Taking R to the Limit (High Performance Computing in R), Part 2 -- Large Datasets, LA R Users' Group 8/17/10

Los Angeles R Users’ Group

Taking R to the Limit, Part II:
Working with Large Datasets

Ryan R. Rosario

August 17, 2010

Ryan R. Rosario
Taking R to the Limit: Part II - Large Datasets Los Angeles R Users’ Group

The Brutal Truth
We are here because we love R. Despite our enthusiasm, R has two
major limitations, and some people may have a longer list.
1 Regardless of the number of cores on your CPU, R will only
use 1 on a default build. (Part I)
2 R reads data into memory by default. Other packages (SAS
particularly) and programming languages can read data from
ﬁles on demand.
Easy to exhaust RAM by storing unnecessary data.
232
The OS and system architecture can only access 10242 = 4GB
of memory on a 32 bit system, but typically R will throw an
exception at 2GB.
Not wise to use more memory than available. System will start
swapping which leads to thrashing and slows your system to a
crawl.

Ryan R. Rosario

The Brutal Truth

There are a couple of solutions:
1 Buy more RAM.
2 Use a database.
3 Build from source and use special build options for 64 bit.
Still not a solution!
1 R, even R64, does not have a int641 data type. Not possible
to index data frames or matrices with huge number of rows or
columns.
4 Sample, resample, or use some Monte Carlo method.
5 Let clever developers solve these problems for you!
We will discuss number 5 above.
1
http://permalink.gmane.org/gmane.comp.lang.r.devel/17281
Ryan R. Rosario

The Agenda

This talk will survey several HPC packages in R with some
demonstrations. We will focus on four areas of HPC:
1 Explicit Parallelism: the user controls the parallelization.
(Part I)
2 Implicit Parallelism: the system abstracts it away. (Part I)
3 Large Memory: working with large amounts of data
without choking R.
4 Map/Reduce: basically an abstraction for parallelism that
divides data rather than architecture.

Ryan R. Rosario

Disclaimer

1 There is a ton of material here. I provide a lot of slides for
your reference.
2 We may not be able to go over all of them in the time allowed
so some slides will go by quickly.
3 Demonstrations will be time permitting.
4 All experiments were run on a Ubuntu 10.04 system with an
Intel Core 2 6600 CPU (2.4GHz) with 2GB RAM. I am
planning a sizable upgrade this summer.

Ryan R. Rosario

Big Data

“Big Data” is a catch phrase for any dataset or data application that does not
ﬁt into available RAM on one system.

R has several a few packages for big data support. We will talk about the
following:
1 bigmemory
2 ff
We will also discuss some uses of parallelism to accomplish the same goal using
Hadoop and MapReduce:
1 HadoopStreaming
2 Rhipe

Ryan R. Rosario

Some Basic Terminology

I will use the word RAM to refer to physical memory, for
simplicity; chips that are installed on the system motherboard.

Virtual memory, or swap is disk space that is used to store
objects that do not ﬁt into RAM and are less frequently accessed.
This is SLOW.

A cluster is a group of systems that communicate with each other
to accomplish a computation.

Ryan R. Rosario

Background bigmemory ﬀ

Large Datasets

R reads data into RAM all at once, if using the usual read.table
function. Objects in R live in memory entirely. Keeping
unnecessary data in RAM will cause R to choke eventually.
Speciﬁcally,
1 on most systems it is not possible to use more than 2GB of
memory.
2 the range of indexes that can be used is limited due to lack of
a 64 bit integer data type in R and R64.
3 on 32 bit systems, the maximum amount of virtual memory
space is limited to between 2 and 4GB.
4 relying on virtual memory will cause the system to grind to a
halt – “thrashing.”

Ryan R. Rosario


Large Datasets

There are two major solutions in R:
1 bigmemory: “It is ideal for problems involving the analysis in
R of manageable subsets of the data, or when an analysis is
conducted mostly in C++.” It’s part of the “big” family,
some of which we will discuss.
2 ff: ﬁle-based access to datasets that cannot ﬁt in memory.
3 can also use databases which provide fast read/write access
for piecemeal analysis.

Ryan R. Rosario


The “big” Family

The big family consists of several packages for performing tasks
on large datasets.
1 bigmemory is our focus.
2 biganalytics provides analysis routines on big.matrix
such as GLM and bigkmeans.
3 synchronicity adds Boost mutex functionality to R.
4 bigtabulate adds table and split-like support for R
matrices and big.matrix memory eﬃciently.
5 bigalgebra provides BLAS and LAPACK linear algebra
routines for native R matrices and big.matrix.
6 bigvideo provides video camera streaming via OpenCV.

Ryan R. Rosario


bigmemory

bigmemory implements several matrix objects.
1 big.matrix is an R object that simply points to a data
structure in C++. Local to a single R process and is limited
by available RAM.
2 shared.big.matrix is similar, but can be shared among
multiple R processes (think parallelism on data!)
3 filebacked.big.matrix does not point to a data structure;
rather, it points to a ﬁle on disk containing the matrix, and
the ﬁle can be shared across a cluster!
Pitfall! Remember that matrices contain only one type of data.
Additionally, the data types for elements are dictated by C++ not
R: double, integer, short, char.

Ryan R. Rosario


bigmemory and Shared Memory
Shared Memory allows us to store data in RAM and share it among multiple
processes. Suppose we want to store some data in shared memory so it can be
read by multiple instances of R. This allows the user the ability to use
multiple instances of R for performing diﬀerent analytics simultaneously.

R Web User 1

Caching

My Dataset
(Shared Memory)

R Web User 2 Web Administrator
Doing Analysis

RAM

Ryan R. Rosario


Demonstration: The Logic

First, construct a big.matrix object. Let us suppose we want to
create a large matrix of 0s and 1s.

We can construct a matrix of type int (4 bytes), short (2 bytes),
double (8 bytes), or char (1 byte). Since all we need is 0 and 1,
we use char. We also zero out the matrix.

> A <- big.matrix(m, n, type="char", init=0, shared=TRUE)
> A
An object of class big.matrix
Slot "address":
<pointer: 0x2d93490>

Ryan R. Rosario


Demonstration: The Logic

We have now created a pointer to a C++ matrix that is on disk.
But, to share this matrix we need to share this descriptor.

Then, we can open a second R session, load the location of the
matrix from disk, and bind the matrix to an R variable!

Ryan R. Rosario


Demonstration: The Code
Session 1
1 library ( bigmemory )
2 library ( biganalytics )
3 options ( bigmemory . typecast . warning = FALSE )
4
5 A <- big . matrix (5000 , 5000 , type = " char " , init =0)
6 # Fill the matrix by r a n d o m l y picking 20% of the p o s i t i o n s
for a 1.
7 x <- sample (1:5000 , size =5000 , replace = TRUE )
8 y <- sample (1:5000 , size =5000 , replace = TRUE )
9 for ( i in 1:5000) {
10 A [ x [ i ] , y [ i ]] <- 1
11 }
12 # Get the l o c a t i o n in RAM of the pointer to A .
13 desc <- describe ( A )
14 # Write it to disk .
15 dput ( desc , file = " / tmp / A . desc " )
16 sums <- colsum (A , 1:20)

Ryan R. Rosario


Session 2
3
4 # Read the pointer from disk .
5 desc <- dget ( " / tmp / A . desc " )
6 # Attach to the pointer in RAM .
7 A <- attach . big . matrix ( desc )
8 # Check our results .
9 sums <- colsum (A , 1:20)

Ryan R. Rosario


Demonstration 1

Demonstration here.

Ryan R. Rosario


bigmemory: Importance of Data Type

Suppose we are not careful in the C++ data type we use for the
big.matrix. Using char, the matrix requires about 24MB of
RAM.

Data Type RAM
char 24 MB
int 96 MB
double 192 MB
short 48 MB

Of course, this assumes that you use dense matrices and there is
no CPU optimization.

Ryan R. Rosario


bigmemory: Other Operations
big.matrix requires its own optimized versions of the R matrix
functions provided in biganalytics:

colmean(x, cols, na.rm)
colmin(x, cols, na.rm)
colmax(x, cols, na.rm)
colvar(x, cols, na.rm)
colsd(x, cols, na.rm)
colsum(x, cols, na.rm)
colprod(x, cols, na.rm)
colna(x, cols)

where x is a big.matrix, cols is a vector of column indices,
na.rm is TRUE if R should remove missing values ﬁrst.
Ryan R. Rosario


The big Example: The Data

Let’s do something useful with some big data. Consider airline
on-time performance from 1987 to 2008. This is the format of the
data:
11GB comma-separated values ﬁle.
120 million rows, 29 columns
Factors coded as integers.
We will estimate the age of an aircraft at each departure.

Ryan R. Rosario


Step 1: Read in the Data

3 x <- read . big . matrix ( " airline . csv " , type = "
integer " , header = TRUE ,
4 backingfile = " airline . bin " ,
5 descriptorfile = " airline . desc " ,
6 extraCols = " Age " )
Initially takes about 28 minutes to run, the ﬁrst time only.
Subsequent accesses are very fast.

Ryan R. Rosario


Step 1: Read in the Data

read.big.matrix inherits from read.table so your favorite
parameters are available for use.
It also adds a few more:
1 type, the C++ type to use for the matrix.
2 separated, separate the columns into individual ﬁles if TRUE.
3 extracols explicitly adds columns to the matrix that you
may use later.
4 backingfile, backingpath and descriptorfile control
where important data about the matrix is stored on disk.

Ryan R. Rosario


Step 2a: Some Initial Tasks

Next, estimate the birthmonth of the plane using the ﬁrst
departure of that plane.
8 birthmonth <- function ( y ) {
9 minYear <- min ( y [ , ’ Year ’] , na . rm = TRUE )
10 these <- which ( y [ , ’ Year ’ ]== minYear )
11 minMonth <- min ( y [ these , ’ Month ’] , na . rm = TRUE )
12 return (12 * minYear + minMonth - 1)
13 }

Ryan R. Rosario


Step 2b v1: Calculate Each Plane’s Birthmonth the Dumb
Way

We could use a loop or possible an apply variant...
14 aircrafts <- unique ( x [ , ’ TailNum ’ ])
15 acStart <- rep (0 , length ( aircrafts ) )
16 for ( i in aircrafts ) {
17 acStart [ i ] <- birthmonth ( x [ mwhich (x , ’ TailNum ’ , i , ’ eq ’) ,
18 c ( ’ Year ’ , ’ Month ’) , drop = FALSE ])
19 }

...this takes about 9 hours...

Ryan R. Rosario


Step 2b v2: Calculate Each Plane’s Birthmonth the big
Way

First separate/divide the data by tail number (aircraft ID).
14 library ( bigtabulate )
15 acindices <- bigsplit (x , ’ TailNum ’)

Each entry i of acindices contains a vector of indices
corresponding to TailFin i.

bigsplit runs about twice as fast as split (6s) and requires
about 2/3 peak RAM.

Ryan R. Rosario


Step 2c v1: Compute an Estimate of Birthmonth using
sapply

Now that we have used bigsplit to quickly split up the
big.matrix, we can use the native sapply:
16 acStart <- sapply ( acindices , function ( i )
birthmonth ( x [i , c ( ’ Year ’ , ’ Month ’) , drop =
FALSE ]) )
which took a mere 8 seconds!

Ryan R. Rosario


Step 2c v2: Compute an Estimate of Birthmonth using
foreach

We can also use this opportunity to revisit foreach:
16 library ( doMC )
17 registerDoMC ( cores =2)
18 acStart <- foreach ( i = acindices , . combine = c )
% dopar % {
19 return ( birthmonth ( x [i , c ( ’ Year ’ , ’ Month ’) ,
drop = FALSE ]) )
20 }
Both cores share access to the same instance of the data smoothly.
Year and Month are cached in RAM. This example took 0.663s.

Ryan R. Rosario


Step 3: Finally, Compute an Estimate of Age

21 x [ , ’ Age ’] <- x [ , ’ Year ’] * as . integer (12) + x [ , ’ Month ’] -
as . integer ( acStart [ x [ , ’ TailNum ’ ]])

Here, arithmetic is conducted using R vectors that are extracted
from the big.matrix. Careful use of as.integer helps reduce
memory overhead.

This data management task requires 8 minutes.

Ryan R. Rosario


Before we Continue: The Point

The point of this demonstration:
1 show how, by using ﬁle descriptors and ﬁle-backed matrices,
working with big data is easy.
2 show how bigmemory can be integrated with parallelism
packages relatively easily.
3 show how other big packages assist with the purpose of the
bigmemory package to provide the user the familiar R
interface after performing some initial maintenance.

Ryan R. Rosario


The big Example 2: Linear Models

Now we will try to predict arrival delay (ArrDelay) using Age and
Year.

ˆ ˆ ˆ
ArrDelay = β0 + β1 Age + β2 Year
biganalytics provides a wrapper to the biglm package.

Ryan R. Rosario



Fitting a linear model looks relatively familiar...
2 blm <- biglm . big . matrix ( ArrDelay ~ Age + Year ,
data = x )

1 Without bigmemory, this process would take about 10GB of
RAM.
2 Expected time to completion would be ∞ since most systems
do not have that much RAM.
3 With bigmemory processing took 4.5 minutes with a few
hundred MB of memory overhead.

Ryan R. Rosario



Just like with base lm, we can get information about the ﬁtted
model.
> blm
Large data regression model: biglm(formula = formula, data = data, ...)
Sample size = 6452
> summary(blm)
Large data regression model: biglm(formula = formula, data = data, ...)
Sample size = 6452
Coef (95% CI) SE p
(Intercept) 6580.4311 1916.9282 11243.9340 2331.7515 0.0048
Age 0.2687 0.0804 0.4569 0.0941 0.0043
Year -3.2753 -5.6005 -0.9500 1.1626 0.0048

Ryan R. Rosario


bigmemory: Other Features

can write a big.matrix to ASCII ﬁle using
write.big.matrix.
can bin data for counting, creating histograms or
visualizations using binit.
create a copy of the content using deepcopy.
can create a hash into a big.matrix using hash.mat
search a matrix using mwhich including powerful comparison
operators using C++, not R.
separated columns: columns of a matrix are separated in
RAM, rather than contiguous.

Ryan R. Rosario


bigmemory: Important Considerations

passing a big.matrix to a function is call by reference not
call by value!
Nerd alert! bigmemory provides transparent read/write
locking to big.matrix, so race conditions are minimized.
For more information, see http://www.bigmemory.org, or the
documentation.

Ryan R. Rosario


bigmemory: Conclusion
Advantages
1 can store a matrix in memory, restart R, and gain access to
the matrix without reloading data. Great for big data.
2 can share the matrix among multiple R instances or sessions.
3 access is fast because RAM is fast. C++ also helps!
Disadvantages
1 no communication among instances of R; can use files instead.
2 limited by available RAM, unless using
filebacked.big.matrix.
3 matrix disappears on reboot, unless using
filebacked.big.matrix.
4 filesize limitations on 32 bit systems for filebacked matrices.
Ryan R. Rosario


ff: “fast access files”
ff is another solution that is based on using files.
provides data structures that are stored on disk.
they act as if they are in memory; only necessary/active parts
of the data from disk are mapped into main memory.
supports R standard atomic types: double, logical, raw,
integer,
as well as non-standard atomic types boolean (1 bit), quad
(2 bit unsigned), nibble (4 bit unsigned), byte (1 byte
signed with NA), ubyte (1 byte unsigned), short and ushort
(1 byte signed w/NA and unsigned resp.), and single (4 byte
float with NA).
and non-atomic types such as factor, ordered, POSIXct,
Date, etc.
Ryan R. Rosario


ff

C support for vectors, matrices and arrays. (FAST)
provides an analog for data.frame called ffdf, also with
import/export functionality.
ff objects can be stored and reopened across R sessions.
ff files can be shared by multiple ff R objects in the same, or
different R sessions.
tons of optimizations provide little noticeable overhead.
Virtual functions allow matrix operations without touching a
single byte on disk.
Disk I/O is SLOW. ff optimizes by using binary files.
closely integrated with package bit which can manipulate
and process, well, bits.
Ryan R. Rosario


ff: The Logic

In bigmemory R keeps a pointer to a C++ matrix. The matrix is
stored in RAM or on disk. In ff R keeps metadata about the
object, and the object is stored in a flat binary file.

ff is somewhat difficult to jump into because there are so few
examples and only one tutorial. There is a lot of information about
the technical side of the package.

In any case, the goal is to get rid of the following error message:
1 > x <- rep (0 , 2^31 - 1)
2 Error : cannot allocate vector of length
2147483647

Ryan R. Rosario


ff

The package ff performs the following functionality to the user:
creating and/or opening flat files using ffopen. Using the
parameters length or dim will create a new file.
I/O operations using the common brackets [ ] notation.
Other functions for ff objects such as the usual dim and
length as well as some other useful functions such as sample.

Ryan R. Rosario


ff Example: Introduction

Let’s start with an introductory demonstration. To create a
one-dimensional ﬂat ﬁle,
1 library ( ff )
2 # creating the file
3 my . obj <- ff ( vmode = " double " , length =10)
4 # modifying the file .
5 my . obj [1:10] <- iris $ Sepal . Width [1:10]
Let’s take a look...

Ryan R. Rosario



We can also create a multi-dimensional ﬂat ﬁle.
1 # c r e a t i n g a m u l t i d i m e n s i o n a l file
2 multi . obj <- ff ( vmode = " double " , dim = c (10 , 3) )
3 multi . obj [1:10 , 1] <- iris [1:10 ,1]


Ryan R. Rosario



We can also create an ff data frame made up of ff atomics.
1 Girth <- ff ( trees $ Girth )
2 Height <- ff ( trees $ Height )
3 Volume <- ff ( trees $ Volume )
4 # Create data frame with some added p a r a m e t e r s .
5 fftrees <- ffdf ( Girth = Girth , Height = Height , Volume = Volume )


Ryan R. Rosario


ff Parameters

ff, ffm and to some degree ffdf support some other options you
may need. R is usually pretty good at picking good defaults for
you. Some of them are below:

Parameter Description
initdata Value to use to initialize the object for construction.
length Optional length of vector. Used for construction.
vmode Virtual storage mode (makes big difference in memory overhead).
filename Give a name for the file created for the object.
overwrite If TRUE, allows overwriting of file objects.
If no filename is given, a new file is created. If a filename is given
and the file exists, the object will be loaded into R.

Ryan R. Rosario


ff Saving Data
We can use the ffsave function to save a series of objects to file. The specified
objects are saved using save and are given the extension .RData. The ff files related
to the objects are saved and zipped using an external zip application, and given the
extension .ffData. ffsave has some useful parameters. See ?ffsave for more
information:

ffsave(..., list = character(0L), file = stop("’file’ must be specified",
envir = parent.frame(), rootpath = NULL, add = FALSE,
move = FALSE, compress = !move, compression_level = 6,
precheck=TRUE)}

... the objects to be saved.
the file parameters specifies the root name (no extension) for the file. It is
best to use absolute paths.
add=TRUE to add these objects to an existing archive.
compress=TRUE to save and compress.
safe=TRUE to write a temporary file first for verification, then move to a
persistent location.

Ryan R. Rosario


ff: Other File I/O Operations

Before discussing ffload in a few slides, there are some other
operations worth mentioning.
ffsave.image allows the user to save the entire workspace to
an ffarchive.
ffinfo, when passed the path for an ffarchive (without
extension) displays information about the archive.
ffdrop allows the user to delete an ffarchive (no
extension).

Ryan R. Rosario


ff Example

As we saw earlier, we can use biglm to ﬁt a linear model to big
data.
7 model <- biglm ( log ( Volume ) ~ log ( Girth ) +
log ( Height ) , data = fftrees )
Demonstration here.

Ryan R. Rosario


ff: Conclusion

Advantages
1 allows R to work with multiple HUGE datasets.
2 clean system; does not make a mess with a ton of files.
3 several optimizations show that ff has a bright future.
Disadvantages
1 few examples; somewhat difficult to introduce.
2 performing analysis may require some clever forethought since
not all of the data is in RAM.
3 unzipping files on load is a huge bottleneck.

Ryan R. Rosario


bigmemory vs. ff
Which one to use is a matter of taste. Performance is about the
same: the ﬁrst row of numbers is the initial processing and the
second uses caching:

Source: http://www.agrocampus-ouest.fr/math/useR-2009/slides/Emerson+Kane.pdf
Ryan R. Rosario

Map/Reduce mapReduce HadoopStreaming

MapReduce

MapReduce is a way of dividing a large job into many smaller jobs producing
an output, and then combining the individual outputs into one output. It is a
classic divide and conquer approach that is embarrasingly parallel and can
easily be distributed among many cores or a CPU, or among many CPUs and
nodes. Oh, and it is patented by Google, its biggest user.
Two fundamental steps:
1 map step: perform some operation f , in parallel, on data and output a
key/value pair for each record/row.
2 reduce step: group common elements and compute some summary
statistic, one per group.
The notion of map and reduce comes from Lisp and functional programming,
where map performs some operation on every element in a collection, and
reduce collapses the results of that operation into one result for the collection.

Ryan R. Rosario


Example: Word Counts
Suppose we have 5 manuscripts and we want to count number of times each
word occurs. (This is a common task in data mining and natural language
processing). The map phase parses the text and produces output like word:
count.
The reduce phase then groups records by word and sums (the reduce
operation) the counts.

the: 200 the: 200 the: 200 the: 200 the: 200
it: 750 it: 750 it: 750 it: 750 it: 750
and: 1000 and: 1000 and: 1000 and: 1000 and: 1000
fire: 1 fire: 1 fire: 1 fire: 1 fire: 1
truck: 1 truck: 1 truck: 1 truck: 1 truck: 1

the: 1000
it: 3750
and: 5000
fire: 5
truck: 5

Ryan R. Rosario


mapReduce

mapReduce is a pure R implementation of MapReduce. A matter
of fact, the authors of the package state that mapReduce is simply:

apply(map(data), reduce)

By default, mapReduce uses the same parallelization functionality
as sapply which is not preferrable.

Ryan R. Rosario


mapReduce

The form of the mapReduce function is very simple.

mapReduce(map, ..., data, apply = sapply)

map is an expression that yields a vector that partitions data
into groups. Can use a single variable as well. This is a bit
diﬀerent from Hadoop’s implementation.
... is one or more reduce functions; typically summary
statistics.
data is a data.frame.
apply is the parallelization toolkit to use: papply,
multicore, snow.

Ryan R. Rosario


mapReduce Demonstration

1 data ( iris )
2 mapReduce (
3 map = Species ,
4 mean . sepal . length = mean ( Sepal . Length ) ,
5 max . sepal . length = max ( Sepal . Length ) ,
6 data = iris
7 )

mean.sepal.length max.sepal.length
setosa 5.006 5.8
versicolor 5.936 7.0
virginica 6.588 7.9

Demonstration here.
Ryan R. Rosario


Hadoop

Hadoop is an open-source implementation of MapReduce that has
gained a lot of traction in the data mining community.

http://hadoop.apache.org

Cloudera and Yahoo provide their own implementations of Hadoop,
which may be easier to use in the cloud:
http://www.cloudera.com/developers/downloads/hadoop-distro/

http://developer.yahoo.com/hadoop/distribution/

Ryan R. Rosario


Hadoop Infrastructure
Hadoop is a separate open source project, so will not discuss it in
detail.

1 Hadoop allows use of parallelism to solve problems using big
data.
2 Hadoop is most eﬀective with a cluster and Hadoop assigns
certain machines critical tasks.
3 Hadoop uses its own ﬁlesystem by default, called HDFS.
4 Map and reduce functions are the same as mentioned in the
Intro to Map-Reduce slide.
5 Map, reduce, job control, logging etc. methods are written in
Java.
We will exploit some workarounds to deal with point #5.
Ryan R. Rosario


Hadoop Installation

Installing Hadoop on most systems is simple.
1 Download the tarball from
http://www.apache.org/dyn/closer.cgi/hadoop/core/
2 Extract using tar xf.
3 Add the Hadoop bin directory to your path.
4 Set environment variable JAVA HOME to point to the location
of system Java (Oracle/Sun preferred).
Hadoop is then ready to run locally. To run on a cluster requires
some more conﬁguration beyond the scope of this talk.

Ryan R. Rosario


HadoopStreaming

Typically, Hadoop jobs are written in Java. This was a roadblock
to many developers that would incur time having to port code over
to Java.

Hadoop is distributed with Hadoop Streaming, which allows
map/reduce jobs to be written in any language as long as it can
read and write from stdin and stdout respectively.

This includes R. Cue the HadoopStreaming package!

Ryan R. Rosario


HadoopStreaming

One way to easily use Map/Reduce in R is with the
HadoopStreaming package. The workflow is as follows:
1 Open a connection: to a file, STDIN, or a pipe.
2 Write map and reduce functions either in one file, or in
separate files.
3 Run the job on data either from the command line using R
only, or with Hadoop.

Ryan R. Rosario


Data Mining with R

The data for this demonstration is a large 14GB ﬁle containing a
tweet ID and the content of a user’s tweet.

tweetID tweet_text

The typical “hello world” example for map/reduce is word
counting, so let’s construct the a list of words and the number of
times they appear over all tweets in the sample.

Ryan R. Rosario


HadoopStreaming: The Map Function

The map function takes a line of input, does something with it,
and outputs a key/value pair that is sorted and passed to the
reducer. For this example, I split the text of the tweet and output
each word as it appears, and the number 1 to denote that the word
was seen once.

The key is the word, and the value is the intermediate count (1 in
this case).

Ryan R. Rosario


HadoopStreaming: The Reduce Function

Before data enters the reduce phase, it is sorted on its key. The
reduce function takes a key/value pair and performs an aggregation
on each value associated with the same key and writes it to disk
(or HDFS with Hadoop) in some format speciﬁed by the user.

In the output, the “key” is a word and the “value” is the number
of times that word appeared in all tweets in the ﬁle.

Ryan R. Rosario


HadoopStreaming: Anatomy of a HadoopStreaming Job

First, we create a ﬁle that will contain the map and reduce
functions.
1 # ! / usr / bin / env Rscript # allows script to be
EXECUTABLE .
2
3 library ( Ha do o pS tr ea m in g ) # need the package .
4
5 # user can create own command line a r g u m e n t s .
6 # By default , certain a r g u m e n t s are already parsed for you
.
7 opts <- c ()
8 # Gets a r g u m e n t s from the e n v i r o n m e n t
9 op <- hsCmdLineArgs ( opts , op en Co n ne ct io n s = TRUE )

Ryan R. Rosario



When running an R script, we can pass command line arguments.
For example:

./mapReduce.R -m
Short Name Character Argument Type Description
mapper m None logical Runs the mapper.
reducer r None logical Runs the reducer.
infile i Required character Input file, otherwise STDIN.
outfile o Required character Output file, otherwise STDOUT.
Some selected arguments are above.

Ryan R. Rosario


op is populated with a bunch of command line arguments for you. If -m is passed to a
script, op$mapper is TRUE and the mapper is run.
10 if ( op $ mapper ) {
11 mapper <- function ( x ) {
12 # t o k e n i z e each tweet
13 words <- unlist ( strsplit (x , " [[: punct :][: space
:]]+ " ) )
14 words <- words [ ! ( words == ’ ’) ]
15 # Create a data frame with 1 column : the words .
16 df <- data . frame ( Word = words )
17 # Add a column called count , i n i t i a l i z e d to 1.
18 df [ , ’ Count ’] = 1
19 # Send this out to the console for the reducer .
20 hsWriteTable ( df [ , c ( ’ Word ’ , ’ Count ’) ] , file = op
$ outcon , sep = ’ , ’)
21 }
22 # Read a line from IN .
23 hsLineReader ( op $ incon , chunkSize = op $ chunksize , FUN =
mapper )
24 }
Ryan R. Rosario



op is populated with a bunch of command line arguments for you. If -r is passed to a
script, op$reducer is TRUE and the reducer is run.
25 else if ( op $ reducer ) {
26 # Define the reducer f u n c t i o n .
27 # It just prints the word , the sum of the counts
28 # s e p a r a t e d by comma .
29 reducer <- function ( d ) {
30 cat ( d [1 , ’ Word ’] , sum ( d $ Count ) , ’ n ’ , sep = ’ , ’)
31 }
32 # Define the column names and types for output .
33 cols = list ( Word = ’ ’ , Count =0)
34 hsTableReader ( op $ incon , cols , chunkSize = op $ chunkSize ,
skip =0 ,
35 sep = ’ , ’ , keyCol = ’ Word ’ , singleKey =T , ignoreKey =F ,
36 FUN = reducer )
37 }

Ryan R. Rosario



Finally, we clean after ourselves and close the connections to the in
and out connections.
38 if ( ! is . na ( opts $ infile ) ) {
39 close ( opt $ incon )
40 }
41
42 if ( ! is . na ( opt $ outfile ) ) {
43 close ( opt $ outcon )
44 }

Ryan R. Rosario


HadoopStreaming: Running a HadoopStreaming Job

HadoopStreaming allows the user to pass data to map/reduce
from the command line using a pipe, or using a ﬁle. To input data
using a pipe:
cat twitter.tsv | ./count.R -m | sort | ./count.R -r

To input data using a ﬁle:
./count.R -m -i twitter.tsv | sort | ./count.R -r

Ryan R. Rosario



Or we can run the script using, you know, Hadoop Streaming!
hadoop jar $HADOOP_HOME/contrib/streaming/hadoop-0.20.2-streaming.jar
-input /home/ryan/hdfs/in
-output ~/hdfs/out
-mapper "count.R -m"
-reducer "count.R -r"
-file ./count.R

Ryan R. Rosario


Hadoop produces a lot of status and progress output and provides
a web interface that you can explore when using it.

Ryan R. Rosario



Output looks as follows.

a,13
about,2
action,1
acutely,1
adapting,1
affairs,1
after,1
again,2
ah,2
Ah,1

Ryan R. Rosario


HadoopStreaming vs. Rhipe

Rhipe is another R interface for Hadoop that provides a more
“native” feel to it.
1 incorporates an rhlapply function similar to the standard
apply variants.
2 uses Google Protocol Buffers.
3 seems to have great flexibility in modifying Hadoop
parameters.
4 has more use cases and flexibility in how to run jobs and how
to transmit and receive data.
5 but... is more complicated to use and requires a more
in-depth knowledge of Hadoop, which is beyond the scope of
the group’s purpose.

Ryan R. Rosario


The Hadoop Barnyard
Hadoop also has several other projects that run on top of it:
Pig: data-ﬂow (query) language and execution framework for
parallel computing.
ZooKeeper: high-performance coordination service for
distributed applications.
Hive: data warehouse infrastructure providing summarization
and ad-hoc querying.
HBase: a scalable, distributed database supporting structured
data storage for large tables.
Avro: data serialization providing dynamic integration with
scripting languages.
Chukwa: data collection system for managing large
distributed systems.
Ryan R. Rosario


The Hadoop Barnyard

There are other packages that can interface with Hadoop, but are
part of a diﬀerent family.
Mahout: suite of scalable machine learning libraries.
Nutch: provides web search and crawling application software
on top of Lucene.
Lucene: an eﬃcient indexer for information retrieval.

Ryan R. Rosario

When to Use What?

This is my personal opinion based on experience with both R and
Hadoop.
For datasets with size in the range 10GB, bigmemory and ff
handle themselves well.
For “larger” datasets, use Hadoop (integrated with R?)
Not enough research on data on the scale of TB or PB in R.
Hadoop is superior here.

Ryan R. Rosario

Other Solutions: Hardware

There are other solutions to some of these problems.
Buy more RAM... lots of it... lots of fast RAM
Let your system run out of memory and conﬁgure it to use
solid state drives (SSD) for swap?
Use a very expensive SAS drive.
use a GPU.

Ryan R. Rosario

Other Solutions: Software

(use the packages we discussed today)
use databases that can be called from R.
if data is sparse, use sparse matrix packages sparseM or slam.
use a diﬀerent language like C/C++/FORTRAN and
interface with R.
use a diﬀerent language entirely.
Revolution R is beta testing big dataset support.
use Hadoop or Amazon EC2 or Elastic MapReduce with(out)
R
use SAS

Ryan R. Rosario

In Conclusion

1 R provides ways to deal with big data.
2 They are fairly easy to use.
3 Worth learning as R gains popularity and datasets grow huge
in size.

Ryan R. Rosario

Keep in Touch!

My email: ryan@stat.ucla.edu

My blog: http://www.bytemining.com

Follow me on Twitter: @datajunkie

Ryan R. Rosario

The End

Questions?

Ryan R. Rosario

Thank You!

Ryan R. Rosario

References

Most examples and descriptions can from package documentation.
Other references not previously mentioned:
1 “Big Data in R” by G Stat: March 24, 2010.
http://statistics.org.il/wp-content/uploads/2010/04/Big Memory%20V0.pdf
2 “The ff Package: Handling Large Data Sets in R with
Memory Mapped Pages of Binary Flat Files” by Adler,
Nenadic, Zucchini, Glaser: UseR! 2007.
http://user2007.org/program/presentations/adler.pdf
3 “The Bigmemory Project” by Michael Kane and John
Emerson: April 29, 2010.
http://cran.r-project.org/web/packages/bigmemory/vignettes/Overview.pdf
4 “Airline On-time Performance.” Data Expo 2009.
http://stat-computing.org/dataexpo/2009/the-data.html.

Ryan R. Rosario

Taking R to the Limit (High Performance Computing in R), Part 2 -- Large Datasets, LA R Users' Group 8/17/10

Recommended

Recommended

More Related Content

What's hot

What's hot (20)

Similar to Taking R to the Limit (High Performance Computing in R), Part 2 -- Large Datasets, LA R Users' Group 8/17/10

Similar to Taking R to the Limit (High Performance Computing in R), Part 2 -- Large Datasets, LA R Users' Group 8/17/10 (20)

Recently uploaded

Recently uploaded (20)

Taking R to the Limit (High Performance Computing in R), Part 2 -- Large Datasets, LA R Users' Group 8/17/10