Download the HW3 Skeleton before you begin. [130 points]
Setting up Development Environment for Q1 and Q2
Download a preconfigured virtual machine (VM) image from Cloudera (CDH). The virtual image
comes with pre-installed Hadoop and Spark. You will use them for this HW.
Please use 5.8 version of CDH: QuickStart Downloads for CDH 5.8.
You can choose any VM platform, but we recommend VirtualBox 5.1.14; it is free.
Refer to the Instructions to setup and configure the VM. Read the important tips below about
how to improve the VM’s speed and reduce memory consumption. Once you launch the VM,
you will have a GUI environment with cloudera user, which has administrator (sudo) privilege. The
account details are:
Tip: You may want to set up port forwarding to obtain SSH access of guest operating system.
You can use SSH for copying files from your host operating system to your guest operating
system and vice versa.
Important Tips to speed up your CDH VM
1. Try allocating 50% or lesser memory for VM. Many students have laptops with 4GB
RAM. The default memory configuration of CDH VM is 4GB, which is too large in these
cases. Try reducing the memory size of the VM to 2GB (or even 1GB).
2. Switch on hardware acceleration if possible. This is usually enabled by default.
3. Remove services that are not used. Since you would not be using Mahout, Oozie,
Hive, Sqoop, etc. in the assignment, you can remove those services (and others), to
speed up the VM’s overall speed.
Use the following command to remove each service:
sudo yum remove mahout
sudo yum remove oozie oozie- client
sudo yum remove sqoop
sudo yum remove sqoop2- server sqoop2- client
The entire list of services running can be found at:
Loading Data into HDFS
Now, let’s load our dataset into the HDFS (Hadoop Distributed File System), an abstract file system
that stores files on clusters. Your Hadoop or Spark code will directly access files on HDFS. Paths on
the HDFS look similar to those on the UNIX system, but you can’t explore them directly using
standard UNIX commands. Instead, you need to use hadoop fs commands. For example
hadoop fs -ls /
Download the following two graph files: graph1.tsv (~5MB) and graph2.tsv (~900MB). Use the
following commands to setup a directory on the HDFS to store the two graph datasets. Please do not
change the directory structure below (/user/cse6242/) since we will grade your homework using
the scripts which assume the following directory structure.
sudo su hdfs
hadoop fs -mkdir /user/cse6242/
hadoop fs -chown cloudera /user/cse6242/
hadoop fs -put path/to/graph1.tsv /user/cse6242/graph1.tsv
hadoop fs -put path/to/graph2.tsv /user/cse6242/graph2.tsv
Now both files – graph1.tsv and graph2.tsv are on HDFS at /user/cse6242/graph1.tsv and
/user/cse6242/graph2.tsv. To check this, try:
hadoop fs -ls /user/cse6242
Setting up Development Environments
We found that compiling and running Hadoop/Scala code can be quite complicated. So, we have
prepared some skeleton code, compilation scripts, and execution scripts for you that you can use,
in the HW3 skeleton folder. You should use this structure to submit your homework.
In the directories of both Q1 and Q2, you will find pom.xml, run1.sh, run2.sh and the src
● The src directory contains a main Java/Scala file that you will primarily work on. We have
provided some code to help you get started. Feel free to edit it and add your files in the
directory, but the main class should be Q1 and Q2 accordingly. Your code will be evaluated
using the provided run1.sh and run2.sh file (details below).
Graph derived from the Enron email network dataset, with around 77K nodes and 321K edges.
Graph derived from the Portuguese Wikipedia link dataset, with around 1M nodes and 53M edges.
● pom.xml contains the necessary dependencies and compile configurations for each
question. To compile, you can simply call Maven in the corresponding directory (i.e., Q1 or
Q2 where pom.xml exists) by this command:
It will generate a single JAR file in the target directory (i.e., target/q2-1.0.jar). Again, we
have provided you some necessary configurations to simplify your work for this homework,
but you can edit them as long as our run script works and the code can be compiled using
mvn package command.
● run1.sh, run2.sh are the script files that run your code over graph1.tsv (run1.sh) or
graph2.tsv (run2.sh) and download the output to a local directory. The output files are
named based on its question number and graph number (e.g. q1output1.tsv). You can use
these run scripts to test your code. Note that these scripts will be used in grading.
Here’s what the above scripts do:
1. Run your JAR on Hadoop/Scala specifying the input file on HDFS (the first
argument) and output directory on HDFS (the second argument)
2. Merge outputs from output directory and download to local file system.
3. Remove the output directory on HDFS.
Q1 [25 pts] Analyzing a Graph with Hadoop/Java
a) [15 pts] Writing your first simple Hadoop program
Imagine that your boss gives you a large dataset which contains an entire email communication
network from a popular social network site. The network is organized as a directed graph where each
node represents an email address and the edge between two nodes (e.g., Address A and Address B)
has a weight stating how many times A wrote to B. Your boss wants to find out the people most
frequently contacted by others. Your task is to write a MapReduce program in Java to report the
largest weight among all the weighted inbound edges for each node in the graph.
First, go over the Hadoop word count tutorial to get familiar with Hadoop and some Java basics. You
will be able to complete this question with only some knowledge about Java . You should have
already loaded two graph files into HDFS and loaded into your HDFS file system in your VM. Each file
stores a list of edges as tab-separated-values. Each line represents a single edge consisting of three
columns: (source node ID, target node ID, edge weight), each of which is separated by a tab (\t).
Node IDs are positive integers, and weights are also positive integers. Edges are ordered randomly.
src tgt weight
Some of you may ask “Why should I learn Java?” A main reason is that most (fast) production code is written in
C++ or Java. For enterprise software, Java is used extensively.
117 51 1
194 51 1
299 51 3
230 151 51
194 151 79
51 130 10
Your code should accept two arguments upon running. The first argument (args) will be a path for
the input graph file on HDFS (e.g., /user/cse6242/graph1.tsv), and the second argument (args) will
be a path for output directory on HDFS (e.g., /user/cse6242/q1output1). The default output
mechanism of Hadoop will create multiple files on the output directory such as part-00000,
part-00001, which will be merged and downloaded to a local directory by the supplied run script.
Please use the run scripts for your convenience.
The format of the output should be such that each line represents a node ID and the largest weight
among all its inbound edges. The ID and the largest weight must be separated by a tab (\t). Lines do
not need be sorted. The following example result is computed based on the toy graph above. Please
exclude nodes that do not have incoming edges (e.g., those email addresses that never get contacted
For the toy graph above, the output is as follows.
Test your program on graph1.tsv and graph2.tsv. To demonstrate how your MapReduce procedure
works, use the inline example above, trace the input and output of your map and reduce functions.
That is, given the above graph as the input, describe the input and output of your map and reduce
function(s) and how the functions transform/process the data (provide examples whenever
appropriate). Write down your answers in description.pdf. Your answers should be written in 12pt
font with at least 1” margin on all sides. Your pdf (with answers for both parts a and b) should not
exceed 2 pages. You are welcome to explain your answers using a combination of text and images.
b) [10 pts] Designing a MapReduce algorithm (and thinking in MapReduce)
Design a MapReduce algorithm that accomplishes the following task: given an unordered collection of
two kinds of records, the algorithm will join (combine) record pairs from the collection. In practice, if
you need join a lot of data on Hadoop, you would typically higher level tools like Hive and Pig, instead
of writing code to perform the joins yourself. Learning how to write the code here will help you gain
deeper understanding of the innerworking of the join operations, and will help you decide what kinds
of joins to use under different situations.
We highly recommend that you read this excellent article about joins using MapReduce, which will
give you some hints to correctly complete this question.
NOTE: You only need to submit pseudo code, a brief explanation of your algorithm, and the trace
of input and output of your map and reduce functions for the example given below. No coding is
Input of your algorithm:
Student, Alice, 1234
Student, Bob, 1234
Department, 1123, CSE
Department, 1234, CS
Student, Joe, 1123
The input is an unordered collection of records of two types: Student and Department.
The Student record is of the form <Student, Name, Department_ID>
The Department record is of the form <Department, Department_ID, Department_Name>
Output of your algorithm:
1123, Joe, CSE
1234, Bob, CS
1234, Alice, CS
Your algorithm will join a department record and a student record if they have the same
Department_ID. Each output record is of the form <Department_ID, Name, Department_Name> . For example, Joe’s student record Student, Joe, 1123 should be joined
with Department, 1123, CSE , producing the output record 1123, Joe, CSE .
1. The number of output records should be the same as the number of students. We assume
each student appears exactly once in the input data.
2. The output records should be sorted in ascending order of Department_ID. It is not necessary
to perform any secondary sort (e.g., no need to further sort the records by name).
3. Note that we want student name (e.g., Alice) to always precede department name (e.g., CS) in
each output record. (Does a reduce function typically guarantee its input data be sorted?)
4. You can assume that there is enough RAM to store all the students belonging to a department.
Explain your algorithm by using the above example data and tracing the input and output of your
map and reduce functions. That is, given the above records as the input, describe and explain the
input and output of your map and reduce function(s). Write down your answers in the same
description.pdf. You are welcome to explain your answers using a combination of text and images.
1. [5 pts] Your Maven project directory including Q1.java. Please see detailed submission
guide at the end of this document. You should implement your own MapReduce procedure
and should not import external graph processing library.
2. [2 pts] q1output1.tsv: the output file of processing graph1.tsv by run1.sh.
3. [3 pts] q1output2.tsv: the output file of processing graph2.tsv by run2.sh.
4. [15 pts] description.pdf: Answers for parts a and b. Your answers should be written in 12pt
font with at least 1” margin on all sides. Your pdf must not exceed 2 pages.
Q2 [25 pts] Analyzing a Large Graph with Spark/Scala
Tutorial: Please go over this Spark word count tutorial to get more background about Spark/Scala.
Your task is to calculate the gross accumulated node weights for each node in graph1.tsv and
graph2.tsv from edge weights using Spark and Scala. Assume the graph to be a representation of a
network flow where each edge represents the number of items flowing from source to target. The
gross accumulated node weight for a node is now defined as the number of items
produced/consumed by the node, and can be evaluated using the following formula:
∑(all incoming edge weights) − ∑(all outgoing edge weights)
You should perform this task using the DataFrame API in Spark. Here is a guide that will help you get
started on working with data frames in Spark.
Load your input data into a data frame by inferring the schema using reflection (Refer to the guide
above). Filter out (ignore) all edges whose edge weights equal 1 i.e., only consider edges whose edge
weights do not equal 1. Then use other DataFrame operations to calculate the gross accumulated
node weight for each node.
You may find some of the following DataFrame operations helpful:
toDF, filter, join, select, groupBy
Note: You must use only DataFrame operations for this task. You will lose points if you use SQL
Consider the following example:
src tgt weight
1 2 40
2 3 100
1 3 60
3 4 1
3 1 10
1 -90 = (10) – (40 + 60)
2 -60 = (40) – (100)
3 150 = (100 + 60) – (10)
Notice here that the edge from 3 to 4 is ignored since its weight is 1.
Your Scala program should handle the same two arguments as in Question 1 for input and output
from the console, and should generate the same formatted output file on the supplied output directory
(tab-separated-file). Please note that the default Spark saveastextfile method uses a saving format
that is different from Hadoop’s, so you need to format the result before saving to file (Tip: use map
and mkString). The result doesn’t need to be sorted.
1. [10 pts] Your Maven project directory including Q2.scala. Please see the detailed
submission guide at the end of this document. You may not use any external graph processing
2. [4 pts] q2output1.tsv: the output file of processing graph1 by run1.sh.
3. [6 pts] q2output2.tsv: the output file of processing graph2 by run2.sh.
4. [5 pts] description.txt: describe your approach and refer to line numbers in your code to
explain how you’re performing each step in not more than 150 words.
Q3 [35 pts] Analyzing Large Amount of Data with Pig on AWS
You will try out Apache Pig for processing n-gram data on Amazon Web Services (AWS). This is a
fairly simple task, and in practice you may be able to tackle this using commodity computers (e.g.,
consumer-grade laptops or desktops). However, we would like you to use this exercise to learn and
solve it using distributed computing on Amazon EC2, and gain experience (very helpful for your future
career in research or industry), so you are prepared to tackle the more complex problems.
The services you will primarily be using are Amazon S3 storage, Amazon Elastic Cloud Computing
(EC2) virtual servers in the cloud, and Amazon Elastic MapReduce (EMR) managed Hadoop
For this question, you will only use up a very small fraction of your $100 credit. AWS allows you to use
up to 20 instances in total (that means 1 master instance and up to 19 core instances) without filling
out a “limit request form”. For this assignment, you should not exceed this quota of 20 instances.
Refer to details about instance types, their specs, and pricing. In the future, for larger jobs, you may
want to use AWS’s pricing calculator.
Please read the AWS Setup Guidelines provided to set up your AWS account.
In this question, you will use subsets of the Google books n-grams dataset (full dataset for
reference), on which you will perform some analysis. An ‘n-gram’ is a phrase with n words; the full
n-gram dataset lists n-grams present in the books on books.google.com along with some
You will perform your analysis on two custom datasets, extracted from the Google books bigrams
(2-grams), that we have prepared for you: a small one s3://cse6242-2017spring-bigrams-small
(~1GB) and a large one s3://cse6242-2017spring-bigrams-big (~60GB).
VERY IMPORTANT: Both the datasets are in the US-Standard (US-East) region. Using machines in
other regions for computation would incur data transfer charges. Hence, set your region to US East
(N. Virginia) in the beginning (not Oregon which is the default). This is extremely important otherwise
your code may not work and you may be charged extra.
The files in these two S3 buckets are stored in a tab (\t) separated format. Each line in a file has the
n-gram TAB year TAB occurrences TAB books NEWLINE
An example for 2-grams (or bigram) would be:
I am 1936 342 90
I am 1945 211 25
I am 1951 47 12
very cool 1923 118 10
very cool 1980 320 100
very cool 2012 994 302
very cool 2017 1820 612
This tells us that, in 1936, the bigram ‘I am’ appeared 342 times in 90 different books. In 1945, ‘I am’
appeared 211 times in 25 different books. And so on.
Output the 10 bigrams having the highest average number of appearances per book along with
their corresponding averages, in tab-separated format, sorted in descending order. Only consider
entries with at least 100 occurrences and at least 20 books. If multiple bigrams have the same
average, order them alphabetically. For the example above, the output will be:
I am 4.80
very cool 3.09
Refer to the calculations given below
I am (342 + 211) / (90 + 25) = 4.80
very cool (320 + 994 + 1820) / (100 + 302 + 612) = 3.09
To help you evaluate the correctness of your output, we have uploaded the output for the small
dataset on T-Square.
Using PIG (Read these instructions carefully)
There are two ways to debug PIG on AWS: (all instructions are in the AWS Setup Guidelines)
1. Use the interactive PIG shell provided by EMR to perform this task from the command line
(grunt). Refer to Section 8: Debugging in the AWS Setup Guidelines for a detailed step-by-step
procedure. You should use this method if you are using PIG for the first time as it is easier to
debug your code. However, as you need to have a persistent ssh connection to your cluster
until your task is complete, this is suitable only for the smaller dataset.
2. Upload a PIG script with all the commands which computes and direct the output from the
command line into a separate file. Once you verify the output on the smaller dataset, use this
method for the larger dataset. You don’t have to ssh or stay logged into your account. You can
start your EMR job, and come back after a few hours when the job is complete!
Note: In summary, verify the output for the smaller dataset with Method 1 and submit the results for
the bigger dataset using Method 2.
Sample Commands: Load data in PIG
To load the data from the s3://cse6242-2017spring-bigrams-small bucket into a PIG table, you can
use the following command:
grunt> bigrams = LOAD ‘s3://cse6242-2017spring-bigrams-small/*’ AS
(bigram:chararray, year:int, occurrences:int, books:int);
● Refer to other commands such as LOAD, USING PigStorage, FILTER, GROUP, ORDER BY,
FOREACH.. GENERATE, LIMIT, STORE, etc.
● Copying the above commands directly from the PDF and pasting on console/script file may
lead to script failures due to the stray characters and spaces from the PDF file.
● Your script will fail if your output directory already exists. For instance, if you run a job with the
output folder as s3://cse6242-output, the next job which you run with the same output folder
will fail. Hence, please use a different folder for the output for every run.
● You might also want to change the input data type for occurrences and books to handle
floating point values.
● While working with the interactive shell (or otherwise), you should first test on a small subset of
the data instead of the whole data (the whole data is over hundreds of GB). Once you believe
your PIG commands are working as desired, you can use them on the complete data and
…wait… since it will take some time.
● pig-script.txt: The PIG script for the question (using the larger data set).
● pig-output.txt: Output (tab-separated) (using the larger data set).
Note: Please follow the formatting requirements as we would be using an autograder.
Q4 [35 pts] Analyzing a Large Graph using Hadoop service on Microsoft
Your task is to write a MapReduce program to calculate the degree distribution of a graph. Note that
this question shares some similarities with Question 1 (e.g., both are analyzing graphs). Question 1
can be completed using your own computer. This question is to be completed using Azure. We
recommend that you first complete Question 1.
You will use data files small.tsv (~75MB) and large.tsv (~3GB), for this question. Each file stores a
list of edges as tab-separated-values. Each line represents a single edge consisting of two columns:
(Node A, Node B), each of which is separated by a tab. Node IDs are positive integers and the rows
are already sorted by Node A.
Node A Node B
Your code should accept two arguments upon running. The first argument (args) will be a path for
the input graph file, and the second argument (args) will be a path for output directory. The default
output mechanism of Hadoop will create multiple files on the output directory such as part-00000,
part-00001, which will have to be merged and downloaded to a local directory.
The format of the output should be as follows. Each line represents the degree and its frequency. The
degree and the frequency of the degree must be separated by a tab(\t), and lines don’t have to be
sorted. The following example result is computed based on the toy graph above.
Explanation: Nodes 200, 230 and 299 have a degree of 1 (1 is the degree for 3 nodes). Node 130,
151 and 194 have a degree of 2 (2 is the degree for 3 nodes). Node id 51 is the only node with a
frequency of 3 (3 is the degree for only one node).
subset of LiveJournal data
subset of Friendster data
Hint: One way of doing it is using mapreduce procedure twice. First for finding the degree of each
node and second for calculating the frequency of each degree. You will have to make appropriate
changes in the skeleton code for this.
Setting up Development Environments
To start with, you should create your Azure free trial account using your gatech email id
(email@example.com), this comes up with $200 worth of Azure credits and should be more than enough
for this question. If you run out of credit, you can request us for extra Azure credit (we will announce
how, on Piazza).
Note: Your virtual machine over Azure should only be enabled when you are using Hadoop service. If
you are not running any computation over Azure, you should deallocate (it is different from a stopped
state) your virtual machine using azure management portal, to avoid getting charged. Microsoft’s logic
here is that you are still occupying some resources over cloud, even if you are not using them for
running any computation.
In the Q4 folder of the hw3-skeleton, you will find the following files we have prepared for you:
● src directory contains a main Java file that you will work on. We have provided some code
to help you get started. Feel free to edit it and add your files in the directory, but the main
class should be called “Q4”.
● pom.xml contains necessary dependencies and compile configurations for the question.
To compile, you can run the command:
mvn clean package
in the directory which contains pom.xml.
This command will generate a single JAR file in the target directory (i.e. target/q4-1.0.jar).
Creating Clusters in HDInsight using the Azure portal
Azure HDInsight is an Apache Hadoop distribution. This means that it handles any amount of data
on demand. The next step is to use azure’s web-based management tool to create a Linux cluster.
Follow the documentation here and create a new cluster. At the end of this process, you would
have created and provisioned a New HDInsight Cluster (the provisioning will take some time
depending on how many nodes you chose to create). Please remember to note the cluster login
credentials and the SSH credentials.
Uploading data files to HDFS-compatible Azure Blob storage
Here is the documentation for uploading data files to your Azure Blob storage. We have listed the
main steps from the documentation here:
1. Install Azure CLI.
2. Open a command prompt, bash, or other shell, and use azure login command to authenticate
to your Azure subscription. When prompted, enter the username and password for your
3. azure storage account list command will list the storage accounts for your subscription.
4. azure storage account keys list command should return Primary and
Secondary keys. Copy the Primary key value because it will be used in the next steps.
5. azure storage container list -a -k command will list
your blob containers.
6. azure storage blob upload -a -k
command will upload the source file to your blob storage container.
Using these steps, upload small.tsv and large.tsv to your blob storage container. After that write your
hadoop code locally and convert it to a jar file using the steps mentioned above.
Uploading your Jar file to HDFS-compatible Azure Blob storage
Azure Blob storage is a general-purpose storage solution that integrates with HDInsight. Your
Hadoop code should directly access files on the Azure Blob storage.
Upload the jar file created in the first step to Azure storage using the following command:
scp /q4-1.0.jar USERNAME@CLUSTERNAME-ssh.azurehdinsight.net:
Replace USERNAME with your SSH user name for the cluster. Replace CLUSTERNAME with the
HDInsight cluster name.
SSH into the cluster using the following command:
Run the ls command to make sure that the q4-1.0.jar file is present.
To run your code on the small.tsv file, run the following command:
yarn jar q4-1.0.jar edu.gatech.cse6242.Q4
Command format: yarn jar jarFile packageName.ClassName dataFileLocation outputDirLocation
The output will be located in the
wasbs://@.blob.core.windows.net/smalloutput. If there are multiple
output files, merge the files in this directory using the following command:
hdfs dfs -cat wasbs://@.blob.core.windows.net/smalloutput/* >
Command format: hdfs dfs -cat location/* >outputFile
Download the merged file to the local machine (this can be done either from https://portal.azure.com/
or by using the scp command from the local machine). Here is the scp command for downloading this
output file to your local machine:
scp @-ssh.azurehdinsight.net:/home//small.out .
Using the above command from your local machine will download the small.out file into the current
directory. Repeat this process for large.tsv.
1. [15pt] Q4.java & q4-1.0.jar: Your java code and converted jar file. You should implement your
own map/reduce procedure and should not import external graph processing library.
2. [10pt] small.out: the output file generated after processing small.tsv.
3. [10pt] large.out: the output file generated after processing large.tsv.
Q5 [10 pts] Binary classification: Credit risk prediction, using Azure ML
This question’s main purpose is to introduce you to Microsoft Azure Machine Learning Studio and
familiarize you with its basic functionality and typical machine learning workflow. Go through the
“Binary Classfication: Credit risk prediction” tutorial and complete the tasks below.
1. [3pts] Repeat the experiment mentioned in the tutorial and report the accuracy results as
mentioned in the ‘results’ section of the tutorial.
2. [3pts] Repeat the same experiment, change the ‘Fraction of rows in the first output’ value in
the split module to 0.7 (originally set to 0.5) and report the corresponding results.
3. [4pts] After changing the split value back to 0.5, change a parameter of your choice in the
Boosted Decision Tree algorithm in the given tutorial and report the results along with the
justification for the results in no more than 50 words (i.e., try and explain why the results have
changed). You may want to read the documentation for Boosted Decision Tree.
Optional reading: note that the above tutorial does NOT cover how to use cross validation to
evaluate models. To learn how to perform cross validation in Azure ML studio, please read the
“How to evaluate model performance in Azure Machine Learning“ tutorial.
1. results.txt: a text file containing results for all the three parts.
2. explanation.txt: a text file containing explanation for third part.
Submit the deliverables as a single zip file named HW3-Lastname-Firstname.zip. Specify the
name(s) of any students you have collaborated with on this assignment, using the text box on the
T-Square submission page.
The directory structure of the zip file should be exactly as below (the unzipped file should look like
(do not attach target directory)
(do not attach target directory)
q4-1.0.jar (from target directory)
(do not attach target directory)