4. Challenges at Large Scale
● Single node can't handle due to limited
resource
○ Processor time, Memory, Hard drive space, Network
bandwidth
○ Individual hard drives can only sustain read speeds between
60-100 MB/second, so multicore does not help that much
● Multiple nodes needed, but probability of
failure increases
○ Network failure, Data transfer failure, Node failure
○ Desynchronized clock, Lock
○ Partial failure in distributed atomic transaction
5. Hadoop Approach (1/4)
● Data Distribution
○ Distributed to all the nodes in the cluster
○ Replicated to several nodes
6. Hadoop Approach (2/4)
● Move computation to the data
○ Whenever possible, rather than moving data for
processing, computation is moved to the node that
contains the data
○ Most data is read from local disk straight into the
CPU, alleviating strain on network bandwidth and
preventing unnecessary network transfers
○ This data locality results in high performance
8. Hadoop Approach (4/4)
● Isolated execution
○ Communication between nodes is limited and done
implicitly
○ Individual node failures can be worked around by
restarting tasks on other nodes
■ No message exchange needed by user task
■ No roll back to pre-arranged checkpoints to
partially restart the computation
■ Other workers continue to operate as though
nothing went wrong
11. HDFS (1/2)
● Storage component of Hadoop
● Distributed file system modeled after GFS
● Optimized for high throughput
● Works best when reading and writing large files
(gigabytes and larger)
● To support this throughput HDFS leverages unusually
large (for a filesystem) block sizes and data locality
optimizations to reduce network input/output (I/O)
12. HDFS (2/2)
● Scalability and availability are also key
traits of HDFS, achieved in part due to data
replication and fault tolerance
● HDFS replicates files for a configured
number of times, is tolerant of both
software and hardware failure, and
automatically re-replicates data blocks on
nodes that have failed
14. MapReduce (1/2)
● MapReduce is a batch-based, distributed
computing framework modeled
● Simplifies parallel processing by abstracting
away the complexities involved in working
with distributed systems
○ computational parallelization
○ work distribution
○ dealing with unreliable hardware and software
17. Hadoop Installation
● Local mode
○ No need to communicate with other nodes, so it
does not use HDFS, nor will it launch any of the
Hadoop daemons
○ Used for developing and debugging the application
logic of a MapReduce program
● Pseudo Distributed Mode
○ All daemons running on a single machine
○ Helps to examine memory usage, HDFS
input/output issues, and other daemon interactions
● Fully Distributed Mode
18. Hadoop Configuration
File name Description
hadoop-env.sh ● Environment-specific settings go here.
● If a current JDK is not in the system path you’ll want to come here to configure your
JAVA_HOME
core-site.xml ● Contains system-level Hadoop configuration items
○ HDFS URL
○ Hadoop temporary directory
○ script locations for rack-aware Hadoop clusters
● Override settings in core-default.xml: http://hadoop.apache.org/common/docs/r1.
0.0/core-default.html.
hdfs-site.xml ● Contains HDFS settings
○ default file replication count
○ block size
○ whether permissions are enforced
● Override settings in hdfs-default.xml: http://hadoop.apache.org/common/docs/r1.
0.0/hdfs-default.html
mapred-site.xml ● Contains HDFS settings
○ default number of reduce tasks
○ default min/max task memory sizes
○ speculative execution
● Override settings in mapred-default.xml: http://hadoop.apache.
org/common/docs/r1.0.0/mapred-default.html
19. Installation
Pseudo Distributed Mode
● Setup public key based login
○ ssh-keygen -t rsa -P '' -f ~/.ssh/id_rsa
○ cat ~/.ssh/id_rsa.pub >> ~/.ssh/authorized_keys
● Update the following configuration
○ hadoop.tmp.dir and fs.default.name at core-site.
xml
○ dfs.replication at hdfs-site.xml
○ mapred.job.tracker at mapred-site.xml
● Format NameNode
○ bin/hadoop namenode -format
● Start all daemons
○ bin/start-all.sh
21. Hadoop FileSystem
File
System
URI
Scheme
Java Impl. (all under org.
apache.hadoop)
Description
Local file fs.LocalFileSystem Filesystem for a locally
connected disk with client-side
checksums
HDFS hdfs hdfs.DistributedFileSystem Hadoop’s distributed filesystem
WebHDFS webhdfs hdfs.web.
WebHdfsFileSystem
Filesystem providing secure
read-write access to HDFS over
HTTP
S3 (native) s3n fs.s3native.
NativeS3FileSystem
Filesystem backed by Amazon
S3
S3 (block
based)
s3 fs.s3.S3FileSystem Filesystem backed by Amazon
S3, which stores files in blocks
(much like HDFS) to overcome
S3’s 5 GB file size limit.
GlusterFS glusterfs fs.glusterfs.
GlusterFileSystem
Still in beta
https://github.
com/gluster/glusterfs/tree/master
/glusterfs-hadoop
22. Installation
Fully Distributed Mode
Three different kind of hosts:
● master
○ master node of the cluster
○ hosts NameNode and JobTracker daemons
● backup
○ hosts Secondary NameNode daemon
● slave1, slave2, ...
○ slave boxes running both DataNode and TaskTracker
daemons
23. Hadoop Configuration
File Name Description
masters ● Name is misleading and should have been called secondary-masters
● When you start Hadoop it will launch NameNode and JobTracker on the local
host from which you issued the start command and then SSH to all the nodes
in this file to launch the SecondaryNameNode.
slaves ● Contains a list of hosts that are Hadoop slaves
● When you start Hadoop it will SSH to each host in this file and launch the
DataNode and TaskTracker daemons
24. Recipes
● S3 Configuration
● Using multiple disks/volumes and limiting
HDFS disk usage
● Setting HDFS block size
● Setting the file replication factor
25. Recipes:
S3 Configuration
● Config file: conf/hadoop-site.xml
● To access S3 data using DFS command
<property>
<name>fs.s3.awsAccessKeyId</name>
<value>ID</value>
</property>
<property>
<name>fs.s3.awsSecretAccessKey</name>
<value>SECRET</value>
</property>
● To use S3 as a replacement for HDFS
<property>
<name>fs.default.name</name>
<value>s3://BUCKET</value>
</property>
26. Recipes:
Disk Configuration
● Config file: $HADOOP_HOME/conf/hdfs-site.xml
● For multiple locations:
<property>
<name>dfs.data.dir</name>
<value>/u1/hadoop/data,/u2/hadoop/data</value>
</property>
● For limiting the HDFS disk usage, specify reserved
space for non-DFS (bytes per volume)
<property>
<name>dfs.datanode.du.reserved</name>
<value>6000000000</value>
</property>
27. Recipes:
HDFS Block Size (1/3)
● HDFS stores files across the cluster by
breaking them down into coarser grained,
fixed-size blocks
● Default HDFS block size is 64 MB
● Affects performance of
○ filesystem operations where larger block sizes
would be more effective, if you are storing and
processing very large files
○ MapReduce computations, as the default behavior
of Hadoop is to create one map task for each data
block of the input files
28. Recipes:
HDFS Block Size (2/3)
● Option 1: NameNode configuration
○ Add/modify dfs.block.size parameter at conf/hdfs-
site.xml
○ Block size in number of bytes
○ Only the files copied after the change will have the
new block size
○ Existing files in HDFS will not be affected
<property>
<name>dfs.block.size</name>
<value>134217728</value>
</property>
29. Recipes:
HDFS Block Size (2/3)
● Option 2: During file upload
○ Applies only to the specific file paths
> bin/hadoop fs -Ddfs.blocksize=134217728 -put data.in /user/foo
● Use fsck command
> bin/hadoop fsck /user/foo/data.in -blocks -files -locations
/user/foo/data.in 215227246 bytes, 2 block(s): ....
0. blk_6981535920477261584_1059len=134217728 repl=1 [hostname:50010]
1. blk_-8238102374790373371_1059 len=81009518 repl=1 [hostname:50010]
30. Recipes:
File Replication Factor (1/3)
● Replication done for fault tolerance
○ Pros: Improves data locality and data access
bandwidth
○ Cons: Needs more storage
● HDFS replication factor is a file-level
property that can be set per file basis
31. Recipes:
File Replication Factor (2/3)
● Set default replication factor
○ Add/Modify dfs.replication property in conf/hdfs-
site.xml
○ Old files will be unaffected
○ Only the files copied after the change will have the
new replication factor
<property>
<name>dfs.replication</name>
<value>2</value>
</property>
32. Recipes:
File Replication Factor (3/3)
● Set replication factor during file upload
> bin/hadoop fs -D dfs.replication=1 -copyFromLocal non-criticalfile.txt
/user/foo
● Change the replication factor of files or file
paths that are already in the HDFS
○ Use setrep command
○ Syntax: hadoop fs -setrep [-R] <path>
> bin/hadoop fs -setrep 2 non-critical-file.txt
Replication 3 set: hdfs://myhost:9000/user/foo/non-critical-file.txt
33. Recipes:
Merging files in HDFS
● Use HDFS getmerge command
● Syntax:
hadoop fs -getmerge <src> <localdst> [addnl]
● Copies files in a given path in HDFS to a
single concatenated file in the local
filesystem
> bin/hadoop fs -getmerge /user/foo/demofiles merged.txt
35. Example:
Advanced Operations
● HDFS
○ Adding new data node
○ Decommissioning data node
○ Checking FileSystem Integrity with fsck
○ Balancing HDFS Block Data
○ Dealing with a Failed Disk
● MapReduce
○ Adding a Tasktracker
○ Decommissioning a Tasktracker
○ Killing a MapReduce Job
○ Killing a MapReduce Task
○ Dealing with a Blacklisted Tasktracker