University of Memphis

University of Memphis Digital Commons

Electronic Theses and Dissertations

7-19-2017

Big Data Analysis Using Hadoop and Spark

Follow this and additional works at: https://digitalcommons.memphis.edu/etd

Recommended Citation
Murthy, Adithya K., "Big Data Analysis Using Hadoop and Spark" (2017). Electronic Theses and
Dissertations. 1700.
https://digitalcommons.memphis.edu/etd/1700

This Thesis is brought to you for free and open access by University of Memphis Digital Commons. It has been
accepted for inclusion in Electronic Theses and Dissertations by an authorized administrator of University of
Memphis Digital Commons. For more information, please contact [email protected].
BIG DATA ANALYSIS USING HADOOP AND SPARK

by

Adithya K Murthy

A Thesis

Submitted in Partial Fulfillment of the

Requirements for the Degree of

Master of Science

Major: Computer Science

The University of Memphis

August 2017
Copyright © 2017 Adithya K Murthy
All rights reserved

ii
 Acknowledgment

Foremost, I would like to cordially thank my advisor, Dr. Dipankar Dasgupta, for

his unconditional encouragement and support throughout my master’s career at the

University of Memphis and guiding me through all efforts in this research project. I am

very grateful to him for giving me an opportunity to work with him and his team,

believing in my work and in me, and advising me like a parent. Besides my advisor, I

would like to equally thank my committee members: Dr. Deepak Venugopal and Dr.

Faith Sen for their valuable time, encouragement, and insightful comments.

My sincere thanks and regards to visiting researcher Dr. Mike Nolen for guiding me in

every part of my research work, providing timely advice and helping me arrive at the

discussed results. I would also like to thank all my colleagues from the Center for

Information Assurance and Fogelman College of Business and Economics for their

support and encouragements. I appreciate the direct or indirect help of my professors and

colleagues in the computer science department during my years at the University of

Memphis.

Lastly, I would like to thank my mother Mrs.Shanthi K Murthy and my father Mr.

R Krishna Murthy for all their love and support throughout my life.

iii
 Abstract

Adithya K Murthy. MS. The University of Memphis. August 2017. Big data
analysis using Hadoop and Spark. Major Professor: Dipankar Dasgupta, Ph. D.

Big data analytics is being used more widely every day for a variety of

applications. These new methods of applying analytics certainly bring innovative

improvements in various fields. To process big data and obtain faster, secure and accurate

results is a challenging task. Hadoop and Spark are two technologies which deal with

large amounts of data and process these data in a distributed environment using parallel

computing. Hadoop and Spark use Map Reduce technique to process large data sets. The

iterative processing capability of Hadoop affects the processing speed on the data. Spark

uses in-memory cluster computing/data storage to enhance the performance for different

datasets.

A series of experiments were conducted on both Hadoop and Spark with different

data sets. To analyze the performance variation in both the frameworks, a comparative

analysis was performed from the results obtained by using both Hadoop and Spark. An

experiment based on financial data (NASDAQ Total view-ITCH) was performed in the

Hadoop environment to analyze stock data and its variations.

iv
 Table of Contents

List of Tables .................................................................................................................... vii

List of Figures .................................................................................................................. viii

List of Acronyms and Abbreviations ................................................................................. ix

Introduction ......................................................................................................................... 1

Problem Statement .......................................................................................................... 2

Motivation ....................................................................................................................... 3

Limitation ........................................................................................................................ 3

Roadmap.......................................................................................................................... 3

Hadoop Framework ......................................................................................................... 4

Hadoop Map Reduce ....................................................................................................... 5

Hadoop Distributed File System (HDFS) ....................................................................... 5

Spark ................................................................................................................................... 7

Components of Spark ...................................................................................................... 8

Resilient Distributed Dataset (RDD)............................................................................... 9

Spark Build on Hadoop ................................................................................................. 10

Comparison between Hadoop and Spark .......................................................................... 11

Cluster Configuration ....................................................................................................... 13

Environment setup......................................................................................................... 13

Experiment and Results .................................................................................................... 14

v
 Dataset-1: Wiki Dump Data .......................................................................................... 14

Dataset-2: Web log data ................................................................................................ 16

Dataset-3: Weather Data ............................................................................................... 17

Dataset-4: Matrix Multiplication ................................................................................... 19

Performance Comparison and Analysis ........................................................................ 20

Financial Data Analysis .................................................................................................... 21

NASDAQ data specifications........................................................................................ 21

Experiment .................................................................................................................... 22

Conclusions ....................................................................................................................... 24

References ......................................................................................................................... 25

vi
 List of Tables

Table Page

1. Time Taken to Process Wiki Dump Data on Both Frameworks .................................. 15

2. Time Taken to Process Web Log Data on Both Frameworks ...................................... 17

3. Time Taken to Process Weather Data on Both Frameworks (80 Gb) .......................... 18

4. Time Taken to Process Matrix Multiplication on Both Frameworks ........................... 20

5. Text Format of NASDAQ Trade Data After Conversion ............................................. 22

vii
 List of Figures

Figure Page

1. Phases of Hadoop Map Reduce ...................................................................................... 5

2. Read and Write Operations on HDFS ............................................................................ 6

3. HDFS Architecture ........................................................................................................ 7

4. Difference between Hadoop and Spark Map Reduce Phases ......................................... 8

5. Spark Components .......................................................................................................... 8

6. Spark and Hadoop in the Same Cluster ....................................................................... 10

7. Three ways of Spark Build on Hadoop ........................................................................ 10

8. Work Flow of Spark and Hadoop Frameworks ............................................................ 12

9. Analysis of Wiki Dump Data Results ........................................................................... 15

10. Analysis of Web Log Data Results ............................................................................. 17

11. Analysis of Weather Data Results .............................................................................. 18

12. Performance Comparison of Spark and Hadoop on Different Data Sets.................... 20

13. Cancellation Statistics ................................................................................................. 23

viii
 List of Acronyms and Abbreviations

HDFS Hadoop Distributed File System

RDD Resilient Distributed Datasets

MLib Machine Learning Libraries

YARN Yet Another Resource Negotiator

SIMR Spark in MapReduce

NCDC National Climatic Data Centre

NTFS New Technology File System

ix
 Introduction

Big data is a collection of large datasets that cannot be processed using traditional

computing techniques. An example of big data might be petabytes (1,024 terabytes) or Exabyte

(1,024 petabytes) of data consisting of billions to trillions of records of millions of people from

different sources (e.g., Web, sales, customer contact center, social media, mobile data and so on).

The data is typically loosely structured which is often incomplete and inaccessible. The

importance of big data doesn’t revolve around how much data we have, but what we do with it

and utilize to analyze. It can be useful for various purposes like cost and time reductions,

optimizations and faster decision makings (SAS, 2017).

Many of the organizations across various industries are being benefited by the analysis of

Big data. Big data techniques are used to detect and prevent cyberattacks. Also, they are used to

improve machine or device performances. For example, big data analytics is used to operate

Google's innovative self-driving car (AUGUR, 2015). The CERN (the European Organization

for Nuclear Research) data center has 65,000 processors to analyze its 30 petabytes of data

across more than 150 data centers worldwide for analyzing nuclear data for science and research

purpose (CERN , 2017). Big data analytics is also used in public health industries. For example,

it is helpful to decode DNA and predict various disease patterns. People are also benefited by the

large amounts of data generated from wearable devices like Fitbit or smart watches to analyze

calories and fitness patterns of the human body. Retail companies, advertising companies,

financial institutions are using Big data to analyze customer behavior and improving decision

making models (Marr, 2015).

1
 Big data analysis is a constructive process which involves different stages such as

collection, storage, process, analyze and visualize. Data can be collected from logs, devices and

different sources to be stored in various platforms like servers, data centers etc. The stored data

will be processed and the raw data is converted to utilizable format leading to the resulting data

sets being stored for data analysis and visualization, statistical predictions etc. (AWS, 2017).

Problem Statement

Data processing is a crucial stage in the big data analysis. As we discussed above Big

data cannot be processed using traditional single node computing techniques. We need to scan

terabytes (or even more) of data within seconds. This is only possible when data is processed

with high parallelism. Various tools are available for processing data in a distributed

environment based on parallel computing methods. Hadoop and Spark are framework tools

which can handle data by distributed processing of large data sets across clusters of computers in

parallel.

Hadoop and Spark use Map Reduce techniques to process different data sets. But the

iterative processing in the Hadoop framework has an effect on the data processing speed. Spark

takes MapReduce to the next level with less expensive shuffles in the data processing with

capabilities like in-memory data storage using which the performance can be several times faster

than other big data technologies (Penchikala, 2015).

We try to analyze different data sets on both Hadoop and Spark by comparing the

efficiency of results and performance through which we can analyze which among the two is a

better tool to process data faster.

2
Motivation

Big data analytics is being used on a large scale in recent times. New methods of

applying analytics is a driving factor towards improving business facilities. It is a challenging

task to process big data and get faster, secure and accurate results. Hadoop is one of the solutions

which processes Big Data in a parallel manner. On the contrary, Hadoop has certain

disadvantages such as large I/O overhead since the data which is being processed depends on

external volatile storage. To overcome the disadvantages of Hadoop, Apache Spark was

introduced which promotes in-memory processing. Spark is a faster cluster computing technique

which is designed to operate on top of the Hadoop Distributed File System (HDFS) and it uses

MapReduce model to perform different computations which includes stream processing and

application querying. A thorough study on both the technologies is essential to understand how

data can be processed and analyzed. To arrive at this result, we perform different experiments

with different datasets to analyze Hadoop and Spark frameworks and establish an acute

comparison among their performances.

Limitation

In this study, we have chosen only five different types of data sets. We mainly focused on

batch data and performance in terms of speed and efficiency. We used Hadoop batch and Spark

core to process the data sets, but Spark also provides Spark streaming and M-Lib for real

streaming time data processing and iterative machine learning.

Roadmap

In the next section, we present a brief introduction to Hadoop and Spark components.

Section “Comparison of Hadoop and Spark” illustrates several framework variations in both

Hadoop and Spark. The next two sections “Cluster Configuration” and “Experiments and

3
Results” will list out all the observation we did and findings as well. Eventually, in the last two

sections, the thesis is summarized highlighting future avenues in this line of research.

Hadoop

Apache Hadoop is an open source software project which enables distributed processing

of large structured, semi-structured, and unstructured data sets across different clusters of

commodity servers. (Rouse, 2016) It was inspired by Google's MapReduce, a software

framework in which an application is broken down into numerous small chunks called fragments

or blocks, and can be run on any node in the cluster (Bappalige, 2014). It is designed with a very

high degree of fault tolerance. Hadoop makes it possible to run applications on systems with

multiple nodes involving thousands of terabytes. Hadoop is designed in such a way to process

large volumes of information by connecting to different computers together so that it can work in

parallel.

Hadoop Framework

Apache Hadoop is an open-source software framework for storage and large-scale data

processing on the clusters. Hadoop is designed based on the concept of moving the code to data

instead of data to code, a write-once, read-many, immutability of the written data (Mohammed,

Far & Naugler, 2014). The Hadoop framework consists of the following modules: (Hadoop,

2014).

• Hadoop Common – contains libraries and utilities required by other Hadoop

modules.

• Hadoop Distributed File System (HDFS) – a distributed file system which stores

data on machines, providing bandwidth across the cluster.

4
 • Hadoop YARN – a resource-management platform which is responsible for

managing resources in clusters and using them to schedule users’ applications.

• Hadoop MapReduce – a programming model for large-scale data processing. It

divides applications into smaller components and distributes them across numerous

machines.

Hadoop Map Reduce

MapReduce is a method for distributing a task across multiple nodes. Each node

processes the data stored on that node to the extent possible. A running Map Reduce job consists

of various phases which is described in the following Figure 1:

Map Sort Shuffle Reduce

Figure 1. Phases of Hadoop Map Reduce

The advantages of using MapReduce, which run over a distributed infrastructure like

CPU and storage are Automatic parallelization and distribution of data in blocks across a

distributed system, fault-tolerance against failure of storage, computation and network

infrastructure, deployment, monitoring and security capability and clear abstraction for

programmers. Most MapReduce programs are written in Java. It can also be written in any

scripting language using the Streaming API of Hadoop (Mohammed, Far & Naugler, 2014).

Hadoop Distributed File System (HDFS)

Hadoop is a distributed master–slave architecture that consists of the Hadoop Distributed

File System (HDFS) for storage and the MapReduce programming framework for computational

capabilities. The cluster runs jobs controlled by the master node, which is known as the Name

Node and it is responsible for chunking/partition of the data, cloning it, sending the data to the

distributed computing nodes (Data Nodes), monitoring the cluster status, and

5
collecting/aggregating the results (Mohammed, Far & Naugler, 2014). The following Figure 2

depicts the data node and name node functioning and read, write operation on HDFS.

Figure 2. Read and Write Operations on HDFS (Singh & Ali, 2016)

HDFS is a scalable, fault-tolerant, highly available and access flexible file system. HDFS

splits data into one or more blocks (chunks of 64MB). These blocks are stored in a set of slaves

or data nodes, to ensure that parallel writes or reads can be performed even on a single file.

Multiple copies of each block are stored per replication factor for making the platform fault

tolerant. HDFS stores each file as a sequence of blocks, with each block stored as a separate file

in the local file system (such as NTFS). Multiple copies or replicas of each block makes it fault

tolerant. If one copy is not accessible or gets corrupted, the data can be read from the other copy

(Singh & Ali, 2016). Figure 3 describes the HDFS architecture in detail:

6
 Figure 3. HDFS Architecture (Hadoop, 2008)

Spark

Apache Spark is an open-source, distributed processing system commonly used for big

data workloads. Apache Spark utilizes in-memory caching and optimized execution for faster

performance, and it supports general batch processing, streaming analytics, machine learning,

graph databases, and ad hoc queries (AWS, 2017).

Apache Spark is a faster cluster computing technique which is designed above the

Hadoop Distributed File System (HDFS) and it uses MapReduce model to perform different

computations and includes Stream processing and query applications. The main advantage of

using Spark in relation to Hadoop is that it supports different tools in a single application. The

following Figure 4 depicts the difference between Hadoop and Spark map reduce operations.

7
 Figure 4. Difference between Hadoop and Spark Map Reduce Phases

Components of Spark

Spark provides different libraries which help in analyzing the data in an effective manner.

Figure 5 depicts the different components which are provided by Spark for different

functionalities (Zaharia, Wendell, Konwinski & Karau, 2015).

Figure 5. Spark Components (Apache Spark, 2017)

• Spark SQL- This is an important component which is installed above Spark Core and it

is helpful for creating a special data abstraction called Schema RDD which is helpful in

providing an environment for structured and semi-structured data.

• Spark Streaming- This component enables fast scheduling capability to perform data

streaming analysis. The data would be divided into smaller fragments and it would help

in creating RDD’s and their transformations on these fragments.

8
 • MLlib (Machine Learning Library)- This component is a distributed machine learning

architecture above Spark and it is based on the memory.

• GraphX – This component is a distributed graph-processing framework above Spark.

This component provides an Application Programming interface for computing graph

models and user defined graphs.

Resilient Distributed Dataset (RDD)

Spark depends on the concept of distributed memory abstraction called resilient

distributed dataset (RDD), which is a fault-tolerant in collection of elements that can be operated

on in parallel. RDDs can be created by parallelizing an existing collection and referencing a

dataset in an external storage system, such as a shared filesystem, HDFS (Spark, 2017).

RDD is the basic data structure in Spark. RDDs perform in-memory computations on

large clusters in a fault-tolerant manner. It is immutable and it is a distributed collection of the

objects. Every dataset present in the RDD is separated into logical partitions and they can be

computed on multiple nodes in the cluster. Different objects related to Java, Scala and python

can be present in an RDD along with the objects defined by the users (Zaharia, Chowdhury, Das,

Dave, McCauley, Franklin,Shenker & Stoica, 2012). Figure 6 illustrates how Spark and Hadoop

can coexist on same nodes in a cluster.

9
 Figure 6. Spark and Hadoop in the Same Cluster (Ramel, 2015)

Spark Build on Hadoop

Spark does not come with its own file management system. Spark always requires to be

integrated with other file systems. Spark can be used or installed on Hadoop as it uses the HDFS

provided by Hadoop. This can be performed in three ways (Stoica, 2014) as shown in Figure 7.

Figure 7. Three ways of Spark Build on Hadoop (Stoica, 2014)

10
 1. Standalone mode: Here Spark would be installed on top of the HDFS where storage

would be specifically allocated for HDFS. In this type of architecture both Spark and

MapReduce would run in parallel to execute all the tasks on the servers (clusters).

2. Hadoop Yarn (Yet Another Resource Negotiator): In this type of architecture Spark

would be executed on Hadoop Yarn without any pre-installation. Through this Spark

can be integrated into the Hadoop environment where different components can be

parallely executed on top of the Spark stack.

3. SIMR (Spark in MapReduce): In this mode Spark jobs are launched with the

standalone deployment of Spark. The user can program in the Spark shell and run

multiple tasks.

Comparison between Hadoop and Spark

The following list describes the comparison between Hadoop and Spark technologies

(Noyes, 2015):

1. Hadoop and Spark perform different operations- Hadoop is a distributed data

infrastructure which reorganizes huge amount of data across multiple servers. This

implies that large number of systems need not be available and the hardware need not be

installed in every server. It can keep track of all the data.

Spark operates on the distributed data which is present in the Hadoop HDFS but it is not

possible for Spark alone to store these distributed data. Spark is not available with a

specific file management system and hence it must be integrated with another file system

for storage. But Hadoop has its own distributed file system and execution engine known

as MapReduce for data analysis.

11
2. Spark is comparatively faster than MapReduce technique as Spark has a different

mechanism to process all the data. MapReduce has certain steps which work in an order,

but Spark addresses all the data at once.

• The workflow of Hadoop MapReduce is as follows: Read data from the cluster,

perform an operation, write results to the cluster, read updated data from the

cluster, perform next operation, write next results to the cluster, etc., and

• Spark has a workflow as follows: Read data from the cluster, perform all the

required analytic operations, write results to the cluster, end the process. The

Figure 8 below describes the above-mentioned point pictorially:

Figure 8. Work Flow of Spark and Hadoop Frameworks

3. Fault recovery- In Hadoop there is a replication factor of 3 for all the data which is stored

on the HDFS. Hence 3 copies of the entire dataset would be stored on multiple nodes.

Spark has RDD’s which perform the function of storing of all the data in memory and

hence it is a better fault tolerance technique in comparison with Hadoop.

4. Hadoop is less fault tolerant to failures as the data is written onto the disk after each

operation. Spark has inbuilt data objects which are stored in RDD’s across the cluster.

12
 RDD’s are stored in memory and hence complete recovery from faults and failures is

available.

Cluster Configuration

We evaluated the performance of Spark and Hadoop with a series of experiments by

setting up a cluster environment to test different data sets. We tried to experiment how much

faster Spark provides the results in comparison with Hadoop on the cluster that we have setup.

Environment setup

The following details explain the cluster configuration we considered for the experiments.

• Brand: Dell

• Product Line: PowerEdge 2950

• Product Type: Rackmount Server

• Form Factor: 2U Rack

• I/O Slots: 6 PCI slots

• Processors: Up to 2x Dual-Core Intel Xeon 5000 series

• Processor Cache: 2x2MB L2 per processor

• Hard Drives: Up to 6x 3.5" SAS/SATA Hard Drives

• Hard Disk drive size of each rack: Name node: 2*146Gb

Data node 1: 2*146Gb

Data node 2: 2*146Gb

Data node 3: 4*146Gb

• RAM installed in GB- (16GB)

• CPU type and speed GHZ –Two 1.86GHz Quad-core processors.

▪ RAID Controller model no: - Perc 5/i Raid 0 (used)

13
 ▪ OS- CentOS release 6.7(Final)

• Hadoop – Cloudera open source ecosystem

• Spark – Apache Spark software foundation

Experiment and Results

Dataset-1: Wiki Dump Data

This dataset consists of a complete copy of wiki database backup in the form of wiki text

source and metadata embedded in XML. It consists the complete dump data of Wikipedia pages

for research purpose. This dump contains multiple formats and our dump file has XML format

which is a semi-structured data. The dump file contains Page title, Revision dates, Timestamp

page Authors and Content. Size of this dump data is about 119 Gigabyte (Uncompressed XML

and text) from 2002 to 2016 database backup dumps. We tried to analyze different factors from

the data such as:

• Total page count

• No of redirected pages (Pages that share same subject)

• Username (Author) count

• Number of revisions made by a user on a page

We tried to test this data on both Hadoop and Spark cluster. We implemented a Map

Reduce application to process the data and obtain desired outputs.

14
 Figure 9. Analysis of Wiki Dump Data Results

The output from the data consists of 39260948 no. of pages, 8896619 redirected titles and

38306114 no. of users as extracted from the Wiki dump data. Figure 9 (graph) describes the

number of pages which were revised by users, which increased on a monthly basis for the years

from 2002 to 2016. We also analyzed the time taken to process this data on both Hadoop and

Spark clusters. The points which are highlighted in red indicate the total number of authors who

are increasingly involved in updating specific pages. This analysis is verified with the output file

by referring to the total number of authors for every particular month (Figure 9) The following

table describes the time variations on both the frameworks. Table 1 describes the time taken to

process the data on both Spark and Hadoop.

Table 1. Time Taken to Process Wiki Dump Data on Both Frameworks

Wiki Dump Data

Spark 9.2 min

Hadoop 19.4 min

15
Dataset-2: Web log data

Analysis of logs have a significant importance in different fields. As an example,

analyzing web log files has become an important task for E-Commerce companies to predict

their customer behavior and to improve their business. For every task in the web operations, logs

are generated for every second. This continuous streaming of logs results in huge amounts of

data. For the experiment, we consider a sample web log data to process in both Spark and

Hadoop environment. The log format that we have chosen is as shown below:

127.0.0.1 - - [10/Oct/2000:13:55:36 -0400] "GET /apache_pb.gif HTTP/1.0" 200 2326

"http://www.example.com/start.html" "Mozilla/4.08 [en] (Win98; I; Nav)"

• 127.0.0.1 - This is the IP address of the client which made the request to the server.

• [10/Oct/2000:13:55:36 -0700] -The time format resembles like [day/month/year:

hour: minute: second zone]

• Get/POST - The request sent from the client. The method used for information

transfer is noted.

• "http://www.example.com/start.html" - This gives the site that the client reports

having been referred from.

• "Mozilla/4.08 [en] (Win98; I; Nav)" - This is the information that the client browser

reports about itself to the server.

We collected 115gb of sample Web log data and implemented a Map Reducer application

to process data on both Spark and Hadoop which calculates the total number of hits and IP

addresses on the website. We extracted the desired data from Hadoop and Spark and analyzed by

plotting a graph. Figure 10 depicts the results of the Web log data process on cluster.

16
 HITS

30

28
27
NO OF HITS

23

23

23
22

22

22
21

20
19
18

17

17

14
13

12

11
9
Hits

8
7

2
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
HOURS OF DAY

Figure 10. Analysis of Web Log Data Results

We also analyzed the time taken to process this data on Hadoop and Spark clusters. The

following table describes the time variations on both frameworks. Table 2 describes the time

taken to process the data on both Spark and Hadoop.

Table 2. Time Taken to Process Web Log Data on Both Frameworks

Web Log Data

Spark 12.3 min

Hadoop 27.5 min

Dataset-3: Weather Data

NCDC (National climatic Data Centre) provides access to daily data from the U.S.

Climate Reference Network / U.S. Regional Climate Reference Network (USCRN/USRCRN).

NCDC has weather data collected from weather sensors all the way back from 1901. We have

considered a sample data for one city (Crossville) from NCDC for experimenting with weather

analysis using Map-Reduce. We have implemented an application which analyzes the weather

17
data for a particular city and the result is the number of days which are Hot or Cold for the city.

The format of the NCDC data as shown below:

0067011990999992015051507004+68750+023550FM12+038299999V0203301N00671220001CN9999999N9+00001+99999.

0043012650999992016032418004+62300+010750FM-12+048599999V0202701N00461220001CN0500001N9-00785+99999.

We collected approximately one year of Crossville city data from the NCDC and

processed on Hadoop and Spark. We Implemented an application which can run data and extract

cold days and hot days in a year. Figure 11 depicts the analysis of the results from the Weather

data process on cluster.

Crossville Weather Data 2015

20

10

-10

Days

Cold-Day Hot-Day

Figure 11. Analysis of Weather Data Results

We also analyzed the time taken to process this data on Hadoop and Spark clusters. The

following table describes the time variations in both the frameworks. Table 3 describes the time

taken to process the data on both Spark and Hadoop.

Table 3. Time Taken to Process Weather Data on Both Frameworks (80 Gb)

Weather Data

Spark 7.2 min

Hadoop 21.4 min

18
Dataset-4: Matrix Multiplication

Some of the wider applications of Matrix multiplication such as running and processing

large matrices is difficult with single processors. Using distributed system and parallel

computing can enhance the speed of the execution of the computations. For example, in the

research of cancer tissues, to compare and analyze the abnormal cell matrix multiplication is

very important and size of the matrices is very large. In these scenarios Spark and Hadoop play a

vital role to get faster and efficient results.

We consider a 100*100 matrix to test the datasets for multiplication operation. Here we

show a small 2*5 and 5*2 matrix and their multiplication which is feasible. The input file format

is as shown below (R-Row, C-Column, V-Value):

R, C, V Output R

0,0,0.0

0,1,1.0

0,2,2.0 (5, [0,1,2,3,4], [45.0,58.0,71.0,84.0,97.0])

0,3,3.0 (5, [0,1,2,3,4], [25.0,30.0,35.0,40.0,45.0])

0,4,4.0 (5, [0,1,2,3,4], [30.0,37.0,44.0,51.0,58.0])

1,0,5.0 (5, [0,1,2,3,4], [40.0,51.0,62.0,73.0,84.0])

1,1,6.0 (5, [0,1,2,3,4], [35.0,44.0,53.0,62.0,71.0])

1,2,7.0

1,3,8.0

1,4,9.0

19
 We also analyzed the time taken to process this data on both Hadoop and Spark

environments. The following table describes the time variations for both the frameworks. Table 4

describes the time taken to process the data on both Spark and Hadoop.

Table 4. Time Taken to Process Matrix Multiplication on Both Frameworks

Matrix Multiplication

Spark 8.1 min

Hadoop 19 min

Performance Comparison and Analysis

We compared the results using different data sets mentioned above and concluded that

Spark produces around 45% faster results than Hadoop. Figure 12 depicts the performance

comparison of Spark and Hadoop framework.

Performance Comparision
Hadoop Spark

30
27.5
25
20 21.4
19 19.4
15
12.3
10 9.2
8.1 7.2
5
0
Matrix Multiplication Web-Log data Weather Data Wiki Dump

Figure 12. Performance Comparison of Spark and Hadoop on Different Data Sets

From the above experiments, for the considered datasets, we analyzed that Spark produces

specific results approximately 2.5 times faster than Hadoop on the configured cluster

environment.

20
 Financial Data Analysis

In the cluster which is set up we used Hadoop as a distributed file system that runs on a

combination of virtual nodes, and is capable of handling large data sets (order of terabytes to

petabytes) with increased number of clusters. Several data analytic tools can be used on top of

Hadoop for analyzing large data sets in parallel. For example, one day’s worth NASDAQ trade

data (Data Feed Total View-ITCH) are of the order 320,000,000 million records requiring 10 GB

(Gigabyte) of space (NASDAQ, 2015). The analysis of these data is possible on a single server

but becomes time consuming when multiple days or weeks’ worth of data needs to be

processed. For data analysis, we have installed Hadoop and software tools provided by Spark to

analyze ITCH data. First challenge was to understand the data set, which is not in a human

readable format rather it is in binary form. Second, there are many record formats and ITCH

documentation provides 18 different record types. In accordance to this we developed our own

ITCH converter which reads the raw binary data and converts it to a common delimited text file

to facilitate processing with Hadoop/Spark. We have recently started working on extract tools,

used to select specific records and data, which would be required to perform analysis in Spark.

NASDAQ data specifications

• The Total View-ITCH data consists of Nano-second level data which contains market

messages for different limit orders such as:

– submissions, cancellations, executions and revisions.

– hidden order executions and revisions.

– other stock and order relevant information for every trading day on NASDAQ.

• Types of variables included in these messages are:

– order reference number

21
 – ticker

– time in nanoseconds (for every order submission, revision, cancellation,

execution)

– number of shares (submitted, executed, deleted)

– buy/sell information

– price (submission, execution)

– market participant ID

• The data is obtained as a zipped .txt files coded in binary which requires a special code to

process the file) and one unzipped file can be around 2-10 Gigabytes(Compressed) and

once it is uncompressed the dataset takes around 20-60Gigabyte of memory.

Normal systems take up to 5-8 hours to open a single file. The distributed cluster

architecture processes the data and can open a file in about 20 minutes. Table 5 depicts the

sample text format of the NASDAQ trade data after conversion from raw data

Table 5. Text Format of NASDAQ Trade Data After Conversion

Message Timestamp (seconds Order Buy/Sell Shares Stocks Price Original Order Match Number Market
Type past midnight) Reference Side Reference Number Participant ID
Number
Add 32435.285091234 335531633 S 300 AMZN 450.5
Add w/ ID 33024.027480234 168914198 B 100 C 48 UBSS
Revision 34011.528407249 336529765 300 450.45 335531633
Exe 35011.180402345 336529765 76 3893405
Exe at diff P 36001.712309875 168914198 100 48 5830295
Part. Cancl 37134.713037569 336529765 100
Full Cancl 38104.371047195 336529765

Experiment

To understand the contents of the data and its functionalities we performed an experiment

which involves the analysis of ITCH data. This experiment was a comparison to the results

obtained by the finance department. The size of the data was about 14GB and was recorded on

23 January 2015 by NASDAQ (after conversion). The data was provided by finance department

22
at the University of Memphis for research purposes. It was processed on Spark framework which

used HDFS and took about 6 min and was considered for a single day transaction. Sample

converted data is stored on the HDFS. On this data set different operations are performed such as

record type counter, hourly addition/deletion of the stocks to find the cancellation stats. In

accordance to this the total number of placed orders (A), total number of executions (E), the total

number of cancelled orders (X) and the total number of deleted fields (D) are considered to

arrive at a result for Placement vs Execution vs Cancellation. The resulting file consisted of

327,000,000 records. We process the data on Hadoop framework and implemented an

application to analyze the hourly history for each type of record. Figure 13 depicts the graph

plotted with time on the x-axis and the total number of messages on the y-axis to compute the

hourly history for each type of the record.

Figure 13. Cancellation Statistics

23
 Conclusions

Spark and Hadoop framework are two popular distributed computing paradigms for

providing an effective solution to handle large amount of data. An in-depth analysis of Hadoop

and Spark frameworks is performed to understand the performance differences and efficient

processing of Big Data. The experimental evaluation confirms that Spark is more efficient and

faster than Hadoop when handling different data in different scales. We expanded our research to

process financial data from NASDAQ. (Total view-ITCH data).

24
 References

AUGUR, H. (2015). How Data Science is Driving The Driverless Car. Retrieved june 15,

2017, from http://dataconomy.com/2015/12/how-data-science-is-driving-the-

driverless-car/

AWS. (2017). Amazon Web Services. Retrieved June 1, 2017, from What is Big Data?:

https://aws.amazon.com/big-data/what-is-big-data/

AWS, A. (2017). Apache Spark on Amazon EMR. Retrieved June 20, 2017, from

https://aws.amazon.com/emr/details/spark/

Bappalige, S. P. (2014). An introduction to Apache Hadoop for big data. Retrieved June

5, 2017, from Opensource: https://opensource.com/life/14/8/intro-apache-hadoop-

big-data

CERN . (2017). Retrieved june 20, 2017, from http://home.cern/about/computing

Emad A Mohammed, B. H. (2014). Applications of the MapReduce programming

framework to clinical big data analysis: current landscape and future trends.

BioMed Central.

Hadoop, A. (2014). What Is Apache Hadoop? Retrieved June 10, 2017, from Apache

Hadoop: http://hadoop.apache.org

Manpreet Singh, A. A. (2016). Big Data Analytics with Microsoft HDInsight (1st ed.).

Sams Teach.

Marr, B. (2015). The Awesome Ways Big Data Is Used Today To Change Our World.

Retrieved May 15, 2017, from

25
 http://www.dataplumbing.datasciencecentral.com/blog/the-awesome-ways-big-

data-is-used-today-to-change-our-world

Matei Zaharia, M. C. (2012). Resilient Distributed Datasets: A Fault-Tolerant Abstraction

for In-Memory Cluster Computing. Presented as part of the 9th USENIX

Symposium on Networked Systems Design and Implementation (NSDI 12.

Matei Zaharia, P. W. (2015). Learning Spark. O'Reilly Media, Inc.

Penchikala, S. (2015). Big Data Processing with Apache Spark – Part 1: Introduction.

Retrieved May 20, 2017, from InfoQ: https://www.infoq.com/articles/apache-

spark-introduction

Rouse, M. (2016). Hadoop. Retrieved May 20, 2017, from TechTarget:

http://searchcloudcomputing.techtarget.com/definition/Hadoop

SAS. (2017). Big Data. Retrieved May 13, 2017, from Big Data Insights:

https://www.sas.com/en_us/insights/big-data/what-is-big-data.html

Spark. (2017). Spark. Retrieved June 2017, 2017 , from Spark Programming Guide:

https://spark.apache.org/docs/latest/programming-guide.html

Spark, A. (2017). Apache Spark™ is a fast and general engine for large-scale data

processing. Retrieved June 10, 2017, from https://spark.apache.org

Stoica, I. (2014). Apache Spark and Hadoop: Working Together. Retrieved June 15,

2017, from Databricks: https://databricks.com/blog/2014/01/21/spark-and-

hadoop.html

26