1. Map reduce
Cs 595
Lecture 11

2. Map Reduce & Hadoop MapReduce: Pioneered by Google
A programming model for expressing distributed computations on massive amounts of data
Execution framework for large scale data processing on clusters of commodity servers
Pioneered by Google
Processes 20PB of data per day
Popularized by the open source Hadoop project
Used by Yahoo!, Facebook, Amazon, …
GFS: Google File System
HDFS: Hadoop File System

3. What is Map Reduce?
Simple data-parallel programming model designed for scalability and fault tolerance
Composed of
Mapping method(s) that performs filtering and sorting of inputs
Ex: Word counting map functions break input strings into tokens and outputs a key/value (word/count) pair for each word.
Reducing method(s) that are called for each key in the key/value pairs generated by the mapping functions.
Ex: in the word counting example, reduce functions take key/value pairs from the mappers, then sums and generates a single output count for each word.

4. What is Map Reduce? Input Map/Combine Shuffle/Sort Reduce Output Map
the quick
brown fox
the fox ate
the mouse
how now
brown cow
Map
Reduce
brown, 2
fox, 2
how, 1
now, 1
the, 3
ate, 1
cow, 1
mouse, 1
quick, 1
the, 1
brown, 1
fox, 1
the, 2

5. Where is Map Reduce used?
In research:
Astronomical image analysis (Washington)
Ocean climate simulation (Washington)
Bioinformatics (Maryland)
Analyzing Wikipedia conflicts (PARC)
Natural language processing (CMU)
Particle physics (Nebraska)
In Industry:
Reporting/Machine learning (Facebook)
Process tweets and log files (Twitter)
User activity, server metrics, transaction logs (LinkedIn)
Scaling tests (Yahoo!)
Search optimization and research (Ebay)

6. Map Reduce design goals
Scalability to accommodate BIG DATA
1000’s of machines with 10,000’s of disks
*Ebay’s Hadoop Cluster

7. Map Reduce design goals
Cost efficiency:
Commodity machines
Cheap, but unreliable
Commodity network
Gigabit Ethernet
Automatic fault tolerance
Fewer administrators needed
Easy to use
Fewer programmers needed

8. Map Reduce challenges Cheap nodes prone to failure
Mean time between failures for one node:
Mean time between failures for 1000 nodes:
Solution: Build fault tolerance into MapReduce
Commodity network = low bandwidth
Solution: Push computation to the data.
Hadoop framework built on top of HDFS
DataNodes in HDFS also responsible for handling computational requests
Requires precise mapping of application computations to specific DataNodes
Programming distributed systems is difficult
Solution: Data-parallel programming model
Users write “map” and “reduce” functions
Hadoop distributes work and handles faults
3 years
1 day

9. File Systems What is a file system?
File systems determine how data is stored and retrieved
Distributed file systems manage the storage across a network of servers
Increased complexity due to networking
GFS/HDFS are distributed file systems

10. GFS Assumptions *HDFS is heavily influenced by GFS
Hardware failures are common
Files are large (GB/TB)
Millions (not billions) of files
Two main types of file reads:
Large streaming reads
Small random reads
Jobs include sequential writes that append data to files
Once written, files are seldom modified, other than append
Random file modification is possible, but not efficient in GFS
High sustained bandwidth has priority over low latency
*HDFS is heavily
influenced by GFS

11. GFS/HDFS GFS/HDFS are not a good fit for:
Low latency (ms) data access
Many small files
Constantly changing file structure/data
*Not all details of GFS are public knowledge

12. Question How would you design a distributed file system? Application
How to write data to the cluster?
How to read data from the cluster?

13. GFS Architecture revisited

14. (chunk_handle, byte_range)
Files on GFS
A single file can contain main objects
Web documents
Logs
Divided into fixed size chunks of 64MB with 64 bit identifiers
264 number of unique files
264MB (Exabyte scale) total filesystem space for data
Uses Linux files
Reading/writing of data specified by they tuple
(chunk_handle, byte_range)

15. Chunks One chunk = 64MB or 128MB (can be changed)
Stored as a plain Linux file on a chunk server
Advantages of large, but not too large, chunk size:
Reduced need for client/master interaction
One request per chunk
Clients can cache all chunk locations for large data sets
Reduced size of metadata on the master
Kept in RAM
Disadvantage
Chunkserver can become a hotspot for popular file(s)
How could chunkserver
Hotspots be resolved?

16. HDFS: Hadoop Distributed File System

17. GFS vs. HDFS GFS HDFS Master NameNode ChunkServer DataNode
Operation Log
Journal, Edit Log
Chunk
Block
Random file writes possible
Only append is possible
Multiple writer/reader model
Single writer/multiple reader model
Default chunk size: 64MB
Default block size: 128MB

18. Hadoop architecture NameNode
Master of HDFS, directs DataNode services to perform low-level I/O tasks
Keeps track of:
File/block location
Location of replicas
Files sharing multiple blocks
Secondary NameNode
Replicates processes running in the primary NameNode for fault tolerance
DataNode
Services and stores files in HDFS

19. Hadoop name and data nodes

20. Hadoop cluster architecture

21. Fault tolerance in Hadoop
If a task crashes:
Retry task on another node
OK for map tasks because it has no dependencies
OK for reduce tasks because map outputs are on disk
If the same task fails repeatedly
Ignore that input block, if possible
Fail the job and notify the user

22. Fault tolerance in Hadoop
If a node crashes:
Relaunch its current tasks on other nodes
Rerun any map tasks that the failed node ran
Necessary because the output files were lost when the node failed

23. Fault tolerance in Hadoop
If a task is going slowly (Straggler):
Launch a second copy of the task on another node
Take the output of whichever copy finishes first, and terminate the other process.
Very important in large Hadoop clusters
Stragglers occur frequently due to failing hardware, software bugs, misconfiguration, etc.
Single stragglers may noticeably slow down a job.

24. Fault tolerance in Hadoop
Advantages of MapReduce/Hadoop:
Using the data-parallel programming model, MapReduce can control job execution in useful ways:
Automatic division of job into tasks
Automatic placement of computation near data
Automatic load balancing
Recovery from failures and stragglers
Users of the MapReduce/Hadoop system focus their efforts on the application instead of the complexities of distributed computing.

25. Hadoop in the wild: Yahoo!
HDFS cluster with 4,500 nodes
NameNodes
Up to 64GB RAM each
40 DataNodes/rack, one switch per rack
16GB RAM each
Gigabit Ethernet
70% of disk space allocated to HDFS
Total storage:
9.8PB  3.2PB with 3x replication
60 million files, 63 million blocks
54k blocks per DataNode
1-2 nodes lost per day
Time for cluster to re-replicate
lost blocks: two minutes