Topic 6: MapReduce Applications

6: MapReduce Applications

Zubair Nabi

zubair.nabi@itu.edu.pk

April 18, 2013

Zubair Nabi 6: MapReduce Applications April 18, 2013 1 / 27

Outline

1 The Anatomy of a MapReduce Application

2 MapReduce Design Patterns

3 Common MapReduce Application Types


Outline





MapReduce job phases

A MapReduce job can be divided into 4 phases:
1 Input split: The input dataset is sliced into M splits, one per map task



2 Map logic: The user-supplied map function is invoked



In tandem a sort phase is also applied that ensures that map output is
locally sorted by key



In addition, the key space is also partitioned amongst the reducers



3 Shufﬂe: Map output is relayed to all reduce tasks



4 Reduce logic: The user-provided reduce function is invoked



4 Reduce logic: The user-provided reduce function is invoked
Before the application of the reduce function, the input keys are merged
to get globally sorted key/value pairs


Of mappers and reducers

In the common case, programmers only need to write a map and a
reduce function



reduce function
The user-provided map function is invoked for every line (can be
modiﬁed) in the input ﬁle and is passed the line number as key and line
contents as value



reduce function
The user-provided map function is invoked for every line (can be
modiﬁed) in the input ﬁle and is passed the line number as key and line
contents as value
The user-provided reduce function is invoked for each key output by
the map phase and is passed the set of associated values as iterable
values


Wordcount: High-level view

Input: A text corpus such as Wikipedia dump, books from Gutenberg,
etc.



etc.
The map function is invoked once for each text line



etc.
Map output: Words as keys and 1 as values



etc.
Reduce input: Key/value pairs of words and values (1)



etc.
The reduce function is invoked once for each word with a list of 1s



etc.
The reduce function is invoked once for each word with a list of 1s
Reduce output: Words and their ﬁnal counts


Wordcount: Low-level view

A new process is created for each map, called MapRunner



MapRunner has a RecordReader instance that is used to read the
input ﬁle



input ﬁle
RecordReader reads the input ﬁle in chunks and parses the chunks
into lines



input ﬁle
into lines
MapRunner also has a Mapper instance with a map function,
WordCountMapper in this case



input ﬁle
into lines
MapRunner also has a Mapper instance with a map function,
WordCountMapper in this case
For each line parse by RecordReader, MapRunner calls
WordCountMapper.map() and passes it the line


Wordcount: Low-level view (2)

WordCountMapper has an OutputCollector instance which maintains
an in-memory buffer for each output partition (one partition per reduce)



Each time WordCountMapper.map() is invoked it, it tokenizes the line
into words



into words
For each word, it writes the word as key and 1 as value to
OutputCollector



into words
OutputCollector
OutputCollector uses the Partitioner instance to select a partition
buffer for each key



into words
OutputCollector
buffer for each key
Whenever the size of a partition buffer exceeds a configurable
threshold, its contents are first sorted by key and then flushed to disk



into words
OutputCollector
buffer for each key
Whenever the size of a partition buffer exceeds a configurable
threshold, its contents are first sorted by key and then flushed to disk
This process is repeated till the map logic has been applied to all lines
within the input file



Once all maps have completed their execution, the reduce phase is
started



started
For each reduce task, a ReduceRunner process is created



started
Each reduce task fetches its input partitions from machines on which
map tasks were run



started
Each reduce task fetches its input partitions from machines on which
map tasks were run
All input partitions are then merged to get a globally sorted partition of
key/value pairs



ReduceRunner contains a Reducer instance with a reduce function,
WordCountReducer in this case



For each word, ReduceRunner invokes WordCountReducer.reduce()
and passes it the word and a list of its values (1s)



WordCountReducer also has an OutputCollector instance with an
in-memory buffer



in-memory buffer
WordCountReducer.reduce() sums the list of values it is passed and
writes the word and its ﬁnal count to the OutputCollector



in-memory buffer
This process is repeated till the reduce logic has been applied
key/value pairs



in-memory buffer
This process is repeated till the reduce logic has been applied
key/value pairs
At the end of the entire job, each reduce produces an output ﬁle with
words and their number of occurrences


Wordcount map in Java

1 public void map( Object key , Text value , Context context ) {
2 StringTokenizer itr = new StringTokenizer ( value . toString ());
3 while (itr. hasMoreTokens ()) {
4 word.set(itr. nextToken ());
5 context .write (word , one );
6 }
7 }


Wordcount reduce in Java

1 public void reduce (Text key , Iterable < IntWritable > values ,
2 Context context ) {
3 int sum = 0;
4 for ( IntWritable val : values ) {
5 sum += val.get ();
6 }
7 result .set(sum );
8 context .write(key , result );
9 }


Wordcount map in Python

1 def map(self , key , value ):
2 [self. _output_collector . collect (word , 1) for word in value . split (’ ’)]


Wordcount reduce in Python

1 def reduce (self , key , values ):
2 sum__ = 0
3 for value in values :
4 sum__ += value
5 self. _output_collector . collect (key , sum__ )


Outline





Bird’s-eye view

The MapReduce paradigm is amenable to divide-and-conquer
algorithms


Bird’s-eye view

algorithms
One way to look at MapReduce is that it is just a large-scale sorting
platform


Bird’s-eye view

algorithms
platform
User-logic is only involved at speciﬁc hook points


Bird’s-eye view

algorithms
platform
Algorithms must be expressed in terms of a small number of speciﬁc
components that ﬁt together in preset ways


Bird’s-eye view

algorithms
platform
Like putting together a jigsaw puzzle in which all the other pieces have
already been assembled and you only need to add two pieces: The map
and the reduce pieces


Bird’s-eye view

algorithms
platform
Like putting together a jigsaw puzzle in which all the other pieces have
already been assembled and you only need to add two pieces: The map
and the reduce pieces
Fortunately a large number of algorithms easily ﬁt this rigid pattern


Programmer control

The programmer has no control over
1 The location of a map or reduce task in terms of nodes in the cluster


Programmer control

2 The start and end time of a map or a reduce task


Programmer control

3 The input key/value pairs processed by a speciﬁc map task


Programmer control

3 The input key/value pairs processed by a speciﬁc map task
4 The intermediate key/value pairs processed by a speciﬁc reduce task


Programmer control (2)

The programmer does have control over
1 The data structures to be used as keys and values



2 Initialization code at the beginning of map/reduce tasks and
termination code at the end



3 Preservation of state across multiple invocations of map/reduce tasks



4 The sort order of intermediate keys and in turn, the order in which a
reducer encounters keys



4 The sort order of intermediate keys and in turn, the order in which a
reducer encounters keys
5 Partitioning of key space and in turn, the set of keys that a particular
reducer encounters


Multi-job algorithms

Many algorithms cannot be easily expressed as a single MapReduce
job



job
Complex algorithms need to be decomposed into a sequence of jobs



job
The output of one job becomes the input to the next



job
The output of one job becomes the input to the next
Most interactive algorithms need to be run by an external driver
program that performs the convergence check


Local aggregation

Network and disk latencies are expensive compared to other
operations


Local aggregation

operations
Decreasing the amount of data transferred over the network during the
shufﬂe phase results in efﬁciency


Local aggregation

operations
Aggressive user of combiners for commutative and associative
algorithms can greatly reduce intermediate data


Local aggregation

operations
Aggressive user of combiners for commutative and associative
algorithms can greatly reduce intermediate data
Another strategy, dubbed “in-mapper combining” can not only decrease
the amount of intermediate data but also the number of key/valur pairs
emitted by the map tasks


Outline





Counting and Summing

1 Problem
A number of documents with a set of terms



1 Problem
Need to calculate the number of occurrences of each term (word count)
or some arbitrary function over the terms (average response time in log
ﬁles)



1 Problem
ﬁles)
2 Solution
Map: For each term, emit the term and “1”



1 Problem
ﬁles)
2 Solution
Map: For each term, emit the term and “1”
Reduce: Take the sum (or any other operation) of each term values


Collating

1 Problem
A number of documents with a set of terms and some function of one
item


Collating

1 Problem
item
Need to group all items that have the same value of function to either
store items together or perform some computation over them


Collating

1 Problem
item
2 Solution
Map: For each item, compute given function and emit function value as
key and item as value


Collating

1 Problem
item
2 Solution
Reduce: Either save all grouped items or perform further computation


Collating

1 Problem
item
2 Solution
Reduce: Either save all grouped items or perform further computation
Example: Inverted Index: Items are words and function is document ID


Filtering, Parsing, and Validation

1 Problem
A set of records



1 Problem
A set of records
Need to collect all records that meet some condition or transform each
record into another representation



1 Problem
A set of records
2 Solution
Map: For each record, emit it if passes the condition or emit its
transformed version



1 Problem
A set of records
2 Solution
transformed version
Reduce: Identity



1 Problem
A set of records
2 Solution
transformed version
Reduce: Identity
Example: Text parsing or transformation such as word capitalization


Distributed Task Execution

1 Problem
Large computational problem



1 Problem
Need to divide it into multiple parts and combine results from all parts to
obtain a ﬁnal result



1 Problem
2 Solution
Map: Perform corresponding computation



1 Problem
2 Solution
Reduce: Combine all emitted results into a ﬁnal one



1 Problem
2 Solution
Reduce: Combine all emitted results into a ﬁnal one
Example: RGB histogram calculation of bitmap images


Sorting

1 Problem
A set of records


Sorting

1 Problem
A set of records
Need to sort records in some order


Sorting

1 Problem
A set of records
2 Solution
Map: Identity


Sorting

1 Problem
A set of records
2 Solution
Map: Identity
Reduce: Identity


Sorting

1 Problem
A set of records
2 Solution
Map: Identity
Reduce: Identity
Also possible to sort by value, either perform a secondary sort or
perform a key-to-value conversion


References

1 Jimmy Lin and Chris Dyer. 2010. Data-Intensive Text Processing with
MapReduce. Morgan and Claypool Publishers.
2 MapReduce Patterns, Algorithms, and Use Cases:
http://highlyscalable.wordpress.com/2012/02/01/
mapreduce-patterns/


Topic 6: MapReduce Applications

More Related Content

Similar to Topic 6: MapReduce Applications (20)

More from Zubair Nabi (20)

Recently uploaded (20)

Topic 6: MapReduce Applications