Memory Management in BigData: A Perpective View

@ IJTSRD | Available Online @ www.ijtsrd.com
ISSN No: 2456
International
Research
Memory Management in BigData: A Perpective View
Bhavna Bharti
1
PG Scholar,
Department of CSE,
ABSTRACT
The requirement to perform complicated statistic
analysis of big data by institutions of engineering,
scientific research, health care, commerce, banking
and computer research is immense. However, the
limitations of the widely used current desktop
software like R, excel, minitab and spss gives a
researcher limitation to deal with big data and big data
analytic tools like IBM BigInsight, HP Vertica, SAP
HANA & Pentaho come at an overpriced license.
Apache Hadoop is an open source distributed
computing framework that uses commodity hardware.
With this project, I intend to collaborate Apache
Hadoop and R software to develop an an
platform that stores big data (using open source
Apache Hadoop) and perform statistical analysis
(using open source R software).Due to the limitations
of vertical scaling of computer unit, data storage is
handled by several machines and so analysis
distributed over all these machines. Apache Hadoop is
what comes handy in this environment. To store
massive quantities of data as required by researchers,
we could use commodity hardware and perform
analysis in distributed environment.
Keyword: Minitab, SPSS, Machine learning, IBM
BigInsight, HP Vertica, SAP HANA, Pentaho, Apache
Hadoop, R, Big Data
1. INTRODUCTION
The memory management technique to deal with big
data to perform linear regression and similar
predictive analysis with ease and prove to be very
helpful for engineering research, business, health care,
scientific research, banking & finance and machine
learning where complicated statistical analysis can be
@ IJTSRD | Available Online @ www.ijtsrd.com | Volume – 2 | Issue – 4 | May-Jun
ISSN No: 2456 - 6470 | www.ijtsrd.com | Volume
International Journal of Trend in Scientific
Research and Development (IJTSRD)
International Open Access Journal
Bhavna Bharti1
, Prof. Avinash Sharma2
PG Scholar, 2
Assistant Professor and MTech Co-ordinator
of CSE, Millennium Institute of Technology and Science
Bhopal, Madhya Pradesh, India
The requirement to perform complicated statistic
analysis of big data by institutions of engineering,
health care, commerce, banking
and computer research is immense. However, the
limitations of the widely used current desktop
software like R, excel, minitab and spss gives a
researcher limitation to deal with big data and big data
BigInsight, HP Vertica, SAP
HANA & Pentaho come at an overpriced license.
Apache Hadoop is an open source distributed
computing framework that uses commodity hardware.
With this project, I intend to collaborate Apache
Hadoop and R software to develop an analytic
platform that stores big data (using open source
Apache Hadoop) and perform statistical analysis
(using open source R software).Due to the limitations
of vertical scaling of computer unit, data storage is
handled by several machines and so analysis becomes
distributed over all these machines. Apache Hadoop is
what comes handy in this environment. To store
massive quantities of data as required by researchers,
we could use commodity hardware and perform
initab, SPSS, Machine learning, IBM
BigInsight, HP Vertica, SAP HANA, Pentaho, Apache
The memory management technique to deal with big
data to perform linear regression and similar
predictive analysis with ease and prove to be very
helpful for engineering research, business, health care,
scientific research, banking & finance and machine
arning where complicated statistical analysis can be
performed. Analysis of large data that is very
complicated for traditional analytic environment is
done with ease in distributed environment without
undermining on the quality of the result.
Entrepreneurship these days demands the gathering of
information that may extend to even petabytes.
Statistics based on these customer feedback data will
help expand businesses and a company that has such
data to its disposal, surely has a far stronger feel on
the pulse of the market.
The following presentation discusses the scope of
large scale data analysis and developing systems that
support it for application at industrial level. Since
most software packages available today focuses on
the main memory of an ave
single server, a researcher is forced to utelise vertical
scalability or choose a random sample data to work
upon. Vertical scalability is costly and even if this
factor is overlooked, there is a limit upto which
vertical scalability can be performed. But instead if
one chooses a random sample, it may not always be
the best representative of all the data collected.
Even complicated data analysis that require the entire
data collected by the researcher, may be dealt with in
this platform. These results are definitely more
accurate than that given by the analysis of randomized
sample [1][2]. Statistical software packages like the
R, SPSS, SAS, or Matlab that are useful in the
analysis of only modest quantities of data forces the
user to either use more powerful computers that are
rare and expensive or pick random samples that may
cause results that are compromised on. Though the
development of many Data Management Systems has
occurred over the years, since the emphasis is on data
Jun 2018 Page: 1993
6470 | www.ijtsrd.com | Volume - 2 | Issue – 4
Scientific
(IJTSRD)
International Open Access Journal
nd Science,
performed. Analysis of large data that is very
complicated for traditional analytic environment is
done with ease in distributed environment without
undermining on the quality of the result.
urship these days demands the gathering of
information that may extend to even petabytes.
Statistics based on these customer feedback data will
help expand businesses and a company that has such
data to its disposal, surely has a far stronger feel on
The following presentation discusses the scope of
large scale data analysis and developing systems that
support it for application at industrial level. Since
most software packages available today focuses on
the main memory of an average sized dataset in a
single server, a researcher is forced to utelise vertical
scalability or choose a random sample data to work
upon. Vertical scalability is costly and even if this
factor is overlooked, there is a limit upto which
y can be performed. But instead if
one chooses a random sample, it may not always be
the best representative of all the data collected.
Even complicated data analysis that require the entire
data collected by the researcher, may be dealt with in
atform. These results are definitely more
accurate than that given by the analysis of randomized
sample [1][2]. Statistical software packages like the
R, SPSS, SAS, or Matlab that are useful in the
analysis of only modest quantities of data forces the
to either use more powerful computers that are
rare and expensive or pick random samples that may
cause results that are compromised on. Though the
development of many Data Management Systems has
occurred over the years, since the emphasis is on data

International Journal of Trend in Scientific Research and Development (IJTSRD) ISSN: 2456-6470
@ IJTSRD | Available Online @ www.ijtsrd.com | Volume – 2 | Issue – 4 | May-Jun 2018 Page: 1994
storage, processing and simple computing, data
analysis is largely left neglected. The analysis
functionality is overlooked by researchers as they feel
that these Data Management Systems are competent
enough at data storage, data processing and simple
computing. Even so, the void of an apt analytically
functional statistical package is undeniable. By
collaborating R and Hadoop, I intend to create a
scalable platform for intense analytics and this project
is work of ongoing research in generating the resultant
Memory Management in Big Data.
2. LITERATURE SURVEY
2.1 Open Source R Programming Language
The R[6][7][8][10] is a free software programming
language used among statisticians and data miners for
statistical computing, graphics and several such
applications. When Ross Ihaka and Robert Gentleman
created “R” in the year 1993, the appreciation and
world wide acceptance [14] it generated, prompted
them to make it available to the general public over
the internet under the GPL. With this move R became
open to a wider public and in no time, it took over
much of the statistical analysis over the machine and
development of the language with time made it the
most eligible and updated software programming
language in the industry. R was released on 29th
February 2000 can be used freely, distributed and sold
to any receiver who possess the same rights. The
flexibility of the language due to it’s being based on a
formal computer language, makes it the most sought
after software programming language in the market.
R is rich in various statistical analysis packages.
There are 5922 packages available in CRAN packages
repository [16]. This is again increasing
exponentially. There are many R users and many
forums as well as groups from where you can be get
concept mind as well as tutorial and support available.
2.2 Open Source Apache Hadoop
Apache Hadoop[3][4][5][7][11][12] readies us with
an easy to use, dependable and accurate distributed
computing software whose main purpose is to enable
the researcher to shift from single server to multiple
connected machines each functioning as a separate
unit as far as data storage and computing is
concerned. By utilizing the Apache Hadoop software
library that has several thousand computers over a
cluster, one can be cent percent assured of the
accuracy of analysis as each unit is so programmed as
to be able to spot and rectify errors.
Hadoop stores massive quantities of data over several
systems in the cluster and comes up with solutions
through a highly scalable, distributed batch processing
system. In this way, a large workload is distributed
over the cluster of several inexpensive datasets and
big data analysis is done over with in a very short
time. The Apache Hadoop framework is composed of
the following modules:
1. Hadoop Common: This is the basic module that is
concerned with libraries and ther such utilities that are
essential for the optimal functioning of other hadoop
modules.
2. Hadoop Distributed File System (HDFS™): This is
the module that concerns itself with data storage over
several storage systems wherein information is
submitted to commodity machines in a classified
fashion. It is under the functionality of this module
that the bandwidth of a cluster is kept at a high level.
3. Hadoop YARN: Where YARN stands for Yet
Another Resource Negotiator, is as the name suggests
a task assigning and resource managing module
performing at the cluster level. It is a generic platform
required for any distributed application to run on.
4. Hadoop MapReduce: A sub project of the Apache
Hadoop project, this project is a yarn based system
that is concerned with distributing input and output
data in chunks for parallel processing.
2.3 RHadoop
RHadoop[7][10][14][15][17] is an open source
project aimed at large scale data analysts to empower
them to use the horizontal scalability oh Hadoop using
the R language. ravro, plymr, rmr, rhdfs and rhbase,
the 5 R packages enable its users to manage and
analyze massive quantities of information using
Hadoop.
1. ravro – It is the R package which permits the
reading and writing of files in avro format, to R.
2. Plyrmr – It is a more recent R package that makes
R and Hadoop perform in near perfect if not perfect
harmony, in the analysis of higher lever plyr like data.
3. Rmr – It is an R package that came into being to
allow users to write map reduce programs in R since it
is more productive and far easier.
4. Rhdfs – It is an R package that gives
administration of HDFS files from within R. It uses
Hadoop common to give access to map reduce file
services.
5. Rhbase – It is an R package that allows its users
basic connectivity to Hbase. Administrating a
database for Hbase by using R is made possible by
rhbase.

3. SYSTEM MODEL
This system is composed of three vital components R
statistical software to perform statistical analysis,
Hadoop framework to distribute data in the cluster
computing and Rhadoop that traverse over and link R
to Hadoop. Using this system, one can load
information from a local system to a hadoop library
through the hadoop distributed file system that reads
and analyses the input data. The four integral
elements that work together to associate R language
with hadoop are RAVRO, RMR, RHDFS, PLYRMR
& RHBASE.
The hadoop distributed file system that facilitates the
user to store, read, and write in hadoop distributed file
system. Given the inability of R in the statistical
analysis of massive information despite its excellent
performance with moderately sized data, when linked
with Hadoop, the resultant system is very promising
for several industries including stock market, artificial
intelligence designing and scientific research and
many more where business analytics need to be
performed.
Figure 1.1: System Model
3.1 Memory Management
Since apache hadoop is a cluster comprising of
several machines all performing on commodity
hardware, looking elsewhere for data storage is out of
question as it is simply possible to do horizontal
scaling of memory in each worker node or even the
master node.
The 2 challenges met with by R users due to its
constitutional property of reading memory by default
are
3.1.1 Reading data from Hadoop distributed file
system- Rhdfs and Rhbase both developed as parts of
the Rhadoop project helps researches using R, utilize
Hadoop hdfs and Hadoop base better. Of these,
Rhbase is what gives the cardinal connectivity to
hadoop distributed file systems. The tables in hbase
can be read, written and modified by a researcher
using R.
3.1.2. Computation Memory Management- Rhdfs
and by processing it either on the device or near it
reduces the area of conveyance. This is done in 2
steps
1. Map step: When a problem is assigned to a cluster,
the master nodes distributes it to smaller questions
and poses it to worker nodes that may distribute it to
smaller portions and distribute it among more nodes,
thereby creating a tree like pattern. When these
worker nodes come up with a solution, it is
transmitted back to the master node.
2. Reduce step: The solution s that are passed on to
the master node by the worker nodes are processed
into a single solution as required by the researcher.
As long as each mapping procedure is not intertwined
with another, parallelism is a possibility not conceived
yet due to the restrain in either the number of separate
storage systems or the number of CPUs in close
proximity to the operation site. And as long as
reduction procedures share the same key and are
submitted to the same reducer at the same time, a set
of reducers can perform the reduction phase with
ease.

Figure 1.2: MapReduce task of Wordcount
A petabyte of data can be processed by a cluster using
MapReduce. An advantage of the use of parallelism is
that an incomplete failure of server can be resolved if
the input data is still available, and even if one
mapper or reducer fails, the effort can be scheduled
for later.
4. RESULT AND DISCUSSION
The more concerning challenge has to deal with
massive data as R is constitutionally designed to read
the whole information in memory and scrutinize it all
to reach a solution when posed with a problem,
thereby creating 2 drawbacks that need be solved.
4.1 Data Memory Management
Depending on hardware configuration R objects can
occupy upto 2-4 GB exceeding which, the researcher
will be forced to use R and thereby analysis of big
data will have to be done. To overcome these
limitations, packages like ff[18], ffbase[18],
bigmemory[19] etc comes to play. Ff is a package that
stores large data efficiently on a disc that is quick to
access. However it can map only a portion of this data
n the main memory thereby acting as if in RAM and
sectioning the effective virtual memory consumption
per ff object. To raise performance standards to an
utmost, many improvising and simplifying approaches
like preprocessing and virtualization can be made use
of.
Ff packages could use external storage devices like
hard disk, CD, DVD etc to store binary files instead of
using it’s memory. It allows us to work on massive
data at one go by reading and writing file in a chunk.
The cost of increasing the memory of a system or
cluster is very high and thus it is more reasonable to
use R bearing in mind the fact that while performing
analysis, the entire data in memory will be taken into
consideration. Memory management is highly
efficient in R ff and ff base. But these packages are
compatible only with hadoop distributed file systems.
Figure 1.3: Working Mechanism of ff – package
Bigmemory, Biganalytics[20] and Biglm[21] are
programmes that can boost memory limitations.
Biganalytics has in it multiple analytic functions.
However ff is not compatible with hadoop and Biglm
that provides functions for linear regression is not
compatible with cluster systems.
library(ff)
library(ffbase)
library(biglm)
options(fftempdir = "c:/bin/")
regh<
read.table.ffdf(file="d:/sampledata/201201hourly.txt",
sep=",
",FUN="read.csv",header=T,VERBOSE=T,
next.rows=401000, colClasses=NA, na.strings="NA")
reg<
bigglm(WBAN~Time+StationType+SkyConditionFla
g+Visib
ility+DryBulbCelsiusFlag+DewPointCelsiusFlag,
data=regh)
system.time(reg<

bigglm(WBAN~Time+StationType+SkyConditionFla
g+Visib
ility+DryBulbCelsiusFlag+ DewPointCelsiusFlag,
data=regh))
object.size(regh)
nrow(regh)
ncol(regh)
Above linear regression analysis is performed and
result is measured in core2duo 2.0 GHz processor
with 3 Gb of RAM. The text file is a weather data of
one month with file size of 509 Mb and contain
4192912 rows and 44 columns. Assume if the file size
is 509 Mb for one month then it’ll be around
509*12Mb for 1 year. Definitely, people want to
perform analysis in data of entire year and even for
decade too. To perform analysis of large file is
possible using these packages but it is not good
practice when file size reach to terabytes and
petabytes and working with single node to perform
analysis where vertical scaling of memory is limited.
5. CONCLUSION AND FUTURE WORK
With the help of data memory management and
computational memory management techniques
handling large scale information has taken a new turn
by utterly changing memory requirement. The hurdles
in the processing of large data have thus been
decreased. Neither are we forced to deal with an
inappropriate random sample and compromise over an
incorrect statistical result, nor are we forced with the
limitation of vertical scaling that has a dead end. With
the addition of kmean, correlation, market basket,
time series and other such functionalities, the above
discussed model becomes all the more potent. By
combining opensource R and opensource Apache
hadoop, a cluster attains the marvel that only a
crossover between a cluster framework package and a
highly competent statistical programming language
can achieve. The security, availability, efficiency and
scalability it promises cannot be compared to
traditional database systems. when additional
programs like business intelligence or a link up with
systems like oracle or SAP will without doubt make it
the best analytic tool available in the industry. It
would also prove useful if in future the cluster could
maintain its optimization with a lesser number of
worker odes. In today’s hyper competitive world
where time is money if we can turn to cloud based big
data analytics when vital analysis needs to be
performed in a short time, big wealth too can be
sapped in a short time.
REFERENCES
1. Qingchao Cai, Hao Zhang, Wentian Guo, Gang
Chen, Beng Chin Ooi, Kian-Lee Ta, Weng-Fai
Wong, “MemepiC: Towards a Uniﬁed In-Memory
Big Data Management System”, IEEE
Transactions on Big Data, 2018.
2. Nikolay Laptev, Kai Zeng, Carlo Zaniolo., “Very
Fast Estimation for Result and Accuracy for Big
Data Analytics: the EARL System”, IEEE’s,
ICDE Conference 2013.
3. (Accessed on 12th November 2013). Hadoop
[Online] Available: http://hortonworks.com
4. (Accessed on 14th November 2013) Hadoop
[Online] Available:
http://en.wikipedia.org/wiki/Apache_Hadoop
5. T (Accessed on 14th November 2013) Hadoop
[Online] Available: http://hadoop.apache.org/
6. (Accessed on 12th December 2013) R [Online]
Available: http://www.revolutionanalytics.com
7. Vignesh Prajapati, Big Data Analytics with R and
Hadoop, Packt publication, 2013.
8. (Accessed on 18th January 2014) R [Online]
Available: http://cran.r-project.org/
9. Q. Ethan McCallum & Stephen Weston Parallel
R, O’Reilly, 2012.
10. Josep Adler, R IN A NUTSHELL, second edition,
O’Reilly, 2012.
11. Paul C. Zikopoulos, Chris Eaton, Dirk Deross,
Thomas Deutsch, George Lapis, Understanding
Big Data Analytics for Enterprise Class Hadoop
and Streaming Data, Mc Graw Hill, 2012.
12. Tom White, Hadoop:The Definitive Guide, 3rd
Edition, O’Reilly, 2012.
13. Sudipto Das, Yannis Sismanis Kevin, S. Beyer,
Rainer Gemulla, Peter J. Haas, and John
McPherson, "Ricardo: Integrating R and Hadoop",
In Proc. SIGMOD’10, Indianapolis, Indiana,
USA, June 6–11, 2010.
14. (Accessed on 19th November 2013) RHadoop
[Online] Available:
https://github.com/RevolutionAnalytics/RHadoop/
wiki

15. (Accessed on 15th December 2013) MapReduce
[Online] Availabe:
https://github.com/RevolutionAnalytics/rmr2/blob
/maste r/docs/tutorial.md
16. (Accessed on 26th October 2014) Cran Packages
Repository, [Online]. Available:
http://cran.rproject.org/web/packages/.
17. Antonio Piccolboni, "Rhadoop". [Online].
Available:
https://github.com/RevolutionAnalytics/RHadoop/
wiki. [Accessed: Oct. 11, 2014].
18. (Accessed on 2nd October 2014) ff ffbase
[Online] Available:
http://cran.rproject.org/web/packages/ff/index.htm
l
19. (Accessed on 2nd October 2014) bigmemory
[Online] Available:
http://cran.rproject.org/web/packages/bigmemory/
index.html

Memory Management in BigData: A Perpective View

More Related Content

Memory Management in BigData: A Perpective View