Lecture 4: Data-Intensive Computing for Text Analysis (Fall 2011)

Data-Intensive Computing for Text Analysis CS395T / INF385T / LIN386M

University of Texas at Austin, Fall 2011

Lecture 4 September 15, 2011

Matt Lease

School of Information

University of Texas at Austin

ml at ischool dot utexas dot edu

Jason Baldridge

Department of Linguistics

University of Texas at Austin

Jasonbaldridge at gmail dot com

Acknowledgments

Course design and slides based on Jimmy Lin’s cloud computing courses at the University of Maryland, College Park

Some figures courtesy of the following excellent Hadoop books (order yours today!)

• Chuck Lam’s Hadoop In Action (2010)

• Tom White’s Hadoop: The Definitive Guide, 2nd Edition (2010)

Today’s Agenda

• Practical Hadoop

– Input/Ouput

– Splits: small file and whole file operations

– Compression

– Mounting HDFS

– Hadoop Workflow and EC2/S3

Practical Hadoop

“Hello World”: Word Count

Map(String docid, String text):

for each word w in text:

Emit(w, 1);

Reduce(String term, Iterator<Int> values):

int sum = 0;

for each v in values:

sum += v;

Emit(term, sum);

map ( K( K1=String, V1=String ) → list ( K2=String, V2=Integer ) reduce ( K2=String, list(V2=Integer) ) → list ( K3=String, V3=Integer)

Courtesy of Chuck Lam’s Hadoop In Action (2010), p. 17

Courtesy of Chuck Lam’s Hadoop In Action (2010), pp. 48-49

Courtesy of Chuck Lam’s Hadoop In Action (2010), p. 51

Courtesy of Tom White’s Hadoop: The Definitive Guide, 2nd Edition (2010), p. 191

Command-Line Parsing

Courtesy of Tom White’s Hadoop: The Definitive Guide, 2nd Edition (2010), p. 135

Data Types in Hadoop

Writable Defines a de/serialization protocol.

Every data type in Hadoop is a Writable.

WritableComparable Defines a sort order. All keys must be

of this type (but not values).

IntWritable

LongWritable

Text

…

Concrete classes for different data types.

SequenceFiles Binary encoded of a sequence of

key/value pairs

Hadoop basic types

Courtesy of Chuck Lam’s Hadoop In Action (2010), p. 46

Complex Data Types in Hadoop

How do you implement complex data types?

The easiest way:

Encoded it as Text, e.g., (a, b) = “a:b”

Use regular expressions to parse and extract data

Works, but pretty hack-ish

The hard way:

Define a custom implementation of WritableComprable

Must implement: readFields, write, compareTo

Computationally efficient, but slow for rapid prototyping

Alternatives:

Cloud9 offers two other choices: Tuple and JSON

(Actually, not that useful in practice)

InputFormat &RecordReader

Note re-use key & value objects!

Courtesy of Tom White’s

Hadoop: The Definitive Guide,

2nd Edition (2010), pp. 198-199

Split is logical; atomic

records are never split

Courtesy of Tom White’s

Hadoop: The Definitive Guide,

2nd Edition (2010), p. 201

Input

Courtesy of Chuck Lam’s Hadoop In Action (2010), p. 53

Output

Courtesy of Chuck Lam’s Hadoop In Action (2010), p. 58

Source: redrawn from a slide by Cloduera, cc-licensed

Reducer Reducer Reduce

Output File

RecordWriter

Ou

tpu

tFo

rmat

Output File

RecordWriter

Output File

RecordWriter

Creating Input Splits (White p. 202-203)

FileInputFormat: large files split into blocks

isSplitable() – default TRUE

computeSplitSize() = max(minSize, min(maxSize,blockSize) )

getSplits()…

How to prevent splitting?

Option 1: set mapred.min.splitsize=Long.MAX_VALUE

Option 2: subclass FileInputFormat, set isSplitable()=FALSE

How to process whole file as a single record?

e.g. file conversion

Preventing splitting is necessary, but not sufficient

Need a RecordReader that delivers entire file as a record

Implement WholeFile input format & record reader recipe

See White pp. 206-209

Overrides getRecordReader() in FileInputFormat

Defines new WholeFileRecordReader

Small Files

Files < Hadoop block size are never split (by default)

Note this is with default mapred.min.splitsize = 1 byte

Could extend FileInputFormat to override this behavior

Using many small files inefficient in Hadoop

Overhead for TaskTracker, JobTracker, Map object, …

Requires more disk seeks

Wasteful for NameNode memory

How to deal with small files??

Dealing with small files

Pre-processing: merge into one or more bigger files

Doubles disk space, unless clever (can delete after merge)

Create Hadoop Archive (White pp. 72-73)

• Doesn’t solve splitting problem, just reduces NameNode memory

Simple text: just concatenate (e.g. each record on a single line)

XML: concatenate, specify start/end tags

StreamXmlRecordReader (as newline is end tag for Text)

Create a SequenceFile (see White pp. 117-118)

• Sequence of records, all with same (key,value) type

• E.g. Key=filename, Value=text or bytes of original file

• Can also use for larger files, e.g. if block processing is really fast

Use CombineFileInputFormat

Reduces map overhead, but not seeks or NameNode memory…

Only an abstract class provided, you get to implement it… :-<

Could use to speed up the pre-processing above…

Multiple File Formats?

What if you have multiple formats for same content type?

MultipleInputs (White pp. 214-215)

Specify InputFormat & Mapper to use on a per-path basis

• Path could be a directory or a single file

• Even a single file could have many records (e.g. Hadoop archive or

SequenceFile)

All mappers must have the same output signature!

• Same reducer used for all (only input format is different, not the

logical records being processed by the different mappers)

What about multiple file formats stored in the same

Archive or SequenceFile?

Multiple formats stored in the same directory?

How are multiple file types typically handled in general?

e.g. factory pattern, White p. 80

Data Compression

Big data = big disk space & I/O (bound) transfer times

Affects both intermediate (mapper output) and persistent data

Compression makes big data less big (but still cool)

Often 1/4th size of original data

Main issues

Does the compression format support splitting?

• What happens to parallelization if an entire 8GB compressed file has

to be decompressed before we can access the splits?

Compression/decompression ratio vs. speed

• More compression reduces disk space and transfer times, but…

• Slow compression can take longer than reduced transfer time savings

• Use native libraries!

White 77-86, Lam 153-155

Slow; decompression can’t keep pace disk reads

Courtesy of Tom White’s

Hadoop: The Definitive Guide,

2nd Edition (2010), Ch. 4

Compression Speed

LZO 2x faster than gzip

LZO ~15-20x faster than bzip2

http://arunxjacob.blogspot.com/2011/04/rolling-out-splittable-lzo-on-cdh3.html

http://www.cloudera.com/blog/2009/11/hadoop-at-twitter-part-1-splittable-lzo-compression/

Splittable LZO to the rescue

LZO format not internally splittable, but we can create

a separate, accompanying index of split points

Recipe

Get LZO from Cloudera or elsewhere, and setup

See URL on last slide for instructions

LZO compress files, copy to HDFS at /path

Index them: $ hadoop jar /path/to/hadoop-lzo.jar

com.hadoop.compression.lzo.LzoIndexer /path

Use hadoop-lzo’s LzoTextInputFormat instead of TextInputFormat

Voila!

Compression API for persistent data

JobConf helper functions –or– set properties

Input

conf.setInputFormatClass(LzoTextInputFormat.class);

Persistent (reducer) output

FileOutputFormat.setCompressOutput(conf, true)

FileOutputFormat.setOutputCompressorClass(conf, LzopCodec.class)

Courtesy of Tom White’s

Hadoop: The Definitive Guide,

2nd Edition (2010), p. 85

Compression API for intermediate data

Similar JobConf helper functions –or– set properties

conf.setCompressMapOutput()

Conf.setMapOutputCompressClass(LzopCodec.class)

Courtesy of Chuck Lam’s

Hadoop In Action(2010),

pp. 153-155

SequenceFile & compression

Use SequenceFile for passing data between Hadoop jobs

Optimized for this usage case

conf.setOutputFormat(SequenceFileOutputFormat.class)

With compression, one more parameter to set

Default compression per-record; almost always preferable to

compress on a per-block basis

See White p. 50;

hadoop: src/contrib/fuse-dfs

From “hadoop fs X” -> Mounted HDFS

Hadoop Workflow

Hadoop Cluster You

1. Load data into HDFS

2. Develop code locally

3. Submit MapReduce job 3a. Go back to Step 2

4. Retrieve data from HDFS

On Amazon: With EC2

You

1. Load data into HDFS

2. Develop code locally

3. Submit MapReduce job 3a. Go back to Step 2

4. Retrieve data from HDFS

0. Allocate Hadoop cluster

EC2

Your Hadoop Cluster

5. Clean up!

Uh oh. Where did the data go?

On Amazon: EC2 and S3

Your Hadoop Cluster

S3 (Persistent Store)

EC2 (The Cloud)

Copy from S3 to HDFS

Copy from HFDS to S3