http://www.mmds.

org

Introduction to Data Science

Yongfeng Zhang, Rutgers University
¡ Much of the course will be devoted to
large scale computing for data mining
¡ Challenges:
§ How to distribute computation?
§ Distributed/parallel programming is hard

¡ Map-reduce addresses all of the above

§ Google’s computational/data manipulation model
§ Elegant way to work with big data

2
 CPU
Machine Learning, Statistics
Memory

“Classical” Data Mining

Disk

3
¡ 20+ billion web pages x 20KB = 400+ TB
¡ 1 computer reads 30-35 MB/sec from disk
§ ~4 months to read the web
¡ ~1,000 hard drives to store the web
¡ Takes even more to do something useful
with the data!
¡ Today, a standard architecture for such
problems is emerging (data center):
§ Cluster of commodity Linux nodes
§ Commodity network (ethernet) to connect them
4
 2-10 Gbps backbone between racks
1 Gbps between Switch
any pair of nodes
in a rack
Switch Switch

CPU CPU CPU CPU

Mem … Mem Mem … Mem

Disk Disk Disk Disk

Each rack contains 16-64 nodes

In 2016 it was estimated that Google had 2.5M machines, http://bit.ly/Shh0RO

5
6
¡ Large-scale computing for data mining
problems on commodity hardware
¡ Challenges:
§ How do you distribute computation?
§ How can we make it easy to write distributed
programs?
§ Machines fail:
§ One server may stay up 3 years (1,000 days), mostly due
to disk breakdown
§ If you have 1,000 servers, expect to loose 1/day
§ People estimated Google had ~2.5M machines in 2016
§ 2,500 machines fail every day!

7
¡ Issue: Machines may fail due to disk breakdown
¡ Idea: Store files multiple times for reliability
¡ Issue: Copying data over a network takes time
¡ Idea: Bring computation close to the data
¡ Map-reduce addresses these problems
§ Google’s computational/data manipulation model
§ Elegant way to work with big data
§ Storage Infrastructure – File system
§ Google: GFS
§ Programming model
§ Map-Reduce

¡ Some other open-source systems: Apache Hadoop and Spark

8
¡ Problem:
§ If nodes fail, how to store data persistently?
¡ Answer:
§ Distributed File System (DFS):
§ Google GFS; Hadoop HDFS;
¡ Typical usage pattern of DFS
§ Huge files (100s of GB to TB)
§ Data is rarely updated in place (e.g., Google, Amazon)
§ Reads and appends are common

9
¡ Chunk servers
§ File is split into contiguous chunks
§ Typically each chunk is 16-64MB
§ Each chunk is replicated (usually 2x or 3x)
§ Try to keep replicas in different racks
¡ Master node
§ a.k.a. Name Node in Hadoop’s HDFS
§ Stores metadata about where files are stored
§ Might be replicated
¡ Client library for file access
§ Talks to master to find chunk servers
§ Connects directly to chunk servers to access data
10
¡ Reliable distributed file system
¡ Data kept in “chunks” spread across machines
¡ Each chunk replicated on different machines
§ Seamless recovery from disk or machine failure

C0 C1 D0 C1 C2 C5 C0 C5

C5 C2 C5 C3 D0 D1 … D0 C2
Chunk server 1 Chunk server 2 Chunk server 3 Chunk server N

Bring computation directly to the data!

Chunk servers also serve as compute servers
11
Warm-up task:
¡ We have a huge text document

¡ Count the number of times each

distinct word appears in the file

¡ Sample application:
§ Analyze web server logs to find popular URLs

12
Assumption:
§ File too large for memory, but all <word, count> pairs fit in
memory
A naïve method:
¡ Count occurrences of words:
§ words(doc.txt) | sort | uniq -c
§ where words takes a file and outputs the words in it, one per line
¡ The method captures the essence of MapReduce
§ Great thing is that it is naturally parallelizable

13
¡ Sequentially read a lot of data
¡ Map:
§ Extract something you care about
¡ Group by key: Sort and Shuffle
¡ Reduce:
§ Aggregate, summarize, filter or transform
¡ Write the result

Outline stays the same, Map and Reduce

change to fit the problem

14
 Big document
MAP:
Read input and
produces a set of
key-value pairs

Group by key:
Collect all pairs with
same key
(Hash merge, Shuffle,
Sort, Partition)

Reduce:
Collect all values
belonging to the
key and output

15
 Provided by the Provided by the
programmer programmer
MAP: Group by key: Reduce:
Read input and Collect all values
Collect all pairs
produces a set of belonging to the
with same key
key-value pairs key and output

data
The crew of the space

reads
shuttle Endeavor recently
(The, 1) (crew, 1)

read the
returned to Earth as (crew, 1) (crew, 1)
ambassadors, harbingers of (crew, 2)
a new era of space (of, 1) (space, 1)

sequential
exploration. Scientists at
(space, 1)
(the, 1) (the, 1)
NASA are saying that the (the, 3)

Sequentially
recent assembly of the (space, 1) (the, 1)
Dextre bot is the first step in (shuttle, 1)
a long-term space-based (shuttle, 1) (the, 1)
man/mache partnership.
(recently, 1)
(Endeavor, 1) (shuttle, 1)
'"The work we're doing now …

Only
-- the robotics we're doing - (recently, 1) (recently, 1)
- is what we're going to
need ……………………..
…. …
Big document (key, value) (key, value) (key, value)
16
map(key, value):
// key: document name; value: text of the document
for each word w in value:
emit(w, 1)

reduce(key, values):
// key: a word; value: an iterator over counts
result = 0
for each count v in values:
result += v
emit(key, result)

17
 Intermediate
Input key-value pairs
key-value pairs on each chunk server

k v
map
k v
k v
map
k v
k v

… …

k v k v

18
 Output
Intermediate Key-value groups key-value pairs
key-value pairs
reduce
k v k v v v k v
reduce
Group
k v k v v k v
by key

k v

… … …

k v k v k v

19
¡ Input: a set of key-value pairs
¡ Programmer specifies two methods:
§ Map(k, v) ® <k’, v’>*
§ Takes a key-value pair and outputs a set of key-value pairs
§ E.g., key is the filename, value is a chunk of the file
§ There is one Map call for every (k,v) pair
§ System group by k’
§ Reduce(k’, <v’>*) ® <k’, v’’>*
§ All values v’ with same key k’ are reduced together
and processed in v’ order
§ There is one Reduce function call per unique key k’
20
Map-Reduce environment takes care of:
¡ Partitioning the input data
¡ Scheduling the program’s execution across a
set of machines
¡ Performing the group by key step
¡ Handling machine failures
¡ Managing required inter-machine
communication
C0 C1 D0 C1 C2 C5 C0 C5

C5 C2 C5 C3 D0 D1 … D0 C2
Chunk server 1 Chunk server 2 Chunk server 3 Chunk server N 21
 Big document
MAP:
Read input and
produces a set of
key-value pairs

Group by key:
Collect all pairs with
same key
(Hash merge, Shuffle,
Sort, Partition)

Reduce:
Collect all values
belonging to the
key and output

22
All phases are distributed with many tasks doing the work
23
¡ Programmer specifies:
§ Map and Reduce and input files Input 0 Input 1 Input 2

¡ Workflow:
§ Read inputs as a set of key-value-
pairs Map 0 Map 1 Map 2
§ Map transforms input kv-pairs into a
new set of k'v'-pairs
Shuffle
§ Sorts & Shuffles the k'v'-pairs to
output nodes
§ All k’v’-pairs with a given k’ are sent
to the same reduce Reduce 0 Reduce 1

§ Reduce processes all k'v'-pairs

grouped by key into new k''v''-pairs
§ Write the resulting pairs to files Out 0 Out 1

¡ All phases are distributed with

many tasks doing the work
24
¡ Input and final output are stored on a global
distributed file system (DFS):
§ Scheduler tries to schedule map tasks “close” to
physical storage location of input data

¡ Intermediate results are stored on local FS

of Map and Reduce workers
¡ Output is often input to another
MapReduce task

25
¡ Master node takes care of coordination:
§ Task status: (idle, in-progress, completed)
§ Idle tasks get scheduled as workers become
available
§ When a map task completes, it sends the master
the location and sizes of its R intermediate files,
one for each reducer
§ Master pushes this info to reducers

¡ Master pings workers periodically to detect

failures
26
¡ Map worker failure
§ Map tasks completed or in-progress at
worker are reset to idle (because intermediate files are
no longer accessible on map workers)
§ Reschedule the task on another map worker
§ Reduce workers are notified when task is rescheduled
on another worker
¡ Reduce worker failure
§ Only in-progress tasks are reset to idle (because
reducer outputs are stored in global file system)
§ Reduce task is restarted
¡ Master failure
§ MapReduce task is aborted and client is notified
27
¡ M map tasks, R reduce tasks
¡ Rule of a thumb:
§ Make M equal to or larger than the number of
nodes in the cluster
§ One DFS chunk per map is common
§ Improves dynamic load balancing and speeds up
recovery from worker failures
¡ Usually R is smaller than M
§ Because output is summarization of map tasks

28
¡ Problem
§ Slow workers significantly lengthen the job
completion time:
§ Other jobs on the machine
§ Bad disks
§ Weird things
¡ Solution
§ Near end of phase, produce multiple backup copies
of tasks
§ Whichever one finishes first “wins”
¡ Effect
§ Save the effort of rerun a task due to machine failure
29
¡ Often a Map task will produce many pairs of the
form (k,v1), (k,v2), … for the same key k
§ E.g., popular words in the word count example
¡ Can save network time by
pre-aggregating values in
the mapper:
§ combine(k, list(v1)) à v2
§ Combiner is usually same
as the reduce function
¡ Works only if reduce
function is commutative and associative
§ Good example: calculating sum of values
§ Bad example: averaging 1, 2, 2, 4, 4 using three mappers
30
¡ Back to our word counting example:
§ Combiner combines the values of all keys of a
single mapper (single machine):

§ Much less data needs to be copied and shuffled!

31
¡ Want to control how keys get partitioned
§ Inputs to map tasks are created by contiguous
splits of input file
§ Reducer needs to ensure that records with the
same intermediate key end up at the same worker
¡ System uses a default partition function:
§ hash(key) mod R
§ Same key in the same reducer
§ load balancing
¡ Sometimes useful to override the hash
function:
§ E.g., hash(hostname(URL)) mod R ensures URLs
from a host end up in the same output file
32
¡ Suppose we have a large web corpus
¡ Look at the metadata file
§ Lines of the form: (URL, size, date, …)
¡ For each host, find the total number of bytes
§ That is, the sum of the page sizes for all URLs from
that particular host

¡ Other examples:
§ Link analysis and graph processing
§ Machine Learning algorithms

34
¡ Statistical machine translation:
§ Need to count number of times every 3-word
sequence occurs in a large corpus of documents

¡ Very easy with MapReduce:

§ Map:
§ Extract (3-word sequence, count) from document
§ Reduce:
§ Combine the counts

35
¡ Compute the natural join R(A,B) ⋈ S(B,C)
¡ R and S are each stored in files
¡ Tuples are pairs (a,b) or (b,c)

A B B C A C
a1 b1
⋈
b2 c1 a3 c1
a2
a3
b1
b2
b2 c2 = a3 c2
b3 c3 a4 c3
a4 b3
S
R

36
¡ Use a hash function h from B-values to 1...k
§ Hash a B-value into one of the k reduce workers
¡ A Map process turns:
§ Each input tuple R(a,b) into key-value pair (b,(a,R))
§ Each input tuple S(b,c) into (b,(c,S))

¡ Map processes send each key-value pair with

key b to Reduce process h(b)
§ Hadoop does this automatically; just tell it what k is.
¡ Each Reduce process matches all the pairs
(b,(a,R)) with all (b,(c,S)) and outputs (a,b,c).
37
¡ In MapReduce we quantify the cost of an
algorithm using
1. Communication cost = total I/O of all
processes
2. Elapsed communication cost = max of I/O
along any path
3. (Elapsed) computation cost analogous, but
count only running time of processes
Note that here the big-O notation is not the most useful
(adding more machines is always an option)

38
¡ For a map-reduce algorithm:
§ Communication cost = input file size + 2 ´ (sum of
the sizes of all files passed from Map processes to
Reduce processes) + the sum of the output sizes of
the Reduce processes.
§ Suppose 2 reduce workers.

§ Elapsed communication cost is the sum of:

§ the largest input + output for its map process, plus the
same for its reduce process

39
¡ Either the I/O (communication) or processing
(computation) cost dominates
§ Ignore one or the other

¡ Elapsed cost is wall-clock time using

parallelism

¡ Total cost tells what you pay in rent from

your friendly neighborhood cloud

40
¡ Total communication cost
= O(3(|R|+|S|)+|R ⋈ S|)=O(|R|+|S|+|R ⋈ S|)
¡ Elapsed communication cost = O(s)
§ We’re going to pick k (#reducer) and the number of Map
processes so that the I/O limit s is respected
§ We put a limit s on the amount of input or output that any
one process can have. s could be:
§ What fits in main memory
§ What fits on local disk (i.e., make full usage of each machine)
¡ With proper scheduling, computation cost is linear in
the input + output size
§ i.e., No machine idles and waists time

41
¡ Google
§ Not available outside Google
¡ Hadoop
§ An open-source implementation in Java
§ Use HDFS for stable storage
§ Download: http://lucene.apache.org/hadoop/
¡ Aster Data
§ Cluster-optimized SQL Database that also
implements MapReduce

43
¡ Ability to rent computing by the hour
§ Additional services e.g., persistent storage

¡ Amazon’s “Elastic Compute Cloud” (EC2) and

Amazon Web Services (AWS)
¡ Aster Data and Hadoop can both be run on
EC2

44
¡ Jeffrey Dean and Sanjay Ghemawat:
MapReduce: Simplified Data Processing on
Large Clusters
§ http://labs.google.com/papers/mapreduce.html

¡ Sanjay Ghemawat, Howard Gobioff, and Shun-

Tak Leung: The Google File System
§ http://labs.google.com/papers/gfs.html

45
¡ Hadoop Wiki
§ Introduction
§ http://wiki.apache.org/lucene-hadoop/
§ Getting Started
§ http://wiki.apache.org/lucene-
hadoop/GettingStartedWithHadoop
§ Map/Reduce Overview
§ http://wiki.apache.org/lucene-hadoop/HadoopMapReduce
§ http://wiki.apache.org/lucene-
hadoop/HadoopMapRedClasses
§ Eclipse Environment
§ http://wiki.apache.org/lucene-hadoop/EclipseEnvironment
¡ Javadoc
§ http://lucene.apache.org/hadoop/docs/api/
46
¡ Releases from Apache download mirrors
§ http://www.apache.org/dyn/closer.cgi/lucene/had
oop/
¡ Nightly builds of source
§ http://people.apache.org/dist/lucene/hadoop/nig
htly/
¡ Source code from subversion
§ http://lucene.apache.org/hadoop/version_control
.html

47
¡ Programming model inspired by functional language
primitives
¡ Partitioning/shuffling similar to many large-scale sorting
systems
§ NOW-Sort ['97]
¡ Re-execution for fault tolerance
§ BAD-FS ['04] and TACC ['97]
¡ Locality optimization has parallels with Active
Disks/Diamond work
§ Active Disks ['01], Diamond ['04]
¡ Backup tasks similar to Eager Scheduling in Charlotte
system
§ Charlotte ['96]
¡ Dynamic load balancing solves similar problem as River's
distributed queues
§ River ['99]
48