Mansoura University

Faculty of Computers and Information

Department of Information System
Second Semester

[AI3502] BIG DATA ANALYTICS

Grade: 3rd AI
Dr. Amira Rezk
Map-Reduce
Map-Reduce

 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
3
SINGLE NODE ARCHITECTURE

CPU
Machine Learning, Statistics
Memory

“Classical” Data Mining

Disk

4
MOTIVATION: GOOGLE EXAMPLE

 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:
 Cluster of commodity Linux nodes
 Commodity network (ethernet) to connect them
5
CLUSTER ARCHITECTURE
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

6

In 2011 it was guestimated that Google had 1M machines, http://bit.ly/Shh0RO

7
LARGE-SCALE COMPUTING

 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)
 If you have 1,000 servers, expect to loose 1/day
 People estimated Google had ~1M machines in 2011
 1,000 machines fail every day! 8
IDEA AND SOLUTION
 Issue: Copying data over a network takes time
 Idea:
 Bring computation close to the data
 Store files multiple times for reliability
 Map-reduce addresses these problems
 Google’s computational/data manipulation model
 Elegant way to work with big data
 Storage Infrastructure – File system
 Google: GFS. Hadoop: HDFS
 Programming model
9
 Map-Reduce
STORAGE INFRASTRUCTURE

 Problem:
 If nodes fail, how to store data persistently?

 Answer:
 Distributed File System:
 Provides global file namespace
 Google GFS; Hadoop HDFS;

 Typical usage pattern

 Huge files (100s of GB to TB)
 Data is rarely updated in place
 Reads and appends are common
10
DISTRIBUTED FILE SYSTEM

 Chunk servers
 File is split into contiguous chunks
 Typically, each chunk is 16-64MB
 Each chunk 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
11
 Connects directly to chunk servers to access data
DISTRIBUTED FILE SYSTEM

 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!

12

Chunk servers also serve as compute servers

PROGRAMMING MODEL: MapReduce

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

13
TASK: WORD COUNT

Case 1:
 File too large for memory, but all <word, count> pairs fit in memory

Case 2:
 Count occurrences of words:
 words(doc.txt) | sort | uniq -c
 where words takes a file and outputs the words in it, one per a line

 Case 2 captures the essence of MapReduce

 Great thing is that it is naturally parallelizable
14
MapReduce: OVERVIEW

 Sequentially read a lot of data

 Map:
 Extract something you care about
Outline stays the same, Map and Reduce
 Group by key: Sort and Shuffle
change to fit the problem
 Reduce:
 Aggregate, summarize, filter or transform
 Write the result

15
MapReduce: THE MAP STEP

Input Intermediate
key-value pairs key-value pairs

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

… …

k v k v
16
MapReduce: THE REDUCE STEP
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

17
MORE SPECIFICALLY

 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 single line in the file

 There is one Map call for every (k,v) pair

 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’ 18
MapReduce: WORD COUNTING
Provided by the Provided by the
programmer programmer
MAP: Group by Reduce:
Read input and key: Collect all values
produces a set of Collect all pairs belonging to the
key-value pairs with same key key and output

data
The crew of the space
(The, 1) (crew, 1)

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

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

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

Sequentially
recent assembly of the
Dextre bot is the first step in (shuttle, 1)
a long-term space-based (shuttle, 1) (the, 1)
(recently, 1)
man/mache partnership. (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 ……………………..
…. …
19
Big document (key, value) (key, value) (key, value)
WORD COUNT USING MapReduce

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)
20
MAP-REDUCE: ENVIRONMENT

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

21
MAP-REDUCE:
A DIAGRAM
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
MAP-REDUCE:
IN PARALLEL

All phases are distributed with many tasks doing the work
 MAP-REDUCE
 Programmer specifies: Input 0 Input 1 Input 2

 Map and Reduce and input files

 Workflow:
Map 0 Map 1 Map 2
 Read inputs as a set of key-value-pairs
 Map transforms input kv-pairs into a new set of k'v'-pairs
 Sorts & Shuffles the k'v'-pairs to output nodes Shuffle

 All k’v’-pairs with a given k’ are sent to the same reduce

 Reduce processes all k'v'-pairs grouped by key into new k''v''- Reduce 0 Reduce 1
pairs
 Write the resulting pairs to files

Out 0 Out 1
 All phases are distributed with many tasks doing the work
DATA FLOW

 Input and final output are stored on a distributed file system (FS):
 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
COORDINATION: MASTER

 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
DEALING WITH FAILURES

 Map worker failure

 Map tasks completed or in-progress at
worker are reset to idle
 Reduce workers are notified when task is rescheduled on another worker

 Reduce worker failure

 Only in-progress tasks are reset to idle
 Reduce task is restarted

 Master failure
 MapReduce task is aborted, and client is notified
27
HOW MANY MAP AND REDUCE JOBS?

 M map tasks, R reduce tasks

 Rule of a thumb:
 Make M much 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 spread across R files
28
TASK GRANULARITY & PIPELINING

 Fine granularity tasks: map tasks >> machines

 Minimizes time for fault recovery
 Can do pipeline shuffling with map execution
 Better dynamic load balancing

29
REFINEMENTS: BACKUP TASKS

 Problem
 Slow workers significantly lengthen the job completion time:
 Other jobs on the machine
 Bad disks
 Weird things

 Solution
 Near end of phase, spawn backup copies of tasks
 Whichever one finishes first “wins”

 Effect
 Dramatically shortens job completion time

30
 REFINEMENT: COMBINERS
 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)) → v2
 Combiner is usually same as the reduce
function
 Works only if reduce function is 31
commutative and associative
REFINEMENT: COMBINERS
 Back to our word counting example:
 Combiner combines the values of all keys of a single mapper (single machine):

32
 Much less data needs to be copied and shuffled!
REFINEMENT: PARTITION FUNCTION

 Want to control how keys get partitioned

 Inputs to map tasks are created by contiguous splits of input file
 Reduce 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

 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 33
PROBLEMS SUITED FOR
MAP-REDUCE
EXAMPLE: HOST SIZE

 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 35
EXAMPLE: LANGUAGE MODEL

 Statistical machine translation:

 Need to count number of times every 5-word sequence occurs in a large corpus of documents

 Very easy with MapReduce:

 Map:
 Extract (5-word sequence, count) from document

 Reduce:
 Combine the counts
36
EXAMPLE: JOIN BY MAP-REDUCE

 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 37
S
R
MAP-REDUCE JOIN

 Use a hash function h from B-values to 1...k

 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). 38
COST MEASURES FOR ALGORITHMS

 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) 39
EXAMPLE: COST MEASURES

 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.
 Elapsed communication cost is the sum of the largest input + output for any
map process, plus the same for any reduce process

40
WHAT COST MEASURES MEAN

 Either the I/O (communication) or processing (computation) cost dominates

 Ignore one or the other

 Total cost tells what you pay in rent from

your friendly neighborhood cloud

 Elapsed cost is wall-clock time using parallelism

41
COST OF MAP-REDUCE JOIN

 Total communication cost

= O(|R|+|S|+|R ⋈ S|)
 Elapsed communication cost = O(s)
 We’re going to pick k 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

 With proper indexes, computation cost is linear in the input + output size
 So computation cost is like comm. cost 42
QUESTIONS