Big Data Computing

Prof. Rajiv Misra

Computer Science andEngineering,IIT Patna

Lecture 09
Parallel programming with Spark
Parallel programming with the Spark.

Refer slide time :( 0:17)

Preface: Content of this lecture. In this lecture we will discuss overview of Spark, fundamental of Scala
and functional programming, Spark concepts, Spark operations and the Job execution.

Refer slide time :( 0:30)

Introduction to Spark.

Refer slide time :( 0:32)

What is this Spark? It is fast expressive cluster computing systems, which is compatible to Apache
Hadoop, now this particular Spark system works, with any Hadoop supported storage system such as
HDFS, S3, sequential file. And so on which improves, the efficiency through in-memory computational,
primitives and general computational graph. So, it also improves the usability through rich collection of
API, is in the form of scale a Java Python, it has the interactive cell. So, this all comprises of the Spark,
scenario. So, using this in memory computation, it is 100 times faster, compared to the previous
generation MapReduce systems and also with the interactive, shell it has reduced often (2-10 x) less code
here in this particular system.
Refer slide time :( 01:44)

So, how to run it basically using local multi-core system or using with the, with a private cluster using
Mesos, YARN and standalone mode

Refer slide time :( 01:57)

So, Spark originally was written in Scala, which allows concise function that is syntax and interactive
use, there are APIs, which are available for Java and Scala and Python now. So, interactive shells are
available in a Scala and Python.

Refer slide time :( 02:18)

Now let us introduce, to the Scala functional programming language.
Refer slide time :( 02:27)

So, it is a high-level language, for the java virtual machine, that is it can compile through the java virtual
machine, byte code and it is statically, typed that is and then it is interoperates with the Java.

Refer slide time :( 02:46)

Declaring variables:
var x: Int = 7
var x = 7 // type inferred
val y = “hi” // read-only

Functions:
def square(x: Int): Int = x*x
def square(x: Int): Int = {
x*x
}
def announce(text: String) {
println(text)
}

And so, here we are going to see the quick tour, of functional programming language, that is the Scala. In
Scala, the variables are defined using where function, where there are two ways, you can define the
variable one is by specifying the type of the variable:
Refer slide time :( 05:41)

is the generic types.

Generic types:
var arr = new Array[Int](8)

var lst = List(1, 2, 3)

// type of lst is List[Int]

Indexing:
arr(5) = 7
println(lst(5))

Refer slide time :( 06:33)

Now processing collection with the functional programming,
Processing collections with functional programming:
val list = List(1, 2, 3)
list.foreach(x => println(x)) // prints 1, 2, 3
list.foreach(println) // same
list.map(x => x + 2) // => List(3, 4, 5)
list.map(_ + 2) // same, with placeholder notation
list.filter(x => x % 2 == 1) // => List(1, 3)
list.filter(_ % 2 == 1) // => List(1, 3)
list.reduce((x, y) => x + y) // => 6
list.reduce(_ + _) // => 6

Refer slide time :( 10:25)

Here in this case now let us see the Scala closure syntax more of it.

(x: Int) => x + 2 // full version

x => x + 2 // type inferred
_ + 2 // when each argument is used exactly once
x => { // when body is a block of code
val numberToAdd = 2
x + numberToAdd
}
// If closure is too long, can always pass a function
def addTwo(x: Int): Int = x + 2

list.map(addTwo)

Refer slide time :( 11:54)

 Inside another function, there similarly there are other collection, methods which are available in the
Scala. So, a Scala collection provides many other functional methods, for example as Google for a Scala
sequence we can see on it. So, map function is can be applied and this will pass each element, through a
function f and similarly flat map one-to-many map function, it will apply similarly for the filter it will
keep the element passing f and exist means a true if one element passes, for all reduce group by and sort
by all different methods are available.

Refer slide time :( 12:35)

Now let us see the Spark concepts.

Refer slide time :( 12:40)

So, here the goal of Spark is to provide a distributed, collection and here the concept of Spark ,is to
support, this distributed computation in the form of resilient, distributed data sets and RDDs are
immutable collection of objects which are spread across the clusters and these RDDs they are built
through a transformation, such as map filter etcetera and these particular, RDDs they are automatically
rebuilt, on the failure, that means a lineage is automatically generated and whenever there is a failure it
will be reconstructed, similarly it is also controllable, persistence that is the cache.

Refer slide time :( 13:29)

 In the rhyme is being done. So, the main primitives is RDDs which are immutable partition collection, of
objects various transformations are applied on the RDDs such as, map filter and group by join and these
are all lazy operations to build RDDs from, other RDDs and besides transformations, actions also can be
defined on our RDD such as count collect, save and it will return the values or it will write it on the under
disk.

Refer slide time :( 13:58)

 Now let us see one example of, of mining the console logs, using the Scala program written, in Spark and
here, it what it will do it will load the error messages, from a log into the memory and then interactively
search for a pattern. So, let us see that, the Spark program now you, you, you know that it runs in a form
of a master that is the driver and the worker, system and the first line written is to read this particular log
files, that is in the from the HDFS, it will read the text file and this particular lines after reading it, it will
become in memory. So, it will generate the, the base RDD after reading the file and then what it does is
apply the filter, on the lines which are now red, into the in memory, it will apply the filter operation that
we have seen, on the lines and in this filter operation, what it does is it will identify the, the, the errors
which are appearing, in the in the in the log files and this will perform a transformation, after filtering and
they are collected in the form of, another RDD that is errors RDD and using this errors, RDD down we
can apply, the map function and which says that using this map function. Now you can split this, s to the
wherever the taps are there and they will be now dividing, these into the different error messages and
these error messages are now cached.
So, apply the filter on these error messages that wherever this, foo is available or appearing, in the
message and you have to make a count, of it how many such foo messages are appearing, in this error
message. So, here the count is the action and before that all were transformations. So, let us see how it
will be done. So, wherever these messages are restoring on that, this ‘foo’ will be counted so, just see that
the driver will perform a task, to count. So, task is to count, this particular task will be communicated by
the driver to all the workers and they will count and then return the result back to the driver and also will
be in the cache, these messages will be in the cache. Now the next time when you want to filter the, the
particular string that is called a, ‘bar’, from, the message then, it will be performed in the cache itself it
will not go and look up into the into, the disk it will not be done in the disk, it will be now returned back
from the from the cache itself. So, the results will be quickly returned back because so, therefore the full,
text search of a Wikipedia can be done, in a less than one second if it is in the cache or, or it is 20 second
if it is done through the disk access. So, therefore if the data is scaled to one terabyte, then it will take five
to seven seconds, if it is to be accessed through the cache or it is 170 seconds if it is on disk data
.
Refer slide time :( 17:49)

So, in all cases this entire big corpus of terabyte of data can be processed very efficiently and in a very
quickly. Now let us see the RDDs fault-tolerance. So, RDDs will track the transformations, used to build
them through the lineage to recompute the last data. So, here we can see that, once we specify these filter
that is using, filter which contains the error and then it will split, all those messages, which are tab
separated and this will be collected in the form of messages. So, let us see that, this particular filtered
RDDs, will contain the information and they are now stored in the HDFS file system. So, RDDs keep
track of the transformations, used to build them, their lineage to recompute the last data.

Refer slide time :( 18:55)

So, fault recovery, test if we see here is that so, the iteration time and the failure happens, is recovery is
quite efficiently and done.

Refer slide time :( 19:09)

Now behavior with less RAM, can see that if it is fully cached, then it will iteration time will be quite
less here in this case.
Refer slide time :( 19:18)

Now the question is what language you can use, obviously a scalar will be performing, better one let us
see the two of Spark further operations. So, easiest way, to use the Spark is by the interpreter over, the
shell and it runs in a local mode with, only one thread by default what control, can be with the master one
and cluster. So, the first stop is through, the to the Spark context, is the main entry point to the Spark
functionality which is created through the Spark shell.
Refer slide time :( 19:18)
And let us see how to create, this entire operation:

# Turn a local collection into an RDD

sc.parallelize([1, 2, 3])
# Load text file from local FS, HDFS, or S3
sc.textFile(“file.txt”)
sc.textFile(“directory/*.txt”)
sc.textFile(“hdfs://namenode:9000/path/file”)
# Use any existing Hadoop InputFormat
sc.hadoopFile(keyClass, valClass, inputFmt, conf)
Refer slide time :( 20:37)
The data and creating the RDDs now the basic transformations are shown over here:

nums = sc.parallelize([1, 2, 3])

# Pass each element through a function
squares = nums.map(lambda x: x*x) # => {1, 4, 9}
# Keep elements passing a predicate
even = squares.filter(lambda x: x % 2 == 0) # => {4}
# Map each element to zero or more others
nums.flatMap(lambda x: range(0, x)) # => {0, 0, 1, 0, 1, 2}

Refer slide time :( 23:45)

And this is the basic transformation and now we are going to see some of the actions,

nums = sc.parallelize([1, 2, 3])

# Retrieve RDD contents as a local collection
nums.collect() # => [1, 2, 3]
# Return first K elements
nums.take(2) # => [1, 2]
# Count number of elements
nums.count() # => 3
# Merge elements with an associative function
nums.reduce(lambda x, y: x + y) # => 6
# Write elements to a text file
nums.saveAsTextFile(“hdfs://file.txt”)

Refer slide time :( 25:29)

And now let us see how to work with key value pairs, now is part distributed reduced transformation x,
on RDDs of the key value pairs and key value pairs means here,

Python: pair = (a, b)

pair[0] # => a

pair[1] # => b

Scala: val pair = (a, b)

pair._1 // => a

pair._2 // => b

Java: Tuple2 pair = new Tuple2(a, b); // class scala.Tuple2 pair._1 // => a

pair._2 // => b
Refer slide time :( 25:59)
Now let us see some more key value pair operations, for example

pets = sc.parallelize([(“cat”, 1), (“dog”, 1), (“cat”, 2)])

pets.reduceByKey(lambda x, y: x + y)
# => {(cat, 3), (dog, 1)}
pets.groupByKey()
# => {(cat, Seq(1, 2)), (dog, Seq(1)}
pets.sortByKey()
# => {(cat, 1), (cat, 2), (dog, 1)}
reduceByKey also automatically implements combiners on the map side
Refer slide time :( 27:31)
Now let us see the full use of full execution of, the word count program. So, what count program means

lines = sc.textFile(“hamlet.txt”)
counts = lines.flatMap(lambda line: line.split(“ ”)) \
.map(lambda word: (word, 1)) \
.reduceByKey(lambda x, y: x + y)

So, these words means map function, will what it will do it will for every word it will generate what
comma 1. So, 2 it will generate 2 comma 1 B, B comma 1 or, or 1 not comma 1 and 2 comma 1 and B
comma 1. So, after having generated this map function, now it will reduce by the key. So, reduced by the
key in the sense it will say that for, a particular key it will do the summation. So, for the same keys for
example B, is the key which is appearing in this particular, map output and here also this is the output,
they will be collected back and that their values will be aggregated, according to the plus sign. So, 2 will
be there similarly, 2 is also appearing two times and 2 will be gathered over here and their values also will
be, added up and whereas all others are appearing only once. So, not will be collected over there and R
will be connected over there. So, this is the reduced function, which will be applied.

Refer slide time :( 29:41)

There are other key value operations and now we can see about these operations, which basically are
mentioned, were here like, like join and cogroup operations.

Refer slide time :( 30:01)

And now, we can all the pair's RDDs operations take the optional second parameter for the number of
tasks and this we have shown over here.

Refer slide time :( 30:13)

And external variables you can you use in the closure will automatically be shipped to the to the cluster.

Refer slide time :( 30:24)

So, all these are the implementation issues.

Refer slide time :( 30:27)

Refer slide time :( 30:29)

So, for other RDD operations the programming guide is there.

Refer slide time :( 30:27)

 And that can be seen now let us see about, the job execution in his Spark it has various software
components and when you say Spark context, will be creating it will be created, in the master job master
and then it will create the worker threads, in the form of power context and then executors will execute
them.

Refer slide time :( 30:56)

Similarly there will be at a scheduler Task, you will support the general tasks graphs, internally and
pipelines will be also pipeline functions we are possible, they will be created by the task scheduler and
cache aware data reuse and locality and partition aware things are done to avoid the shuffles.
Refer slide time :( 31:21)

So, more information about resources, on Scala is available.

Refer slide time :( 31:28)

And let the Hadoop is Spark can read and write to any storage system format that has plug into the
Hadoop. And API is like a Spark on text file supports, while the Spark context Hadoop RDD allows pass
any Hadoop job, to configure the input file.

Refer slide time :( 31:45)

 And this is to create the Spark context, now Spark context when we say it has different arguments. So,
the first one is called, ‘Master URL’, is nothing but to create the URL or a local oblique in local node n.
so, next one next thing is the application name and then, Spark will now install the path on the cluster,
with the name is Spark home and finally the list of jar files also has to be given in the Spark context
automatically it creates a Spark context using the Java jar file.

Refer slide time :( 32:25)

So, now this is the complete word count program, which we have discussed, shown over here as the local
and what count is the name of the program arguments are there and this argument number one.
Refer slide time :( 32:40)

And

Refer slide time :( 32:41)

Now let us see an example

Refer slide time :( 32:43)

Of a PageRank algorithm how we can execute in the
Refer slide time :( 32:48)

Refer slide time :( 32:48)

Spark system.
Refer slide time :( 32:51)

So, let us start let us see the algorithm,

1. Start at each page the rank of one.

2. On each iteration, have page p contribute rankp / |neighborsp| to its neighbors
3. Set each page’s rank to 0.15 + 0.85 × contribs

So, here we can see that, here we can see that, this particular node, is basically this particular node has,
two links out so, the contribution of one will be divided equally 0.5 and 0.5 similarly this page also has
two outgoing lines links so, it will be also having a contributions of 0.5 and this will have only one. So, it
will be having the contribution of entire 1 and this page is also having 1. So, it will be having contribution
of 1, now let us see that this particular page, we can see that it is now incoming. So, it is incoming with
0.5 so, now we have to calculate according to this 0.15 + 0.85 x 0.5. So, that will be the new page rank of
this and as far as this particular page, is concerned the incoming is once 0.85 x 1 + 0.15. so, the page rank
will become 1 in that case so, 1 will not be changed so, this will be the page rank 1 and how about this so,
here 1 2 3 different links are coming so, with a with this, this link will be 1 plus this will be 1 and this will
be 0.5 and this also will be the this also will be 0.5 so, this become 2 x 0.85 + 0.15. And this also I will
have the same so, let us see that after doing this particular iterations. So, the PageRank will now be
changed, into as we have seen that it will be 1 it was pointed point, five eight point five eight point eight
five one point eight five and this particular, iterations will continue and here it will stop ,why because it is
not going to change.
Refer slide time :( 35:52)

Further let us see how this entire PageRank algorithm can be implemented using scalar.

val links = // RDD of (url, neighbors) pairs

var ranks = // RDD of (url, rank) pairs
for (i <- 1 to ITERATIONS) {
val contribs = links.join(ranks).flatMap {
case (url, (links, rank)) =>
links.map(dest => (dest, rank/links.size))
}
ranks = contribs.reduceByKey(_ + _)
.mapValues(0.15 + 0.85 * _)
}

ranks.saveAsTextFile(...)
Refer slide time :( 37:23)

So, we see that the page rank performance, with the Spark it is very efficient and very fast compared to
The Hadoop here, if the number of machines are 16 and the iteration time is very less,

Refer slide time :( 37:36)

There are other iterative algorithms which are implemented, in the SPARC such as K-means clustering
logistic regression in all the cases you see that, Spark is very very efficient compared to the Hadoop.
Refer slide time :( 37:53)

In the iterations so, these are some of the references, for the Spark to be.
Refer slide time :( 37:59)

So, conclusion is, Spark offers a rich API is, to make the data analytics fast. Both fast to write and fast to
run. It achieves hundred times speed-up, in the real applications, the growing community with 14
companies are contributing to it and details tutorials are available in the website, www.spark-project.org.
Thank you.