MapReduce

Zuhair Khayyat3/11/2012

Introduction to

MapReduce

CS245 - 2012 Introduction to MapReduce 2

What is MapReduce

● A programming model introduced by Google in OSDI '04 for processing large datasets efficiently.

● Features:

– Automatic parallelization, no parallel experience required.

– Data and process redundancy for failure recovery.

– Auto scheduling and Load balancing.

– Easy to program, based on two simple functions:

● Map● Reduce.


Why MapReduce?

● For a cluster of:

– 2000 machines.

– Total 16 TB Ram (≈ 8 GB each).

– Total 2 PB Disk space (≈ 1 TB each).● Use the maximum capacity of the cluster to:

– Implement a parallel word count for input size 100 TB.


Why MapReduce?

● For a cluster of:

– 2000 machines.

– Total 16 TB Ram (≈ 8 GB each).

– Total 2 PB Disk space (≈ 1 TB each).● Use the maximum capacity of the cluster to:

– Implement a parallel word count for input size 100 TB.

– Implement a parallel sort for the same input file.● Can you use the same code for both applications?


How Fast is MapReduce (Hadoop)

● Sort Benchmark competition (http://sortbenchmark.org/):

– 2009: 100 TB in 173 minutes using 3452 nodes:● 2 x Quad core Xeons @ 2.5 GHz.● 8 GB RAM.

– 2008: 1TB in 3.48 minutes using 910 nodes:● 4 x Dual core Xeons @ 2.0 GHz.● 8 GB RAM.

http://sortbenchmark.org/


Who uses MapReduce?


Map & Reduce functions

● The Mapper (Pick a key):

– Input: Read input from disk.

– Output: Create pairs of <key, value>, known as intermediate pairs.

– More input partitions == More parallel Mappers.● The Reducer (Process values):

– Input: a list of <key,value> pairs with a unique key.

– Output: Single or multiple of <key, values>

– More unique keys == More Parallel Reducers.


How MapReduce Work

1) Partition input file into M partitions.

2) Create M Map tasks, read M partitions in parallel and emits intermediate <key, value> pairs. Store them into local storage.

3) Wait for all Map workers to finish, sort and partition intermediate <key, value> pairs into R regions.

4) Start R reduce workers, each reads a list of intermediate with a unique key from remote disks.

5) Write the output of reduce workers to file(s).


Example – Word count

● Assume an input as following:

cat flower picturesnow cat cat

prince flower sunking queen AC



● Step1: Partition input file into M partitions.

cat flower picturesnow cat cat

prince flower sunking queen AC

cat flower picture

snow cat cat

prince flower sun

king queen AC



● Step2: Create M Map tasks, read M partitions in parallel and emits intermediate <key, value> pairs. Store them into local storage.

cat flower picture

snow cat cat

prince flower sun

king queen AC

Mapper 1

Mapper 2

Mapper 3

Mapper 4

<cat,1> <flower,1> <picture,1>

<snow,1> <cat,1> <cat,1>

<prince,1> <flower,1> <sun,1>

<king,1> <queen,1> <AC,1>



● Step3: Wait for all Map workers to finish, sort and partition intermediate <key, value> pairs into R regions.

<cat,1> <flower,1> <picture,1>

<snow,1> <cat,1> <cat,1>

<prince,1> <flower,1> <sun,1>

<king,1> <queen,1> <AC,1>

<AC,1><cat,1><cat,1><cat,1>

<flower,1><flower,1><king,1>

<picture,1><prince,1><queen,1><snow,1><sun,1>

<cat,1><flower,1><picture,1>

<cat,1><cat,1>

<snow,1><flower,1><prince,1><sun,1>

<AC,1><king,1>

<queen,1>



● Step4: Start R reduce workers, each reads a list of intermediate with a unique key from remote disks.

Reducer 1

Reducer 2

Reducer 3

Reducer 9

<AC,1>

<cat,3>

<flower,2>

<sun,1>

<AC,1><cat,1><cat,1><cat,1>

<flower,1><flower,1><king,1>




● Step5: Write the output of reduce workers to file(s).

<AC,1>

<cat,3>

<flower,2>

<sun,1>

<AC,1><cat,3>

<flower,2><king,1>


<king,1>

<picture,1>


MapReduce framework


MapReduce Failure Recovery

● The framework works as master worker paradigm.

● The master keeps records of the work done on each worker.

● If a worker fails, the master assigns the same work to another worker.

● If a worker is late, another copy of the same work is assigned to another worker.

● If the master fails, another backup copy of the master can pick up and continue execution from the last check points.


Advantages of MapReduce

● Parallel IO: hides disk latency.

● Parallel Processing:

– Map functions works independently in parallel, each process one unique partition.

– Reduce functions work independently in parallel, each on a unique intermediate key.

● Using large clusters of commodity machines gives better results than small expensive clusters.


Advantages of MapReduce

● Parallel IO: hides disk latency.

● Parallel Processing:

– Map functions works independently in parallel, each process one unique partition.

– Reduce functions work independently in parallel, each on a unique intermediate key.

● Using large clusters of commodity machines gives comparable results than small expensive clusters.


Hadoop vs. others

● Algorithm: Sorting 100 TB data.

Hadoop DEMSort TritonSort

Nodes Count 3452 195 47

Processor 2x Quad-core Xeons @ 2.5 GHz

2x Quad-core Xeons @ 2.6 GHz

2x Quad-core Xeons @ 2.27 GHz

Memory 8 GB 16 GB 24 GB

Network 1 Gigabit Ethernet InfiniBand 10 Gigabit Fiber

Throughput 0.578 TB/Min 0.564 TB/Min 0.582 TB/Min


MapReduce weak points

● Overhead of MapReduce is huge.

● Data dependent applications may need multiple iterations of MapReduce, for example:

– K-means.

– PageRank.● Complex algorithms can be very hard to implement.

– Range Queries.● Sensitive to <key,value> pairs' skewed distribution


Implementations of MapReduce

● Hadoop in Java.

● Mars in C++ & CUDA.

● Skynet in Ruby.

● Phoenix in C++

● Microsoft Dryad:

– Schedule multiple levels of “MapReduce” like operations..


MapReduce in Database


MapReduce in Database - Ex1

● Select Name from Students where age = 23;

Name ID Age

Ahmed 1177 23

Bob 1131 20

Sara 1197 22

Students:



● Select COUNT(Name) from Students where age > 20 group by Name;

Name ID Age

Ahmed 1177 23

Bob 1131 20

Sara 1197 22

Students:



● Select Name, Term from Students, Enrolment where ID = SID and age != 20;

CID SID Term

CS290 1177 042

CS260 1177 052

ME222 1131 051

AMCS220 1197 051

Enrolment:Name ID Age

Ahmed 1177 23

Bob 1131 20

Sara 1197 22

Students:



● Select Name, Term from Students, Enrolment where ID != SID;

● What if the condition ID > SID?

CID SID Term

CS290 1177 042

CS260 1177 052

ME222 1131 051

AMCS220 1197 051

Name ID Age

Ahmed 1177 23

Bob 1131 20

Sara 1197 22

Enrolment:Students:



● Select Name, Term from Students, Enrolment where ID = SID and Admission != Term;

CID SID Term

CS290 1177 042

CS260 1177 052

ME222 1131 051

AMCS220 1197 051

Enrolment:Name ID Age Admission

Ahmed 1177 23 042

Bob 1131 20 051

Sara 1197 22 042

Students: Enrolment:Students:



● Select y from R, S, T where R.x = S.x and T.a = S.a;

a b xS:

x y zR:

m n aT:


MapReduce in Academic Papers

● NIPS '07: Map-Reduce for Machine Learning on Multicore.

● Escience '08: CloudBLAST: Combining MapReduce and Virtualization on Distributed Resources for Bioinformatics Applications.

● KDD '09: Large-scale behavioral targeting.

● GCC '09: Spatial Queries Evaluation with MapReduce.

● SIGIR '09: On single-pass indexing with MapReduce.

● MDAC '10: A novel approach to multiple sequence alignment using hadoop data grids.

● VLDB Endowment '11: Social Content Matching in MapReduce.

● VLDB '12: Building Wavelet Histograms on Large Data in MapReduce.


Links

● http://code.google.com/edu/parallel/mapreduce-tutorial.html

● http://hadoop.apache.org/mapreduce/

● http://www.cse.ust.hk/gpuqp/Mars.html

● http://skynet.rubyforge.org/

● http://mapreduce.stanford.edu/

● http://wiki.apache.org/hadoop/PoweredBy

● http://atbrox.com/2011/05/16/mapreduce-hadoop-algorithms-in-academic-papers-4th-update-may-2011/

Education

MapReduce