloud Data Centers hosted by files,
and accessing most frequent files are replicated files
on ranking based techniques in clouds server.
Figure.1 Proposed architecture
Data Centers are configured and located in
different geo-locations. Files are stored in that
configured Data Centers. The previous work of this
research proposed an algorithm (Dynamically
Reduced Replica for Rarely Accessed file (DRRRA)
Algorithm [15]) reduce the replica based on file
popularity (Least accessed files) which is the result
of Ranking Algorithm (File Accessing Frequency
based Ranking) [14]. The ranking algorithm predicts
the files based on their popularity. But, this paper
focuses the most frequently accessed file which is
the most popular file. If the user need is increased
for a file that files replicated dynamically.
4.2 Mathematical model for replica creation and
placement
The following notation and equations are used in
the prediction process of frequently accessed files
and replica increasing process.
 m: Number of uploaded files.
 n: Number of Data Center (DC).
  : ith DC, i  1,…,n.
  : jth file, j  1,…,m.
  Replica Accessing Frequency (RAF)
(Hit Rate) of ith file in jth DC. It is shown in
the following matrix (m X n ) representation.
   
 







  





  




 FAF: File Accessing Frequency computed
using eq. (1) as,
 FAF

 , i  1,…, n
 j  1,…,m. (1)

Received: November 18, 2017 165
International Journal of Intelligent Engineering and Systems, Vol.11, No.2, 2018 DOI: 10.22266/ijies2018.0430.18
 TFL : Total File list(TFL) in ith file in jth
DC, where i  1,…,n  j  1,…,m.
 AFL : L where is Allocated File List
(AFL) of ith file in jth DC, if

 > 0,
TFL
 NAFL: Non Allocated File List of ith file
in jth DC, if

  0, TFL
 TFLDC :K where is Total File List of
each Data Center‘ of ith DC in jth File,
where i  1,…,n  j  1,…,m.
 AFLDC : D where is Allotted File List of
each Data Center, of ith DC in jth File, if

  0 K.
 DCC : Data Center' Capacity (DCC) of jth
DC, where j  1,…, n .
 FS: File’s Sizes of jth DC, where j  1,…,
m .
 OS: Occupied Space of jth DC is calculated
using the following equation
OS= 

 D, j  1,…,n. (2)
 AS : Available Space of jth DC calculated
using eq. (3) as,
AS = DCC- OS j  1,…,n. (3)
 FA_Th: Frequently Accessed files
Threshold Value.
 Th_LL: Low Limit of threshold value.
 Th_UL: Upper Limit of threshold value
In prediction process, there are two levels of
prediction done to find the files which are really
needed to increase the number of replicas for
meeting the availability enhancement requirements.
In first level prediction, the most frequently
accessed files are predicted using Eq. (4) based on
the FAF of the FAFR model.
The second level prediction able to done based
on the result of the first level prediction. By first
level prediction, some of the most frequently
accessed file is in resulting set. From that files, the
second level prediction finds the files which are
really required to provide seamless availability that
is done using Eq. (5) based on RAF of the FAFR
model.
  FA_Th (4)
 
󰇱 󰇡ASNAFL󰇢 if th_LL



th_UL,
  

(5)
To be exact, Check the frequency of each replica
(if it is equally accessed it will be replicated
otherwise it won’t be replicate). Then the Replica
Placement will be done based on available space of
the data center. That is, the dynamically created
replica will be placed in Datacenter that has more
available space and that file does not exist.
4.3 File accessing frequency based ranking
(FAFR) algorithm
FAFR is a ranking algorithm which proposed in
[14] itself. But in this paper, the new name is given
with some refined work.  is the Replica Accessing
Frequency (RAF) (hit rate). Initially,  value is 0
when the file is uploaded then the  value became 1.
And whenever the file is accessed the  value will
be incremented. Finally, the summative  value is
calculated and considered as a File Accessing
Frequency.
Input: , , M(), k=0,  =0,q
Output: Rank { q’s result set}
1. Done  file in  Data center’s when user
interface triggered
2. If file upload
3.  =1
4. If file access
5.  =  +1;
6. for each  do
7. for each  do
8. K=k+ 
9. End
10. Rank. insert (all Values)
11. If Rank =q
12. Return Rank
Figure.2 Working principle of DRRRA

Received: November 18, 2017 166
International Journal of Intelligent Engineering and Systems, Vol.11, No.2, 2018 DOI: 10.22266/ijies2018.0430.18
In Fig. 2 is shown the working principle of
Dynamically Reduced Replica for Rarely Accessed
files (DRRRA) is described in [15]. FAFR and
DRRRA are jointly worked and predict Rarely
Accessed files then reduce their replica to 2-replica
strategy. Here the minimum replica is 2 for assuring
Reliability and maximum is 3-replica strategy.
4.4 Dynamic replica creation for availability
enhanced storage (DRCAES) algorithm
The following DRCAES algorithm dynamically
finds the most frequently accessed file using FAFR
[14]. Then, based on FAFR result DRCAES
dynamically replicate the replica and place it on DC
that have maximum Available Space (AS) and that
DC does not have that file’s replica on it.
Input : Fi, DCi , NALk,  i,j , FA_Th, Th_LL ,
Th_UL., DC, Replica
Output: Dynamically Increased Replica.
1. Set values to FA_Th, Th_LL and Th_UL.
2. If user access
3. FAF getFAF()
4. If >= FA_Th then
5. Fi  getHighRankFiles()
6. DCi  getHighRankFilesDC()
7. NALk  get_allNonAllocatedList()
8. For each Fi do
9. For each DCj do
10.  i,j  getRAF(F i , D j)
11. If( >= Th_LL and  i,j < Th_UL) then
12. Replica=copyof(Fi)
13. DC=getDCid(max(ASi
getAvailableSpace(NALk )))
14. REPLICATE( DC, Replica)
15. End
16. End
In DRCAES algorithm Step (3) is used to get
most frequently accessed files. For every user access
monitored and when FAF reach the FA_Th valued
that time it performs the second level prediction
which is done in RAF. That is, the every RAF
should satisfy the range which mentions in step 10.
End of this checking, if the result set has files that
files are required to increase their number of
replications, that will be done based on SLA
specification and Available Space of DC which is
defined in Step (12).
Step (11) is used to verify if the file need to
replicate or not. That is, Th_LL is lower limit of
threshold value range and Th_UL is a upper limit of
threshold value range which determine the ranges of
threshold values are decide to replicate or not.Then,
Step(12-14) are work when step (11)’s decision.
Example:
The worked out examples of DRCAES
algorithm is presented in tables 1 to 6, which will be
discussed below one by one. After this discussion,
the reader can easily understand the concept of our
approach clearly.
The Prediction of Frequently Accessed Files
based on FAF process is represented in table 1. The
FA_Th value for validation is 15. When the FAF
reach 15 that file comes under the consideration. For
an example, using Eq. (4) the files 1, 3, and 5 are in
consideration.
In Table 1 shows seven Data Centers are
configured with 5GB Memory and they are located
in different geo-locations. And five different files
are stored on among the 7 DC.
Next, these files are verified by second level
prediction process that is Prediction of highly
needed files based on RAF which is explained in
table 2. In second level prediction done using Eq.
(5) which checks the individual replica frequency, if
it all equally accessed that file will be replicated in
one more data center.
In Eq. (5), there are two threshold values
involved. First one is Th_LL, its value for validation
is 10. The second one is Th_UL, its value for
validation is 20. The file 3’s replicas residing in
DC2, DC4, and DC5, as well as their RAF, is 3 is
8,10,6 respectively. So this files not in the range of
Th_UL and Th_UL. Thus, this file need not be
replicated, but other two files 1 and 5 need to be
replicated because it satisfies the range values.
In table 3 depicted the process of replica
placement which is described in Eq. (7). The file
1’s replicas are residing in DC2, DC4 and DC7
along with their RAF are 14, 20 and 15 respectively.
The replica placement has done based on
Available Space (AS) of the data center which
doesn’t have the replica of the file. Here, the DC1,
DC3, DC5, and DC6 doesn’t have the replica of
file1 as well as their AS is 4.5, 4.9, 4.7 and 4.8
respectively.
The DC3 has the maximum of AS among the
DCs which as mentioned in the previous point. So,
the replica will be placed in DC3, the reflected
changes in metadata are shown in table 4.

Received: November 18, 2017 167
International Journal of Intelligent Engineering and Systems, Vol.11, No.2, 2018 DOI: 10.22266/ijies2018.0430.18
Table 1. Prediction of frequently accessed files based on FAF
S
.
N
o
File Name
File
Type
N
R
File
Size
in
MB
RAF (Replica Accessing Frequency ) ()
F
A
F
DC 1
5GB
DC 2
5GB
DC 3
5GB
DC 4
5GB
DC 5
5GB
DC 6
5GB
DC 7
5GB
1
Array_Java
docx
3
0.07
0
14
0
20
0
0
15
49
2
Tree_ds
pdf
2
0.11
0
0
1
0
0
1
0
2
3
Img_001
jpeg
3
0.31
0
8
0
10
6
0
0
24
4
CS_C
mp3
2
0.55
3
0
0
2
0
0
0
5
5
HelloEnglish
mp4
2
1
0
13
0
0
0
0
19
32
OS (Occupied Space In GB)
0.5
1.4
0.1
0.9
0.3
0.1
1.0
AS ( Available Space in GB)
4.5
3.6
4.9
4.1
4.7
4.9
4.0
Table 2. Prediction of highly needed files based on RAF
S.No
File Name
File
Type
NR
File
Size
in
MB
RAF (Replica Accessing Frequency ) ()
FAF
DC
1
5GB
DC
2
5GB
DC
3
5GB
DC
4
5GB
DC
5
5GB
DC
6
5GB
DC
7
5GB
1
Array_Java
docx
3
0.07
0
14
0
20
0
0
15
49
2
Tree_ds
pdf
2
0.11
0
0
1
0
0
1
0
2
3
Img_001
jpeg
3
0.31
0
8
0
10
6
0
0
24
4
CS_C
mp3
2
0.55
3
0
0
2
0
0
0
5
5
HelloEnglish
mp4
2
1
0
13
0
0
0
0
19
32
OS (Occupied Space In GB)
0.5
1.4
0.1
0.9
0.3
0.1
1.0
AS ( Available Space in GB)
4.5
3.6
4.9
4.1
4.7
4.9
4.0
Table 3. Ex. 1: RAF based replica creation (Before placement)
S.No
File Name
File
Type
NR
File
Size
in
MB
RAF (Replica Accessing Frequency ) ()
FAF
DC
1
5GB
DC
2
5GB
DC
3
5GB
DC
4
5GB
DC
5
5GB
DC
6
5GB
DC
7
5GB
1
Array_Java
docx
3
0.07
0
14
0
20
0
0
15
49
2
Tree_ds
pdf
2
0.11
0
0
1
0
0
1
0
2
3
Img_001
jpeg
3
0.31
0
8
0
10
6
0
0
24
4
CS_C
mp3
2
0.55
3
0
0
2
0
0
0
5
5
HelloEnglish
mp4
2
1
0
13
0
0
0
0
19
32
OS (Occupied Space In GB)
0.5
1.4
0.1
0.9
0.3
0.1
1.0
AS ( Available Space in GB)
4.5
3.6
4.9
4.1
4.7
4.8
4.0

Received: November 18, 2017 168
International Journal of Intelligent Engineering and Systems, Vol.11, No.2, 2018 DOI: 10.22266/ijies2018.0430.18
Table 4. Ex. 1: RAF based replica creation (After placement)
S.No
File Name
File
Type
NR
File
Size
in
MB
RAF (Replica Accessing Frequency ) ( )
FAF
DC
1
5GB
DC
2
5GB
DC
3
5GB
DC
4
5GB
DC
5
5GB
DC
6
5GB
DC
7
5GB
1
Array_Java
docx
4
0.07
0
14
1
21
0
0
15
49
+1
2
Tree_ds
pdf
2
0.11
0
0
1
0
0
1
0
2
3
Img_001
jpeg
3
0.31
0
8
0
10
6
0
0
24
4
CS_C
mp3
2
0.55
3
0
0
2
0
0
0
5
5
HelloEnglish
mp4
2
1
0
13
0
0
0
0
19
32
OS (Occupied Space In GB)
0.5
1.4
0.4
0.9
0.3
0.1
1.0
AS ( Available Space in GB)
4.5
3.6
4.6
4.1
4.7
4.9
4.0
Table 5. Ex. 2: RAF based replica creation (Before placement)
S
.
N
o
File Name
File
Type
NR
File
Size
in
MB
RAF (Replica Accessing Frequency ) ()
FAF
DC 1
5GB
DC 2
5GB
DC 3
5GB
DC 4
5GB
DC 5
5GB
DC 6
5GB
DC 7
5GB
1
Array_Java
docx
4
0.07
0
14
1
20
0
0
15
49 +1
2
Tree_ds
pdf
2
0.11
0
0
1
0
0
1
0
2
3
Img_001
jpeg
3
0.31
0
8
0
10
6
0
0
24
4
CS_C
mp3
2
0.55
3
0
0
2
0
0
0
5
5
HelloEnglish
mp4
2
1
0
13
0
0
0
0
19
32
OS (Occupied Space In GB)
0.5
1.4
0.4
0.9
0.3
0.1
1.0
AS ( Available Space in GB)
4.5
3.6
4.6
4.1
4.7
4.9
4.0
Table 6. Ex. 2: RAF based replica creation (After placement)
S.
N
o
File Name
File
Type
NR
File
Size
in
MB
RAF (Replica Accessing Frequency ) ()
FAF
DC 1
5GB
DC 2
5GB
DC 3
5GB
DC 4
5GB
DC 5
5GB
DC 6
5GB
DC 7
5GB
1
Array_Java
docx
4
0.07
0
14
1
20
0
0
15
49 +1
2
Tree_ds
pdf
2
0.11
0
0
1
0
0
1
0
2
3
Img_001
jpeg
3
0.31
0
8
0
10
6
0
0
24
4
CS_C
mp3
2
0.55
3
0
0
2
0
0
0
5
5
HelloEnglish
mp4
2
1
0
13
0
0
0
0
19
32
OS (Occupied Space In GB)
0.5
1.4
0.4
0.9
0.3
0.1
1.0
AS ( Available Space in GB)
4.5
3.6
4.6
4.1
4.7
4.9
4.0

Received: November 18, 2017 169
International Journal of Intelligent Engineering and Systems, Vol.11, No.2, 2018 DOI: 10.22266/ijies2018.0430.18
Table 7: Comparison of Existing PRC [3], Proposed DRRRA [15] and DRCAES
S.no
FAF
Existing
(PRC)
conventional 3 Replica
Strategy (Minimum Replica 1
Maximum Replica 3)
Proposed DRRRA
Algorithm
(Dynamically Reduced
Replica of Rarely Accessed
files (2-Replica)
Proposed DRCAES
Algorithm
(Dynamically create the
Replica based on User need)
NR
OS=
FS *
NR
Cost= OS *
0.00067
RR_TD
NR
OS=
FS *
NR
Cost= OS *
0.00067
RR_TD
NR
OS=
FS *
NR
Cost= OS *
0.00067
RR_TD
1
>10
3
2.31
0.0015477
3
2
1.54
0.0010318
3
2
1.54
0.0010318
3
2
15-40
3
2.31
0.0015477
3.5
3
2.31
0.0015477
3.5
3
2.31
0.0015477
3.5
3
41-60
3
2.31
0.0015477
4.2
3
2.31
0.0015477
4.2
4
3.08
0.0020636
3.7
4
61-80
3
2.31
0.0015477
7
3
2.31
0.0015477
7
5
3.85
0.0025795
3.9
In table 5 is shown the replica placement process
of file 5. The file 5’s replicas are residing in DC2
and DC7 along with their RAF is 13, and 19
respectively. Here, the highlight point regarding this
file is, the replica of this file is reduced by DRRRA
approach (2-replica strategy) which is discussed in
chapter 6. But, the need for this file is increased. So,
it is going to be replicated.
Here, the DC1, DC3, DC4, DC5, and DC6
doesn’t have the replica of file 5 as well as their AS
is 4.5, 4.6, 4.1, 4.7 and 4.9 respectively. The DC6
has the maximum of AS. So, the replication will be
stored in DC6 which is shown in table 6. It is shown
in Table (5) is after ranking the process replicated to
that files based on server memory.
5. Result and discussion
The reflections in the parameter which involved
in research due to DRCAES approach is presented
in table 7. There are 4 parameters and their value
calculation shown in this table such as NR (number
of Replicas), OS (occupied Space), Cost and
RR_TD (Request-Response Time Delay).
Finally, when the research comparing the
proposed algorithms with the existing algorithm the
DRCAES give better performance in all aspects
such as the number of replicas, Occupied Space
(OS), Cost and Request-Response Time Delay
(RR_TD) based on File Accessing Frequency
Finally, when the research comparing the proposed
algorithms with the existing algorithm the DRCAES
give better performance in all aspects such as the
number of replicas, Occupied Space (OS), Cost and
Request-Response Time Delay (RR_TD) based on
File Accessing Frequency (FAF) for file of File Size
0.77 GB. It is shown in the table 7.
The change in NR value will be reflected to OS,
Cost, and RR_TD parameter values. In the existing
system, the NR value decided based on disk failure
rate benchmark of NR is 3 which is the convention
strategy.
For an example, the above table represents the
file with File Size (FS) 0.77 in GB is uploaded and
accessed in different scenarios.
• In the proposed system, the NR is decided
based on DRRRA and DRCAES approaches.
• The Occupied Space (OS) is calculated using
following way,
Occupied Space (OS) = File Size (FS) *
Number of Replicas (NR)
• The cost is calculated in the following way,
Cost = Occupied Space (OS) * 0.00067
(0.00067 is the amount incurred for 1 GB per
day which is adopts based on Google Drive
Cost plan. This is only for testing purpose)
• The RR_TD values are obtained by the use
of MATLAB tool.
These values are calculated based on different
File Accessing Frequency (FAF). The table values
are presented in graphical representations.
Fig. 3 shown the comparison of changes in
Number of Replicas parameter in different File
Accessing Frequency (FAF). From the graph, we
can clearly understand the NR is standard in existing
PRC, either 2 or 3 in DRRRA approach, and it is
vary based on FAF in DRCAES approach.
Fig. 4 presents the comparison of Occupied
Space (OS) and different File Accessing Frequency
(FAF) range. The graph is boons for the clear
understanding the reflections done due to the NR.
The OS is also standard in existing PRC because of
Standard NR, minimized for rarely accessed files in

Received: November 18, 2017 170
International Journal of Intelligent Engineering and Systems, Vol.11, No.2, 2018 DOI: 10.22266/ijies2018.0430.18
Figure.3 FAF vs. No. of replica
Figure.4 FAF vs. Occupied space
Figure.5 FAF vs. Cost
DRRRA approach, but optimized based on FAF in
DRCAES approach.
Fig. 5, the comparison of Cost in dollars ($) and
different File Accessing Frequency (FAF) range is
presented. The graph is determines the changes in
cost due to the Occupied Space (OS) modification.
The cost is also a regular in existing PRC because of
Standard OS, minimized for rarely accessed files in
Figure.6 FAF vs. Request-response rime delay.
DRRRA approach, but in DRCAES approach, it is
optimized based on FAF.
Fig. 6 shows the Request-Response Time Delay
(RR_TD) in the sec. and different File Accessing
Frequency (FAF) range. The graph values
represented based on MATLAB tool result and the
reflection of the RR_TD values is determined by the
value of NR. The RR_TD is worst in existing PRC
and DRRRA approach for high FAF range because
of Standard NR, but it is reduced in DRCAES
approach, because of optimized NR based on FAF.
Similarly, the table 7 presents the comparison of
parameters such as Number of Replicas (NR),
Occupied Space (OS), Cost, Request-Response
Time Delay (RR_TD) along with reliability and
availability concern of the proposed DRCAES with
DRRRA and exiting PRC algorithm. The optimized
cost obtained without affecting the existing
reliability assured percentage.
From the table, we can understand the DRCAES
provides an efficient data storage on cloud
computing environment with reliability, availability
concerns in a cost-effective manner.
The benefits of DRCAES approach is listed
below which stated in section 1. That all are proved,
 Dynamically predict rarely accessed files and
most frequently accesses the file using FAFR
model.
 Through DRRRA dynamically reduce the
number of replicas of that rarely accessed
files, so that, the cost and occupied space is
minimized. For reduction of the replica, it
finds minimum available Space of DC
among DC’ where that file exists (Removal).
 In DRCAES, dynamically create and place
the new replica for the most accessed file if
the frequency of each replica is equally
accessed otherwise it won’t replicate. The
new replica placed in the data center that has

Received: November 18, 2017 171
International Journal of Intelligent Engineering and Systems, Vol.11, No.2, 2018 DOI: 10.22266/ijies2018.0430.18
Table 8. Comparison of parameters for existing and proposed algorithms
Number
Replica
Occupied
Space
Cost
Reliability
Availability
Request-
Response
Time Delay
Existing
PRC
[3]
1 or 2 or 3
Decide
Based on
Disk Failure
Rate
Minimized
Based on
Replica
Minimized
Based on
Replica
1-Replica No
reliability
2-replica95%
Assured
3-replica99%
Not
Considered
Increased for
more request
Proposed
DRRRA
2 or 3
decide
Based on
FAF
Minimized
for Rarely
Accessed
File.
Minimized
for Rarely
Accessed
File.
2-replica 95%
Assured
[3]
Not
Considered
Increased
for more
request
Proposed
DRCAES
2-Replica is
Minimum
and
maximum is
decided
Based on
SLA
optimized
optimized
2-replica95%
Assured [3]
Enhanced
Decreased
more available space and that file does not
exist.
 Balanced storage retained during removal
and new replica placement. Because, it
analyses all aspects of Storage system like
available Space, SLA, Accessing frequency
of all existing replica.
6. Conclusion
To minimize the request response time and delay
of data placement for time-varying workload
applications, user necessity optimally makes use of
the time difference between storage and network
services across multiple cloud service provider. The
previous work of this research dynamically predicts
rarely accessed files with a help of FAFR algorithm
and reduces the number of replicas for that file, if it
satisfies the time limit using DRRRA algorithm.
Similarly, the proposed DRCAES algorithm
dynamically predicts most frequently accessed files
with the help of the FAFR algorithm. Then, it
creates a new replica for that file and finds the data
center that has more available space and doesn’t
have that file. However, this work achieves the
optimizing occupied space, cost, server performance,
increased server’s service delivery speed and
decreased request-response time delay. Thus,
ultimately the proposed DRCAES provide an
efficient data storage with an optimized cost without
affecting reliability, availability concerns for the
cloud also by optimizing the number of replicas
based on the user need and SLA. The proposed
algorithm achieves better result when compared to
the existing algorithms. In future replica
management during the disaster could be considered
without affecting the reliability and availability
concerns with minimum replica.
References
[1] R. Han, M.M. Ghanem, L. Guo, Y. Guo, and M.
Osmond, “Enabling cost-aware and adaptive
elasticity of multi-tier cloud applications”,
Elsevier, Future generation power systems in
Science Direct, Vol.32, No.1, pp. 82-98, 2014.
[2] M. Du and F. Li, "ATOM: Efficient Tracking,
Monitoring, and Orchestration of Cloud
Resources", IEEE Transactions on Parallel &
Distributed Systems, Vol. 28, No.8 , pp. 2172-
2189, 2017.
[3] W. Li, Y. Yang, and D. Yuan, “Ensuring Cloud
Data Reliability with Minimum Replication by
Proactive Replica Checking”, IEEE Transactions
on Computers, Vol. 65, No. 5, pp. 1494-1506,
2016.
[4] S. Souravlas, and A. Sifaleras, "Binary-Tree
Based Estimation of File Requests for Efficient
Data Replication", IEEE Transactions on
Parallel & Distributed Systems, Vol. 28, No. 7,
pp. 1839-1852, 2017.
[5] R. Han, S. Huang, Z. Wang, and J. Zhan, "CLAP:
Component-Level Approximate Processing for
Low Tail Latency and High Result Accuracy in
Cloud Online Services", IEEE Transactions on

Received: November 18, 2017 172
International Journal of Intelligent Engineering and Systems, Vol.11, No.2, 2018 DOI: 10.22266/ijies2018.0430.18
Parallel & Distributed Systems, Vol. 28, No.8 ,
pp. 2190-2203, 2017.
[6] U. Tos, R. Mokadem, A. Hameurlain, T. Ayav,
and S. Bora “A Performance and Profit Oriented
Data Replication Strategy for Cloud Systems”,
In: Proc. of IEEE Conferences on Ubiquitous
Intelligence & Computing, Toulouse, France,
pp. 780-787, 2016.
[7] D.T. Nukarapu, B. Tang, L. Wang, and S. 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.
[8] A. Kumar, R. Tandon, and T.C. Clancy, “On the
Latency and Energy Efficiency of Distributed
Storage Systems”, IEEE Transactions on Cloud
Computing, Vol. 5, No 2, pp. 221- 233, 2017.
[9] M. Hadji, “Scalable and Cost-Efficient
Algorithms for Reliable and Distributed Cloud
Storage”, Springer International Publishing
Switzerland, Vol.581, No.1, pp. 3-12, 2016.
[10] S.Q. Long, Y.L. Zhao, and W. Chen, “MORM:
A Multi-objective Optimized Replication
Management strategy for cloud storage cluster”,
Journal of Systems Architecture, Vol. 60, No.1,
pp. 234-244, 2014.
[11] S. Rampersaud and D. Grosu, "Sharing-Aware
Online Virtual Machine Packing in
Heterogeneous Resource Clouds", IEEE
Transactions on Parallel & Distributed Systems,
Vol. 28, No. 7, pp. 2046-2059, 2017.
[12] Y. Lin and H. Shen, “EAFR: An Energy-
Efficient Adaptive File Replication System in
Data-Intensive Clusters”, IEEE Transactions on
Parallel and Distributed Systems , Vol. 28, N,
pp. 1017-1030, 2017.
[13] L. Shi, Z. Zhang, and T. Robertazzi, "Energy-
Aware Scheduling of Embarrassingly Parallel
Jobs and Resource Allocation in Cloud", IEEE
Transactions on Parallel & Distributed Systems,
Vol. 28, No. 8, pp. 1607-1620, 2017.
[14] S.A.E. Selvi and R. Anbuselvi, “Ranking
Algorithm Based on File’s Accessing
Frequency for Cloud Storage System”,
International Journal of Advanced Research
Trends in Engineering and Technology, Vol. 4,
No. 9, pp. 29-33,2017.
[15] S.A.E. Selvi and R. Anbuselvi, “Optimizing the
Storage Space and Cost with Reliability
Assurance by Replica Reduction on Cloud
Storage System”, International Journal of
Advanced Research in Computer Science, Vol.
8, No.8, pp.327-332, 2017.