E018142329

IOSR Journal of Computer Engineering (IOSR-JCE)
e-ISSN: 2278-0661,p-ISSN: 2278-8727, Volume 18, Issue 1, Ver. IV (Jan – Feb. 2016), PP 23-29
www.iosrjournals.org
DOI: 10.9790/0661-18142329 www.iosrjournals.org 23 | Page
A Review: Hadoop Storage and Clustering Algorithms
Latika Kakkar1
, Gaurav Mehta2
1
(Department of Computer Science and Engineering, Chitkara University, India)
2
(Department of Computer Science and Engineering, Chitkara University, India)
Abstract : In the last few years there has been voluminous increase in the storing and processing of data,
which require convincing speed and also requirement of storage space. Big data is a defined as large, diverse
and complex data sets which has issues of storage, analysis and visualization for processing results. Four
characteristics of Big data which are–Volume, Value, Variety and Velocity makes it difficult for traditional
systems to process the big data. Apache Hadoop is an auspicious software framework that develops applications
that process huge amounts of data in parallel with large clusters of commodity hardware in a fault-tolerant and
veracious manner. Various performance metrics such as reliability, fault tolerance, accuracy, confidentiality
and security are improved as with the use of Hadoop. Hadoop MapReduce is an effective Computation Model
for processing large data on distributed data clusters such as Clouds. We first introduce the general idea of big
data and then review related technologies, such as could computing and Hadoop. Various clustering techniques
are also analyzed based on parameters like numbers of clusters, size of clusters, type of dataset and noise.
Keywords: Big data, Cloud Computing, Clusters, Hadoop , HDFS, MapReduce
I. Introduction
To analyze data from different aspects and to outline and sum up this data into valuable data i.e.
important information is known as Data Mining. Data Mining is process of analyzing correlations and patterns
among huge data in a database. Major data mining techniques involve classification, regression and clustering.
Clustering is the process of assigning data into groups called clusters such that object in same cluster is more
similar than the objects of other clusters. Clustering is a main task of Data Mining. It common technique for
statistical data analysis used in many fields which includes pattern recognition, analysis of image, information
retrieval. Now a day’s big data is growing rapidly specially those related to internet companies. For example,
Google, Facebook processes petabytes of data within a month.Figure1 shows the increase of the data volume in
a rapid manner. There are various challenges as the datasets are increasing in a drastic manner. Due to
advancement of information technology huge amount of data can be generated. On an average, YouTube
uploads 72 hours of videos in every minute [1]. This creates the problem of collection and integration of
tremendous data from widespread sources of data. The accelerated advancement of cloud computing and the
Internet of Things further contributes the fine increase of data. Through this increase in volume and variety there
is a great accentuation on the existing computing capacity. This huge data causes problem of storing and
managing such large complex datasets with fulfilling the hardware and software infrastructure. The mining of
such complex, heterogeneous and voluminous data at different levels of analyzing, perception, anticipating
should be done carefully to unfold its natural properties so as to filter good decision making. [2]
II. Challenges Of Big Data
2.1 Confidentiality of Data
Various big data service providers use different tools to process and analyze data, due to which there
are security risks. For example, the transactional dataset generally includes a set of complete operating data to
execute the important business processes. Such data consists of lower granularity and delicate information like
credit card numbers. Therefore, preventive measures are taken to protect such sensitive data, to ensure its safety.
2.2 Representation of Data
The datasets are heterogeneous in nature. Representation of Data aims to make meaningful data for
analysis of computers and user comprehension. Improper representation of data affects the data analysis.
Efficient data representation shows data structure, class and technologies which enables useful operations on
different datasets.
2.3 Management of Data Life Cycle
The storage system used normally cannot support huge quantity of data. Data freshness is the key
component which describes the values private in big data. Therefore, to decide which data shall be stored and
which data shall be ignored, a principle associated with the analytical value should be developed.

A Review: Hadoop Storage And Clustering Algorithms
2.4 Analytical Mechanism:
The analytical system processes large amount of heterogeneous data with limited time period.
However, the traditional RDBMSs lack scalability and expandability, by which performance requirement is
affected. Non-relational databases are advantageous in the analysis of unstructured data but there are still some
problems in performance of non-relational databases. Few enterprises have employed hybrid architecture of
database that has advantages of both types of databases e.g. Facebook and Taobao.
2.5 Energy Management
Large amount of electrical energy is consumed for the analysis, storage, and transmission of big data.
Therefore, power-consumption control and management mechanism of system or computer were required for
big data with the constraints that the expansion and accessibility are ensured.
2.6 Scalability
The analytical system designed for big data should be compatible with all types of datasets. It must be
able to process drastically increasing complex datasets.
2.7 Reduction of Redundancy and Data Compression
Redundancy reduction and data compression is very useful to diminish the cost of the entire system on
the basis that there is no affect on the data. For example, mostly the data developed by sensor networks have
high degree of redundancy that may be compressed and filtered. [3-5]
Fig. 1 Composition of Big Data
III. Related Technologies
Fundamental technologies that are closely related to big data:
3.1Cloud Computing
Cloud computing refers to parallel and distributed system that consists of computers that is visualized
and inter-connected. This system is represented as a computing resource based on service agreements
established through intervention between the customers and the service providers. Cloud computing enables the
organizations to consume the resources as a utility rather than developing and maintaining the internal
computing infrastructures.
Benefits of cloud computing:
 Pay per use: Users can pay only for the resources and workload they use which means that the computing
resources are measured at a granular level.
 Self-service provisioning: The computing resources can be operated by end users for the workload on
demand.
 Elasticity: With the increase and the decrease of the computing needs of the companies, they can scale up
and scale down the resources.
3.2 Cloud Computing and Big Data Affiliation
Cloud computing and big data are closely related. The key components of cloud computing are shown
in Fig.2. For the storage and processing of big data cloud computing is the solution. On the other hand, the big
data also increases the development of cloud computing. The technology of cloud computing of distributed
storage efficiently manages the big data and the parallel computing capability of cloud computing improves the
big data analysis.

Fig. 2 Key components of cloud computing
3.3 Hadoop
The storage infrastructure has to provide two basic requirements. First is to provide service for
information storage and secondly to provide interface for efficient query access and data analysis. Traditionally,
structured RDBMSs were used for storing, managing and data analysis but with the increasing growth of big
data, efficient and reliable storage mechanism is needed. Hadoop is a project of the Apache Software
Foundation. Hadoop [6] is an open-source software framework implemented using Java and is designed to be
used on large distributed systems of commodity hardware. The software framework of Hadoop is designed in
such a way that it manages hardware failure automatically. Hadoop is very popular software tool due to it being
open-source. Yahoo, Facebook, Twitter, LinkedIn, and others are currently using Hadoop. [8]
3.3.1 Hadoop Distributed File System (HDFS)
The Hadoop Distributed File System (HDFS) is the file system component of the Hadoop framework.
Hadoop Distributed file System is used to store huge data on commodity hardware [6]. HDFS is designed for a
large amount of big data files. The size of data to be stored using Hadoop Distributed file System could be sized
from gigabytes to terabytes [7]. The property and feature of HDFS is that it can support millions of files and has
scalability to hundreds of nodes. HDFS has separate storage places for metadata and application data [8]. The
metadata is stored in the NameNode server and application data are stored on DataNodes which contain blocks
of data. The communication between the different nodes of the HDFS is done using a TCP-based protocol [8].
Fault tolerance is the ability of a system to function adequately and correctly even after some system
components gets failed. Hadoop and HDFS are highly fault tolerant as HDFS can be spread over multiple nodes
that is cost effective. It duplicates the data so that if one copy of data is lost then still there are backup copies
available. [9]
3.3.2 Mapreduce Processing Model
Hadoop MapReduce processes big data in parallel and provides output with efficient performance.
Map-reduce consist of Map function and Reduce function. Map function executes filtering and sorting of large
data sets. Reduce function performs the summary operation which combines the result and provides the
enhanced output. Hadoop HDFS and Map-Reduce are delineated with the help of Google file system. Google
File System (GFS) is developed by Google is a distributed file system that provide organized and adequate
access to data using large clusters of commodity servers. Map phase: The Master node accepts the input and
then divides a large problem is into smaller sub-problems. It then distributes these sub-problems among worker
nodes in a multi-level tree structure. These sub-problems are then processed by the worker nodes which execute
and sent the result back to the master node. Reduce phase: Reduce function combines the output of all sub-
problems and collect it in master node and produces final output. Each map function is associated with a reduce
function.

Fig 3 A Multinode Hadoop Cluster
Figure 3 shows a multi-node Hadoop cluster. It consists of a Master node and slave nodes. Task tracker
and datanode operate on slave nodes. Master node consists of Task tracker, Datanode, Job tracker and
Namenode. The Namenode is used to operate the file system. The job scheduling is executed by job tracker. The
datanode stores the information which is in knowledge of the Namenode. On user interaction with Hadoop
system, the Namenode becomes active and all the information about datanode data is processed by the
Namenode.
Fig.4 HDFS file system
Figure 4 shows the HDFS file system. Namenode is used to access and store user data. Thus Namenode
becomes a single point of failure. For that a secondary Namenode is always active which takes snapshots from
Namenode and stores all the information itself. On failure of the Namenode, all the data can be recovers from
the secondary Namenode. Hadoops’s HDFS can handle Fault Tolerance, but there are some other points of
Hadoop and HDFS that could require some improvement. There is the problem of scalability of the NameNode
[8]. The Namenode could subject to physical hardware constraints because Namenode is the single node.For
example, because the namespace is kept in memory, there could be probability of HDFS to become
unresponsive if the namespace grows to aa limit that it has to use node’s memory. The solution to this issue is to
use more than one Namenode but this can create problems in communication [8]. Another solution to this
problem is to store only large data files in the HDFS. By using large data files the size of namespace will be
reduced. Moreover the file system like HDFS is designed for large data files. Hadoop also lacks in providing
good high level support. This issue can occur with open source software. Enhancing Hadoop’s functionality on a
system can be difficult without proper support [11]. HDFS is also sensitive to scheduling delays which restricts
it to provide its full potential. Thus the node could have to wait for its new task. This issue is particularly found
in HDFS client code [12].
4 Comparisons of Various Clustering Algorithms
4.1K-Means Algorithm
K-Means is a clustering algorithm that partitions objects into k clusters, where each cluster has a centre
point called centroid clustering algorithm finds the desired number of distinct clusters and their centroids.
Algorithmic steps for K-Means clustering [21]
1) To find the value of K. K is the number of desired clusters.
2) Initialization – Now choose starting points that can be used to estimate centroids of clusters.
3) Classification – To check each object in the dataset and assigning it to the centroid nearest to it.
4) Calculation of centroid – After assignment of the objects in the cluster, new k centroids is calculated.
5) Meeting criteria – The steps 3 and 4 are repeated until the objects stop changing their cluster s and centroid of
all clusters get fixed.

4.2 Fuzzy C-Mean
Algorithmic steps for Fuzzy C-Means clustering [22]
We let X = {x1, x2, x3 ..., xn} be the set of data points and V = {v1, v2, v3 ..., vc} be the set of centers.
1) Randomly ‘c’ cluster centers are selected.
2) The fuzzy membership 'µij' is calculated using:
3) Then fuzzy centers 'vj' are computed using:
4) Steps 2and 3 are repeated until the minimum ‘J’ value is achieved or
|| U (k+1)
-U (k)
||< β, where, ‘k’ is the iteration step, ‘β’ = termination criterion between [0, 1], ‘U’= (µij) n*c’ is the
fuzzy membership matrix, ‘J’ = objective function.
FCM iteratively changes the centers of clusters to exact location in a dataset. When fuzzy logic is
introduced in K-Means clustering algorithm, it becomes Fuzzy C-Means algorithm. FCM clustering is based on
fuzzy behavior. The structure of FCM is basically similar to K-Means. In [20] experimental results showed that
K-Means takes less elapsed time than FCM. It can be concluded that in K-Means algorithm, the number of final
clusters are to be defined beforehand. Such algorithms may be susceptible to local optima, can be outlier-
sensitive and memory space required. The time-complexity of the K-Means algorithm is O(ncdi) and that of
FCM algorithm is O(ndc2i). [20] Concluded that K-Means algorithm is better than FCM algorithm. Fuzzy c-
means requires high computation time than K-Means because in FCM, fuzzy measures are also required to be
calculated. FCM can handles issues related to pattern understanding, noise in data and thus provides faster
solutions. FCM can be used for discovering association rules, retrieval of image and functional dependencies.
4.3 Hierarchical Clustering
Hierarchical clustering is a clustering algorithm which is used to build hierarchy of clusters. There are
two strategies for hierarchical clustering:
Agglomerative: In Agglomerative each object initiates in its own cluster and cluster pairs are merged as we
move up in the hierarchy. Agglomerative is bottom up approach.
Divisive: In divisive all the observations start in one cluster, and when we move downwards, splitting of cluster
is performed iteratively. Divisive is top down approach. Dendogram is used to show results of hierarchical
clustering. The complexity of agglomerative clustering is O(n3
),thus in case of large datasets performance of
agglomerative clustering is too slow. The complexity of divisive clustering is O(2n
), which is more worse.
Hierarchical clustering has limitations linear time complexity. These algorithms can also be used as a combined
approach to produce better results. Agglomerative hierarchical clustering is used because of its quality and K-
means is used because of its run-time efficiency. Comparative analysis of various clustering algorithms, K-
means, Hierarchical, Self Organization map algorithm, Expectation Maximization algorithm and Fuzzy-c
algorithm is shown in fig 5. [21, 22, 23, 24, 25]
4.4 Self-Organization Map (SOM)
The SOM [26] is self organization map is used to find better mapping from high dimensional input
space to a two dimensional node representation. In SOM, object represented by same node are grouped in same
cluster.
SOM algorithm pseudo code [26]:
1. Size and dimension of map is chosen.
2. With every new sample vector:
2.1 The distance between every codebook vector and new vector is computed
2.2 Using learning rate that decreases in time and distance radius on the map, all codebook vectors with new
vector is computed.

Table 1. Comparative analysis of various clustering algorithms
IV. Conclusion
For efficient processing of Big data, Hadoop proves to be a good software. HDFS is a very large
distributed file system that uses commodity clusters and provides high throughput as well as fault tolerance.
This is very useful property of Hadoop because in case of data loss or system crashes, it can be adverse to
business with irreversible consequences. For efficient retrieval of data from Hadoop , comparative analysis of
various clustering algorithm is made based on various input parameters (Table1).
References
[1] Mayer-Schönberger V, Cukier K, Big data: A revolution that will transform how we live, work, and think. Houghton Mifflin
Harcourt, 2013.
[2] Chen M, Mao S, Liu Y, Big data: A survey: Mobile Networks and Applications, 19(2), 2014 Apr 1, 171-209.
[3] Labrinidis A, Jagadish HV, Challenges and opportunities with big data, Proceedings of the VLDB Endowment, 5(12), 2012 Aug
1,2032-3.
[4] Chaudhuri S, Dayal U, Narasayya V, An overview of business intelligence technology, Communications of the ACM,54(8), 2011
Aug 1,88-98.
[5] Agrawal D, Bernstein P, Bertino E, Davidson S, Dayal U, Franklin M, Gehrke J, Haas L, Halevy A, Han J, Jagadish HV,
Challenges and Opportunities with big data, 2011-1.
[6] Apache Hadoop, http://hadoop.apache.org
[7] Borthakur D, The hadoop distributed file system: Architecture and design, Hadoop Project Website, 2007 Aug 11, 1.
[8] Shvachko K, Kuang H, Radia S, Chansler R, The hadoop distributed file system. In Mass Storage Systems and Technologies
(MSST), 2010 IEEE 26th Symposium ,2010 May 3, 1-10.
[9] Evans J. Fault Tolerance in Hadoop for Work Migration. CSCI B34 Survey Paper, 3(28),2011,11.
[10] Bessani AN, Cogo VV, Correia M, Costa P, Pasin M, Silva F, Arantes L, Marin O, Sens P, Sopena J. Making Hadoop MapReduce
Byzantine Fault-Tolerant, DSN, Fast abstract, 2010.
[11] Wang F, Qiu J, Yang J, Dong B, Li X, Li Y. Hadoop high availability through metadata replication. In Proceedings of the first
international workshop on Cloud data management, 2009 Nov 2, 37-44.
[12] Shafer J, Rixner S, Cox AL, The hadoop distributed filesystem: Balancing portability and performance, In Performance Analysis of
Systems & Software (ISPASS), 2010 IEEE International Symposium, 2010 Mar 28,122-133.
[13] Dean J, Ghemawat S, MapReduce: simplified data processing on large clusters, Communications of the ACM. 51(1),2008 Jan
1,107-13.
[14] Xie J, Yin S, Ruan X, Ding Z, Tian Y, Majors J, Manzanares A, Qin X, Improving mapreduce performance through data placement
in heterogeneous hadoop clusters, In Parallel & Distributed Processing, Workshops and Phd Forum (IPDPSW), 2010 IEEE
International Symposium, 2010 Apr 19, 1-9.
[15] Cattell R. Scalable SQL and NoSQL data stores , ACM SIGMOD Record, 39(4), 2011 May 6, 12-27.
[16] Chaiken R, Jenkins B, Larson PÅ, Ramsey B, Shakib D, Weaver S, Zhou J, SCOPE: easy and efficient parallel processing of
massive data sets, Proceedings of the VLDB Endowment, 1(2), 2008 Aug 1, 1265-76.
[17] Grossman RL, Gu Y, Sabala M, Zhang W, Compute and storage clouds using wide area high performance networks, Future
Generation Computer Systems, 25(2), 2009 Feb 28,179-83.
[18] Liu X, Han J, Zhong Y, Han C, He X, Implementing WebGIS on Hadoop: A case study of improving small file I/O performance on
HDFS, In Cluster Computing and Workshops, CLUSTER '09. IEEE International Conference, 2009 Aug 31, 1-8.

[19] Grace RK, Manimegalai R, Kumar SS, Medical image retrieval system in grid using Hadoop framework, In Computational Science
and Computational Intelligence (CSCI), International Conference, Vol. 1, 2014 Mar 10, 144-148.
[20] Ghosh S, Dubey SK, Comparative analysis of k-means and fuzzy c-means algorithms, International Journal of Advanced Computer
Science and Applications,4(4),2013,34-9.
[21] Kanungo T, Mount DM, Netanyahu NS, Piatko CD, Silverman R, Wu AY, An efficient k-means clustering algorithm: Analysis and
implementation, Pattern Analysis and Machine Intelligence, IEEE , 24(7), 2002Jul, 881-92.
[22] Rao VS, Vidyavathi DS. Comparative investigations and performance analysis of FCM and MFPCM algorithms on iris data, Indian
Journal of Computer Science and Engineering,1(2),2010,145-51.
[23] Joshi A, Kaur R, A review: Comparative study of various clustering techniques in data mining, International Journal of Advanced
Research in Computer Science and Software Engineering, 3(3), 2013 Mar, 55-7.
[24] Mingoti SA, Lima JO, Comparing SOM neural network with Fuzzy c-means, K-means and traditional hierarchical clustering
algorithms, European Journal of Operational Research, 174(3), 2006 Nov 1, 1742-59.
[25] Steinbach M, Karypis G, Kumar V, A comparison of document clustering techniques, In KDD workshop on text mining, Vol. 400,
No. 1, 2000 Aug 20,525-526.
[26] Abbas OA, Comparisons between Data Clustering Algorithms, Int. Arab J. Inf. Technol... 5(3), 2008 Jul 1,320-5.

E018142329

More Related Content

What's hot (19)

Viewers also liked (20)

Similar to E018142329 (20)

More from IOSR Journals (20)

Recently uploaded (20)

E018142329