Cloud computing systems today, whether open-source or used inside companies, are built using a common set of core techniques, algorithms, and design philosophies – all centered around distributed systems. Learn about such fundamental distributed computing "concepts" for cloud computing. Some of these concepts include: clouds, MapReduce, key-value/NoSQL stores, classical distributed algorithms, widely-used distributed algorithms, scalability, trending areas, and much, much more! Know how these systems work from the inside out. Get your hands dirty using these concepts with provided homework exercises. In the programming assignments, implement some of these concepts in template code (programs) provided in the C++ programming language. Prior experience with C++ is required. The course also features interviews with leading researchers and managers, from both industry and academia.
3.2. MapReduce Examples
Loading...
From the course by University of Illinois at Urbana-Champaign
Cloud Computing Concepts, Part 1
548 ratings
University of Illinois at Urbana-Champaign
548 ratings
Course 1 of 6 in the Specialization Cloud Computing
From the lesson
Week 1: Orientation, Introduction to Clouds, MapReduce
This course is oriented towards learners with similar backgrounds as juniors and seniors in a CS undergraduate curriculum. Since learners come from various backgrounds, it is critical you view this lecture AND pass the prerequisite test. This will ensure you have many of the assumed prerequisite pieces of knowledge required to enjoy this course.
Meet the Instructors
Indranil GuptaProfessor
Department of Computer Science
In this lecture,
uh, we're going to see some examples
of, uh, applications that use MapReduce.
We'll work with, uh, fairly simple examples,
but these will, these examples will hopefully show you,
uh, the simplicity of programming in MapReduce,
and yet, at the same time, how powerful the paradigm can be.
So the first application is a distributed grep.
You have a large set of files
with a large amount of text in them,
for instance, uh, these might be the works of Shakespeare,
or in a production cluster
these might be, uh, the logs of, uh, uh, applications
that are running,
and you want to grep for, uh, particular words,
say of characters in Shakespeare's works,
uh, or, uh, for error messages, uh,
in, uh, your, uh, production cluster.
So essentially, you may have a pattern
which might be a regular, uh, expression pattern
or just a-a word or a series of words,
and you want to output all the lines that match that pattern.
So here's what, uh, the Map would look like;
um, here's how you would write the Map.
The Map essentially would, uh, look at each line of text
and, uh, check if that line of text
matches the supplied pattern.
It might use regular grep for this, and if it does,
then it emits that line of text as the key,
with the value being, say, one, it doesn't really matter.
The Reduce in turn, uh, simply copies the intermediate data
to the output.
It does not do any processing,
because each line, uh,
essentially, uh, might be unique,
and you don't want to really do
any merging of lines in this case.
And the output of, uh, this particular MapReduce uh, uh,
implementation is essentially the set of lines
that match your pattern.
The advantage of doing distributed grep here
is that since you have a very large data set,
you could have multiple Map tasks and multiple Reduce tasks
which speed up the processing of your, uh, distributed grep,
instead of doing a lot of the grep on one machine,
which might be very, very slow, uh,
because you have to sequentially process all the data,
and also you may need to transfer the data
to that single machine.
With MapReduce you could run your, uh, application MapReduce
distributed grep application even though your data
is distributed out over multiple servers.
The next application is a reverse web-link graph.
So you have a graph, uh,
which is, say, a portion of the web graph,
where you have, uh, edges (a, b)
where a is a webpage and b is a webpage,
and a has a link that points to b.
This is represented, uh, in the input as tuples (a, b),
and that's what your input consists of.
It is just pairs of URLs, uh, (a, b)
where ais a URL and b is a URL
where the webpage at URL a points to the webpage at URL b.
The output you want is the reverse.
You want for each page,
uh, all the list-the list of all the pages that point to it,
okay, so you want to find out all the pages that point to,
say, your homepage.
So here's what the Map does.
The Map processes the web log,
and for each input pair (source, target),
(a, b) that it receives, it just reverses it.
It makes the target as the key and the source as the value.
Why does it do this?
Well it does this so that in the Reduce phase,
all of, uh, the key/value pairs that share this target
as the key will end up at the same Reduce task,
and that Reduce task in turn
can simply list the corresponding sources.
It doesn't need to add them,
because you can't really add URLs.
Uh, it just lists them, makes a list out of the sources,
and outputs that as the key/value pair,
or the final key/value pair.
So this, ah, is essentially the output where for each target
and key/value pair the corresponding value is the list
of all the URLs that, um, have pointed to it.
Now you could take this a step further where, uh,
instead of, um, uh, listing,
you could simply add up the number of URLs.
Okay, in this case, um, you would output, uh,
the number of pages that, uh, point to the target.
In that case, you could also optimize the map, uh,
so that the, the, uh, output of it is (target, 1).
So that kind of dovetails into our next,
where, uh, you have, uh, our next application
where you have a, um, uh, a say a log,
which is a, uh, log of accessed URLs,
say from a proxy server,
and there you have, uh, one URL for, uh, for every webpage
that was fetched by any client or user inside,
for instance, your company or organization.
And you want as output for each URL
the percentage of total accesses for that URL, okay,
so you want to say, for instance,
CNN.com accounted for 13.5% of all accesses, uh,
from this particular company,
and so on and so forth for every URL.
So here's how the Map and the Reduce would work.
The Map processes the web log.
Remember, the web log has a series of URLs,
and for each URL that it encounters,
it outputs the key/value pair (URL, 1).
The Reducers essentially, uh, do the same thing
as we did for the WordCount
where each URL is assigned to exactly one Reducer,
and, uh, for that URL it sums up the, uh, values
that it, uh, receives as, um, values associated with that key.
So at this point we have output
from, uh, the MapReduce function,
where for each URL you have a count
of the number of times that URL was accessed.
Okay, this is essentially the WordCount example,
but this is not what we're looking for.
We're looking for a percentage rather than a count.
So what you do is you chain another, uh, MapReduce
right after this.
So you have the Map phase that leads to the Reduce phase,
and then after the Reduce phase you chain another Map phase
and another Reduce phase after that,
which will do the percentage part.
So the Map receives-receives as input
the output of the previous Reduce phase,
which is the (URL, URL_count) pairs,
uh, and it outputs (1, the value) as its output.
Why does it do this?
Well, you just wanna have one key
so that you have exactly one Reducer
in that second MapReduce phase,
and it simply takes, uh,
all of these key value pairs it has received,
it extracts the URL count, it sums them up.
Once it has summed them up,
it knows the total number of URL counts,
uh, that are there across all the URLs,
and it makes wa-one more pass, uh,
through this, uh, uh, data set,
and it divides the URL count by the overall count,
then it is calculated and then it emits
(URL, URL_count/overall_count)
for each URL that is present in its inputs.
So this second reducer, uh, second-phase Reducer,
there is only one of it,
because, um, you want to calculate
the overall count, right?
You can't calculate the overall count
if there are multiple Reducers,
and essentially the Reducers and the maps
don't talk with each other, uh, in this particular case,
and you won't have exactly one Reducer.
And the Reducer does two passes;
one pass to calculate the overall count
by summing up all the URL counts,
and the second pass to calculate the percentages
by dividing all the counts by the overall count.
The last example is, uh, one of sorting,
and you might think sorting might be fairly complicated, um,
but in fact, uh, the internal MapReduce,
um, uh, technique, uh,
internal MapReduce, uh, engine, uh,
the Hadoop engine as well, uh,
does some more of, uh, sorting already,
and this helps us right sort very, very easily.
Concretely, the output of the Map task,
remember the output of each Map task
is a set of, uh, key/value pairs,
this set of key/value pairs is already sorted
before it is sent out to the Reduces.
This sorting is typically a quick sort implementation
in the Apache Hadoop implementation.
For instance, the keys might be integers or they might be words,
and these are sorted uh, uh,
based on increasing order of integers
or lexicographically using the words.
The Reduce task, a given Reduce task,
also receives a-a lot of input for multiple key/value pairs.
It also sorts all of this input before it starts its processing.
This sorting is actually fairly important
for, uh, the Reduce task,
because, uh, then, uh, the Reduce task can decide, uh,
basically once you sort the key/value pairs by key, uh,
all the sorting here is done by key, uh,
once you sort them by key,
uh, essentially what happens
is that all the key/value pairs associated with a given key
are present contiguously in your sorted, uh, set,
and so you can call, uh, the Reduce function
once for that entire batch of, um, uh, inputs,
Reduce inputs that share the same key.
So, uh, in order to sort
a fairly large data set in MapReduce,
um, the input receives a series of key/value pairs,
and for the output
you want to, uh, output the set of sorted values.
The Map function, uh, that we would write
would take the key/value pairs
and it would simply output value comma whatever
as the key/value pair.
We don't really care
what the value here is for the output key/value pair.
Essentially this is the identity function;
it does not do any processing.
Similarly, the Reduce also does no processing;
it just outputs the same key/value pair.
You might wonder "How does this work?"
Well, it works because the Map output is already sorted
using Quicksort,
and the Reduce output is also sorted using Mergesort,
and so the output of, uh, the Reducers,
as long as the Reducers are also sorted
with respect to each other
and you assign the file names
for each of the Reducer outputs appropriately,
uh, you will get output that is also, uh, sorted, um,
based on the values.
The only concern here is one of partitioning.
You cannot use hash-based partitioning over here.
You want to make sure that the Reducers themselves
are also ordered so that their output-
corresponding output file names can be given increasing
um, um, uh, file names, uh, in increasing order,
and they can be collated appropriately.
So, uh, you want to partition keys across Reducers
based on ranges.
In other words, you want to assign each Reducer
a range of keys,
say Reducer number one gets, uh, uh, keys 0 through 1,000, uh,
Reducer number two gets, uh, keys 1,001 through 2,000,
and Reducer number three gets keys 2,001 to-through 3,000
and so on and so forth.
So, um, this again goes into the partitioning function.
Ah, in addition,
the, uh, values may be, uh, non-uniformly distributed,
so there might be many more keys
in, uh, the range say 2,001 through 3,000,
so you may want to assign more Reducers from that range.
Okay, and so you want to-you may want
to take the data distribution into account
in order to assign, uh, while assigning ranges, uh,
to the keys, uh, to the, uh, Reduce tasks.
Explore our Catalog
Join for free and get personalized recommendations, updates and offers.
Coursera provides universal access to the world’s best education,
partnering with top universities and organizations to offer courses online.
© 2018 Coursera Inc. All rights reserved.
