These are slides used in a guest lecture at Eindhoven Technical University, on May 30th, 2012. Adapted from slides by Jimmy Lin.

1. Guest Lecture Eindhoven University of Technology
Notes on Data-Intensive Processing
with Hadoop MapReduce
Evert Lammerts
May 30, 2012
Image source: http://valley-of-the-shmoon.blogspot.com/2011/04/pushing-elephant-up-stairs.html

2. To start with...
● About me
●
Note on this lecture
● Adapted from Jimmy Lin's Cloud Computing course...
http://www.umiacs.umd.edu/~jimmylin/cloud-2010-Spring/index.html
● … and from Jimmy's slidedeck from the SIKS Big Data course and his talk at UvA
http://www.umiacs.umd.edu/~jimmylin/
● Today's slides available at
http://www.slideshare.net/evertlammerts
●
About you
● Big Data?
● Cloud computing?
● Supercomputing?
● Hadoop and / or MapReduce?

3. The lecture
● Why “Big Data”?
● How “Big Data”?
● MapReduce
● Implementations

4. Why “Big Data”?
The Economist, Feb 25th 2010

5. 1. Science
● The emergence of the 4th paradigm
● http://research.microsoft.com/en-us/collaboration/fourthparadigm/
● CERN stores 15 PB LHC data per year, a fraction of the actual produced
data
● Square Kilometer Array expectation: 10 PB / hour
Adapted from (Jimmy Lin, University of Maryland / Twitter, 2011)

6. 2. Engineering
● Count and normalize
http://infrawatch.liacs.nl/
Adapted from (Jimmy Lin, University of Maryland / Twitter, 2011)

7. 3. Commerce
● Know thy customers
● Data → Insights → Competitive advantages
● Google was processing 20 PB each day... in 2008!
● FaceBook's collected 25 TB of HTTP logs each day... in 2009!
● eBay had ~9 PB of user data, and a growth rate of more than 50 TB /
day in 2011
Adapted from (Jimmy Lin, University of Maryland / Twitter, 2011)

8. IEEE Intelligent Systems, March/April 2009

9. s/knowledge/data/g
Jimmy Lin, University of Maryland / Twitter, 2011

10. Also see
● P. Russom, Big Data Analytics, The Data Warehousing Institute, 2011
● James G. Kobielus, The Forrester Wave™: Enterprise Hadoop
Solutions, Forrester Research, 2012
● James Manyika et al., Big data: The next frontier for innovation,
competition, and productivity, McKinsey Global Institute, 2011
● Dirk de Roos et al., Understanding Big Data: Analytics for Enterprise
Class Hadoop and Streaming Data, IBM, 2011
Etcetera

11. How “Big Data”?

13. Divide and Conquer
“Work”
Partition
w1 w2 w3
“worker” “worker” “worker”
r1 r2 r3
“Result” Combine
Jimmy Lin, University of Maryland / Twitter, 2011

14. Amdahl's Law

15. Challenges in Parallel systems
● How do we divide the work into separate tasks?
● How do we get these tasks to our workers?
● What if we have more tasks than workers?
● What if our tasks need to exchange information?
● What if workers crash? (That's no exception!)
● How do we aggregate results?

16. Managing Parallel Applications
● A synchronization mechanism is needed
● To coordinate communication (like exchanging state) between workers
● To manage access to shared resources like data
● What if you don't?
● Mutual Exclusion
● Resource Starvation
● Race Conditions
● Dining philosophers, sleeping barber, cigarette smokers, readers-writers,
producers-consumers, etcetera
Managing parallelism is hard!

17. Source: Ricardo Guimarães Herrmann

18. Well known tools and patterns
● Programming models Shared Memory Message Passing
Shared memory (pthreads)
Memory
●
● Message passing (MPI)
● Design patterns P1 P2 P3 P4 P5 P1 P2 P3 P4 P5
● Master-slave
● Producer-consumer
● Shared queues
producer consumer
master
work queue
slaves
producer consumer
Jimmy Lin, University of Maryland / Twitter, 2011

19. From Von Neumann...
http://www.lrr.in.tum.de/~jasmin/neumann.html

20. … to a datacenter

22. Where to go from here
● The search for the right level of abstraction
● How do we build an architecture for a scaled environment?
● From HAL to DCAL
● Hiding parallel application management from the developer
● It's hard!
● Separating the what from the how
● The developer specifies the computation
● The runtime environment handles the execution
Barosso, 2009

23. Ideas on scaling
● Scale “out”, don't scale “up”
● Hard upper-bound on the capacity of a single machine
● No upper-bound on the amount of machines you can buy (in theory)
● When dealing with large data...
● Prefer sequential reads over random reads
& rather not store a trillion small files, but a million big ones
– Disk access is slow, but throughput is reasonable!
● Try to understand when a NAS / SAN architecture is really necessary
– It's expensive to scale!

24. MapReduce

25. An abstraction of typical large-data problems
(1) Iterate over a large number of records
(2) Extract something of interest from each
(3) Shuffle and sort intermediate results
(4) Aggregate intermediate results
(5) Generate final output

26. An abstraction of typical large-data problems
(1) Iterate over a large number of records
M
(2) Extract something of interest from each A P
(3) Shuffle and sort intermediate R
results
ED
(4) Aggregate intermediate results U
C
(5) Generate final output E
MapReduce provides a functional abstraction of step 2 and step 4

27. Roots in functional programming
Map(S: array, f())
● Apply f(s ∈ S) for all items in S
Fold(S: array, f())
● Recursively apply f() to each item in S and the result of the previous
operation, or nil if such an operation does not exist
Source: Wikipedia

28. MapReduce
The programmer specifies two functions:
● map(k, v) → <k', v'>*
● reduce(k', v'[ ]) → <k', v'>*
All values associated with the same key are sent to the same reducer
The execution framework handles everything else

29. k1 v1 k2 v2 k3 v3 k4 v4 k5 v5 k6 v6
map map map map
a 1 b 2 c 3 c 6 a 5 c 2 b 7 c 8
Shuffle and Sort: aggregate values by keys
a 1 5 b 2 7 c 2 3 6 8
reduce reduce reduce
r1 s1 r2 s2 r3 s3
Jimmy Lin, University of Maryland / Twitter, 2011

30. MapReduce “Hello World”: WordCount
● Question: how can we count unique words in a given text?
● Line-based input (a record is one line)
● Key: position of first character in the whole document
● Value: a line not including the EOL character
● Input looks like:
Key: 0, value: “a wise old owl lived in an oak”
Key: 31, value: “the more he saw the less he spoke”
Key: 63, value: “the less he spoke the more he heard”
Key: 99, value: “why can't we all be like that wise old bird”
● Output looks like:
(a,1) (an,1) (be,1)
(he,4) (in,1) (we,1)
(all,1) (oak,1) (old,2)
(owl,1) (saw,1) (the,4)
(why,1) (bird,1) (less,2)
(like,1) (more,2) (that,1)
(wise,2) (can't,1) (heard,1)
(lived,1) (spoke,2)

31. MapReduce “Hello World”: WordCount

32. MapReduce
The programmer specifies two functions:
● map(k, v) → <k', v'>*
● reduce(k', v'[ ]) → <k', v'>*
All values associated with the same key are sent to the same reducer
The “execution framework” handles ? everything else ?

33. MapReduce execution framework
● Handles scheduling
● Assigns map and reduce tasks to workers
● Handles “data-awareness”: moves processes to data
● Handles synchronization
● Gathers, sorts, and shuffles intermediate data
● Handles errors and faults
● Detects worker failures and restarts
● Handles communication with the distributed filesystem

34. MapReduce
The programmer specifies two functions:
● map (k, v) → <k', v'>*
● reduce (k', v'[ ]) → <k', v'>*
All values associated with the same key are sent to the same reducer
The execution framework handles everything else...
Not quite... usually, programmers also specify:
● partition (k', number of partitions) → partition for k'
● Often a simple hash of the key, e.g., hash(k') mod n
● Divides up key space for parallel reduce operations
● combine (k', v') → <k', v'>*
● Mini-reducers that run in memory after the map phase
● Used as optimization to reduce network traffic

35. k1 v1 k2 v2 k3 v3 k4 v4 k5 v5 k6 v6
map map map map
a 1 b 2 c 3 c 6 a 5 c 2 b 7 c 8
combine combine combine combine
a 1 b 2 c 9 a 5 c 2 b 7 c 8
partition partition partition partition
Shuffle and Sort: aggregate values by keys
a 1 5 b 2 7 c 2 9 8
3 6
reduce reduce reduce
r1 s1 r2 s2 r3 s3
Jimmy Lin, University of Maryland / Twitter, 2011

36. Quick note...
The term “MapReduce” can refer to:
● The programming model
● The “execution framework”
● The specific implementation

37. Implementation(s)

38. MapReduce implementations
● Google (C++)
● Dean & Ghemawat, MapReduce: simplified data processing on large
clusters, 2004
● Ghemawat, Gobioff, Leung, The Google File System, 2003
● Apache Hadoop (Java)
● Open source implementation
● Originally led by Yahoo!
● Broadly adopted
● Custom research implementations
● For GPU's, supercomputers, etcetera

39. User
Program
(1) submit
Master
(2) schedule map (2) schedule reduce
worker
split 0
(6) write output
split 1 (5) remote read worker
(3) read file 0
split 2 (4) local write
worker
split 3
split 4 output
worker
file 1
worker
Input Map Intermediate files Reduce Output
files phase (on local disk) phase files
Jimmy Lin, Adapted from (Dean and Ghemawat, OSDI 2004)

40. User
Program
(1) submit
Master
(2) schedule map (2) schedule reduce
worker
split 0
(6) write output
split 1 (5) remote read worker
(3) read file 0
split 2 (4) local write
worker
split 3
split 4 output
worker
file 1
worker
Input Map Intermediate files Reduce Output
files phase (on local disk) phase files
Jimmy Lin, Adapted from (Dean and Ghemawat, OSDI 2004)

41. User
Program
(1) submit
Master
(2) schedule map (2) schedule reduce
worker
split 0
(6) write output
split 1 (5) remote read worker
(3) read file 0
split 2 (4) local write
worker
split 3
split 4 output
worker
file 1
worker
Input Map Intermediate files Reduce Output
files phase (on local disk) phase files
How do we get our input data to the map()'s on the workers?
Jimmy Lin, Adapted from (Dean and Ghemawat, OSDI 2004)

42. Distributed File System
● Don't move data to the workers... move workers to the data!
● Store data on the local disks of nodes in the cluster
● Start up the work on the node that has the data local
● A distributed files system is the answer
● GFS (Google File System) for Google's MapReduce
● HDFS (Hadoop Distributed File System) for Hadoop

43. GFS: Design decisions
● Files stored as chunks
● Fixed size (64MB)
● Reliability through replication
● Each chunk replicated across 3+ chunkservers
● Single master to coordinate access, keep metadata
● Simple centralized management
● No data caching
● Little benefit due to large datasets, streaming reads
● Simplify the API
● Push some of the issues onto the client (e.g., data layout)
HDFS = GFS clone (same basic ideas)
Jimmy Lin, Adapted from (Ghemawat, SOSP 2003)

44. From GFS to HDFS
● Terminology differences:
● GFS Master = Hadoop NameNode
● GFS Chunkservers = Hadoop DataNode
● Chunk = Block
● Functional differences
● File appends in HDFS is relatively new
● HDFS performance is (likely) slower
● Blocksize is configurable by the client
We use Hadoop terminology

45. HDFS Architecture
HDFS namenode
Application /foo/bar
(file name, block id)
File namespace block 3df2
HDFS Client
(block id, block location)
instructions to datanode
datanode state
(block id, byte range)
HDFS datanode HDFS datanode
block data
Linux file system Linux file system
… …
Jimmy Lin, Adapted from (Ghemawat, SOSP 2003)

46. Namenode Responsibilities
● Managing the file system namespace:
● Holds file/directory structure, metadata, file-to-block mapping, access
permissions, etcetera
● Coordinating file operations
● Directs clients to DataNodes for reads and writes
● No data is moved through the NameNode
● Maintaining overall health:
● Periodic communication with the DataNodes
● Block re-replication and rebalancing
● Garbage collection

47. Putting everything together
namenode job submission node
namenode daemon jobtracker
tasktracker tasktracker tasktracker
datanode daemon datanode daemon datanode daemon
Linux file system Linux file system Linux file system
… … …
slave node slave node slave node
Jimmy Lin, University of Maryland / Twitter, 2011

48. Questions?