High Performance Computer Architecture

Prof. Ajit Pal

Department of Computer Science and Engineering
Indian Institute of Technology, Kharagpur
Lecture - 41
Cluster, Grid and Cloud Computing
Hello and welcome to this last lecture of this lecture series, on High Performance
Computer Architecture.
(Refer Slide Time: 00:42)

In this lecture I shall consider the topic Cluster, Grid and Cloud Computing, so these are
the three topics which are little inter related, but definitely they are not same. And I shall
consider them one after the other, and as we copper different topics, we shall high light
their differences.

(Refer Slide Time: 01:08)

First let as look at the motivation behind cluster computing, we will find there are many
applications that require high performance computing. So, numerous applications are
now-a-days available are present, which requires high performance computing, and that
can be satisfied with the help of cluster computing. And some examples of these high
performance computing requirement, I mean which requires high performance
computing is given here.
Number 1 is numerous scientific and engineering applications, like modeling, simulation
and analysis of complex systems like climate, galaxies, molecular structures, nuclear
explosion and whether it is forecasting and so on. And also there are applications which
requires high performance computing in business, and internet applications such as Ecommerce and web servers, file servers, databases and so on. And for these applications,
if you want to have a very dedicated computer, custom made computer with custom
software, it is pretty costly. And more over this custom hardware and software does not
allow, and not only they are very expensive, they are not extensible.
That means, they can be extended as the requirement goes off, so I mean some kind of
super computers are needed to satisfy this high performance computer requirement, but
unfortunately they are not extendable. And the cost effective approach is to use cluster,
grid and cloud computing for these high performance computing, to satisfy this

computing requirement, and as you shall see how these are platform in this three
situations.
(Refer Slide Time: 03:19)

And as I have already mentioned examples applications of cluster computing, and which
has I mean, you may be asking in few situation you will use cluster computing, these are
the several situations like, where you require large run times, real time constants that we
satisfied, large memory usage is there. Then you require high I O usage, then you require
fault tolerance and also you want to have high availability, in such situations we will go
for cluster computing.

(Refer Slide Time: 03:57)

So, as I mentioned high performance computing is the main motivator, and it is an

alternative two symmetric multiprocessing to provide high performance and availability.
And as a consequence clustering has become a hottest topic in computer system design,
you may be asking how do you cluster computing, actually it can defined as coordinated
use of interconnected autonomous computers in a machine rooms.
That means, your interested to perform the computing with the help of a interconnected
array of computers inside a room, single room, it is not that they are dispersed
throughout the country or throughout the large geographical area, but in a machine room.
And you notice another term autonomous, by autonomous mean each and every
computer is a complete system, can work independently, without the need of others that
is you call it cluster computing. So, you can say collection of standalone workstations of
PCs that are interconnected by high speed network, may be a local area network, and
work as an integrated collection of sources.
So, you find in addition to this the important feature is you have a large number of
systems, but at the same time they work in a integrated manner. And they provide a
single system image, they have a single system image spanning all it is nodes, so this is a
very important concept you will require a special software layer to provide these as usual
see.

(Refer Slide Time: 06:01)

Actually the cluster computing has become popular, because of two reasons; number one
is nowadays you can have personal computers, which are pretty powerful at a very
affordable price. Similarly, you can have computer networks like switches, different
types of switches, which also available at very low price, so these two is the main
motivation. And so that means, it is now possible to connect cluster of workstations with
latencies and bandwidths comparable to tightly coupled machines, because of the
advancement of your networking technologies.
So, we can build a system by using commodity of the shelf components, what are the
commodity of the shelf components, we shall be using standard PCs or workstations and
we shall be using standard networking components. So, clusters are started to take off in
90's, clusters IBM, Sun, DEC workstations are connected by Ethernet, 100 Megabyte
Ethernet LAN, HP clusters and so on. So, you can see we are using standard local area
network to link the different nodes.

(Refer Slide Time: 07:33)

And as I have already mentioned, although you have got several interconnected
computers, but it gives single system is a image. And it gives you high performance, and
it is very inexpensive or you can say this is low cost alternative to inexpensive
computers. So, a single system image makes a cluster appear like a single machine to the
user.
So, you can say if you look at the different components of a cluster, it consists of
standalone machines with storage a first interconnection network, high speed LAN, low
latency communication protocols. So, not only we will be using hardware, I mean fast
interconnected network, you have to use suitable software which will provide you low
latency, for communication. And software to give system image that is known as cluster
middleware and of course, you will require a host of programming tool to utilize a
cluster.

(Refer Slide Time: 08:47)

And there are possible different causal possible configurations, one is passive standby to
make it reliable or to make it more available. And active secondary, then you can have
separate servers, you can have servers connected to disks, servers with share disk, as it is
shown in this diagram.
(Refer Slide Time: 09:11)

Here you have got two computers shown here, I mean here you have got one you can
say, number of processors connected to a shared bus and this is connected to another
system, which is also I mean multiple processor and which shared memory. So, these

two are shared memory multiprocessors, these two are connected together with a high
speed messaging link to work as a cluster.
(Refer Slide Time: 09:51)

This is another configuration, in which you can have not only high speed messaging link,
but you may have a shared hard disk, so in the form of redundant array of in expensive or
independent disks. So, here a raid is shared by these two systems through I O buses, so
this is stand by server with shared disk.
(Refer Slide Time: 10:15)

And for a cluster you must be having a operating system, and these are the various
design issues in the context of clusters, first offal you have to take care of failure
management. So, whenever you are behaving a large number of computers, some of
them may fail, so it will tolerate failure, then it will be cohere form failure. So, you will
require software for operating system will take care of failure management, then you
have to do load balancing, the load has to be uniform distributed among all the nodes in a
cluster.
Then you will require parallelizing computation, because you will be doing a kind of
parallel processing with the help of the number of commuters. So, you will require
parallelizing compiler parallelizing application which is amenable to application, which
are amenable to parallelism then parametric computation. So, these are the operating
system design issues.
(Refer Slide Time: 11:23)

In the context of clusters, then you can have here are the two basic categories of type one
can be non dedicated clusters, so they are essentially you may say signature. So, for
example, a network of workstations use spare computation cycles of nodes, and
background job distribution is done individual owners of workstations. So, this is non
dedicated, then you can have dedicated clusters with joint ownership having in dedicated
nodes, and which will also allow parallel computing.

(Refer Slide Time: 12:04)

Then you have got, you can have homogenous cluster another way of classifying is
homogenous and heterogeneous, so homogeneous cluster will have similar processors.
So, all the nodes will be identical or similar software operating system, and software
present in all the components, all the nodes in cluster. Alternatively you can have
heterogeneous system with different architecture, data format, computational speed,
system software that means, operating system and so on.
(Refer Slide Time: 12:35)

So, let us see what are the advantages and disadvantages of the clusters obviously, main
advantages is gives you high availability, and resilient failures. That is since, you have
got large number of computers, if one or two fails still it will continue to function and as
a consequence it gives you high availability. And then second important feature is
incremental extensibility, so as your requirement keeps on increasing with time, you
keep on adding nodes in a cluster. So, you can increase the number of nodes in a cluster,
in a incremental manner.
And desk tops are cheap and ubiquitous, so this is another important advantage, so you
are able to provide high performance computing at a very low cost, no need to by
dedicated expensive hardware. So, you can use commodity of self type of desk tops for
your application for developing a cluster, so these are the advantageous. Of course, you
have got several disadvantages, number one is administration complexity, so since you
have got you are administering N node cluster, I mean administering N node cluster is
close to administering N big machines.
The reason for that is each machine is autonomous with hardware, operating system,
application software and so on, so as a consequence each of these nodes are be
administered independently and separately. So, administering N node multiprocessors is
close to administering one big machine, so whenever you have got multiprocessor
systems, shared memory multiprocessor systems, they are you can have only it may be
considered as a one big machine.
But, that is not true in the context of clusters, so and also it gives you higher cost of
ownership and second important disadvantage is, these computers are connected through
I O bus. So, in case of we have seen shared memory multiprocessor system or massively
parallel processing system, the processors are connected through memory box. So,
memory box has got higher bandwidth and smaller latency, on the other hand whenever
they are connected using I O bus, the bandwidth is smaller and latency is larger.
So, lower bandwidth and higher latency you can say, so higher part is missing here, so it
will higher latency, so this is the disadvantage. And another important disadvantage is N
machine cluster have N independent memories and N copies of operating systems as I
mentioned earlier. So, since each is autonomous will be having independent memories,

so that they can work independently and N copies of operating system, so because each
can again operate independently without need of others.
(Refer Slide Time: 15:50)

And here is another advantage large computers of small volumes, so whenever you are
building a large computers, then cost has to be amortize over few systems. So, these
results in higher cost, on the other hand in case of clustering, you have got large number
of low cost system as a result the cost is much smaller. So, administrated complexity can
be mitigated, construction of shared memory multiprocessors and keeps storage outside
the clusters. So, you can keep the storage outside the clusters, as it happens in case of
storage area network.

(Refer Slide Time: 16:41)

So, these are gives you some components which are use like, the local area network that
is been use, you can use high fast Ethernet or Gigabit Ethernet or Myrinet, so these ate
type networks that can be used for connecting or InfiniBand, which gives you 100
Megabytes per second. So, you can see these are the rates 11.2 high Megabits per second
or 110 Megabits per second or 200 Megabits per second or 800 Megabits per second, so
this speed is quite high and distance maximum.
Of course, distance is limited you can it has to be within 100 meters that is why I said
that it has to be within a room, but it did not be really within a room, so 100 meter where
is quite long distance and you can connect about 100 nodes. So, storage area network
tries to optimize network of performance for short distances for example, as it happens in
case of InfiniBand. So, you have much slower protocol over head and much less security
concerned, as it happens whenever you do the communication through internet.

(Refer Slide Time: 17:55)

So, let us have a look at the taxonomy of a clusters, one is network of workstations
example is Beowulf clusters, where it use COTS PCs that we commonly, I mean
commodity of the self PCs with storage network. And you can have cluster forms
existing PCs on a LAN, which can hardly can perform work or super cluster or
constellation, where you can have cluster of cluster that is within a campus. So, these are
possible ways in which you can middle clusters.
(Refer Slide Time: 18:44)

So, these are some of the examples, the club diagram shows uniprocessor clusters, each
processor here you have got a single processor with it is own dedicated memory, and I O.
So, you have got a large number of such processors connected through a Ethernet switch,
1 GP Ethernet switch, here you have two way symmetric multiprocessor cluster, because
each node is having two processors with a shared memory and I O. And these are
connected through 1 GP Ethernet switch and the third diagram that is shown is eight way
SMP cluster.
Here you have got eight processors in each node with a shared memory and I O and
switches connected with the help of again 1 Gigabit Ethernet switch and of course, may
be connected to internet. So, these are the three different types of cluster shown here,
they are shown in the processor nodes to a SMP nodes, are eight way SMP nodes.
(Refer Slide Time: 19:57)

And this is how the connections can be done flat neighborhood networks, this is how the
interconnection can be made and this particular topology is known as Ethernet or flat
neighborhood networks. So, this diagram shows the layout of 25 node FNN, a Flat
Neighborhood Networks, here as you can see you have got 24 nodes, 1 2 up to 23, 0 to
23, 24 nodes each having processors memory and I O devices. And these are connected
with the help of 16 port Gigabit switch, so 16 port Gigabit switch, but one point you
should notice is that each node is connected two switches.

So, you have got multiple network interface card, NIC stands for Network Interface Card
that means, each node is having multiple in this particular case two network interface
card, each network interface card each network interface card is connecting to one of
these switches, so two NICs connecting to two switches. So, in this way the
communication between any two node can be done using a single hop, so it requires
special routing to the setup at each node 348 port switches can be used to connect 64
PCs each using no more 2 NIC's per PCs. So, this is a very interesting topology and it
can be used to realize clusters.
(Refer Slide Time: 21:53)

This is an example Beowulf cluster, NASA build first Linux PC cluster in 1994, so it is a
low cost network PCs computers connected to one another by a private Ethernet
network. So, connection to any external network is through a single gate way computer,
and configuration is you are having commodity of self components, such as inexpensive
computers. And blade components are use that means, the computers mounted on a
mother board that are plugged into connectors on rack.
So, you will be using a rack on the back plane, there are connectors and each of the
blades can be pushed to connect to the system as I shall show you, and shared this can
shared nothing model is possible in whenever you go for this type of cluster.

(Refer Slide Time: 22:50)

This shows a single Beowulf cluster, as you can see this is the connector which can be
connected to the back plane. And here it shows two processor, this is one processor, this
is another processor to two processing module, which are present in a single blade.
(Refer Slide Time: 23:14)

And this Beowulf project of NASA used Linux and public domain software, so they did
not go for any custom software, so custom software is not being used, standard
commodity of the software has been used. So, made some changes to Linux kernel to
support things like channel bonding, so you have seen multiple Ethernet channels are

used to connect, but to connecting to a single virtual channel to overcome bandwidth

limitations. We know that the Ethernet has a limited bandwidth, like 1 Megabit per
second or 10 Megabits per second, whenever you go for Gigabit Ethernet.
But, if you want higher bandwidth, you can use a special type of technique and this is
what is being done known as, with the help of this special technique you can combine
multiple Ethernet channels into a single virtual channel. So, you can have more than 1
Gigabit bandwidth, so this will overcome bandwidth limitation. And Beowulf clusters
have become very popular, and you can have 1000 nodes in a single system.
(Refer Slide Time: 24:40)

Another example is Google infrastructure, Google serves on an average 1000 queries per
second, all of your familiar with I mean use of, when your using the cyber provided by
Google or for E-mail that is G-mail and all those things, so that is service to with the help
of this infrastructure. So, in addition to this the serving the queries a search engine must
crawl the web periodically to have up to date information. So, this is used for web
servers more than 6000 PCs and 12000 disks gives 1 peta byte of disk storage, rather
than using RAID this Google infrastructure relies and redundant storage and sites, so
each pc runs Linux.
And this is only the biggest source of failure is the software, in case of this infrastructure,
because you have got lot of redundant, I mean PCs computers and other things, so only
biggest source of failure is software.

(Refer Slide Time: 26:01)

So, this diagram shows Google infrastructure, you can see here different rack shown 10
racks in here, 10 racks here, we have got 40 racks connected by 4 copper Gigabit. And
that links 128 into 128 switches, so we have 128 into 128 switches and 1 rack contains
80 PCs. So, each rack is containing 80 PCs and you have got 40 such racks, so you can
imagine the total number of such some PCs have represent here.
And you are using that OC 12 and it is OC 45 that gives you 622 Megabits per second or
24988.988 Gigabit per second that I mean, that is the rate at switch you can communicate
with this end switch. So, end switch end gives you faster access to the entire
infrastructure, and you have got two such ends switches at both ends.

(Refer Slide Time: 27:22)

So, this is the diagram which shows the way, this is single rack and close of view of 1
rack PC, so this diagram shows you Google infrastructure.
(Refer Slide Time: 27:42)

There is another view of the cluster, that is implemented with a help of Google
infrastructure.

(Refer Slide Time: 27:51)

Then you can have cluster of shared memory multiprocessors, so highbred cluster of
shared memory multiprocessor can be used, this is often called constellation. So, this is
efficient shared memory programming at the individual node is possible, because each is
having a shared memory multiprocessor. So, this is better suited for implementation of
distributed shared memory multiprocessor, and this is meaningful only p-way node not
much expensive than p individual nodes.
So, this is what has been observe that, if you implement a p-way node that means, each
node is having p computers that is cheaper than to have p individual nodes, and then this
is meaningful and that is what is being done and example is Orion system. So, we have
discussed about one very important topic that is your cluster computing, and how
clusters have been implemented and various topologies and at the end we have shown
some example, now we shall focus or attention to another way of computing.

(Refer Slide Time: 29:08)

Grid computing.
(Refer Slide Time: 29:15)

So, why the name grid computing, so this the term grid computing originated in the early
1990's as a metaphor for making computer power has easy to access, as an electric power
grid, all of are your using electric power. So, as there is a that electric line is connected to
each and every house, and you can access it without much a botheration, so we know
that each plug point, we provide you to 20 volt or 120 volt a depending on where you
are. And that is the main feature, so power grid is transparent to the appliance, so

whether you will be running a computer or will be running some other appliance, it does
not matter you get the power.
And you can run different types of systems, without bothering where from the power is
coming and how the power is available and it is all pervasive. So, it is available
everywhere and any power point provide same power that means, you get 220 volt at
each every point. So, that is power availability is same at different points, to that is what
is the main feature of grid computing, so availability is uniform from different places.
Similarly, just like power grid, in compute grid is transfer rent to the user, so this
compute grid is transparent to the user, although it is distributed throughout the country.
That mean, it is access through wide area network and it is pervasive and computer
power is independent of fair a job is initiated. So, you can initiate a job from any fair and
you will get the same computing power for performing an application.
(Refer Slide Time: 31:26)

So, this question a sadly arises, what do you mean by grid computing, so this is
alternative to traditional large parallel or distribute systems, which can provide better
computational power for high and application, this is known as grid. And this is an
innovative extension of distributed computing technology, so grid computing is a kind of
distributed computing. But, it is not the way traditional distributing computing is done,
so it leverages a combination of hardware and software, virtualization, and distributed
sharing of those virtual resources, these are the two basic concept.

First up all hardware and software virtualization, second is distributed sharing of this
virtualized resources, and these resources can include all elements of computing. So, like
hardware, software, applications, networking services, pervasive devices and complex
footprints of computing power. So, you can see, you can have different types of a
different elements which are virtualized an available, so it can be actually sometimes
grid computing is also referred as CPU scavenging or cycle scavenging or cycle stealing
or shared computing, creates a grid from unused resources or in a network of
participants.
So, what is happening here, you have got large number of participants and all of them
performing their own job however, each of them may have some additional or extra
computing power, and which is been utilized in grid computing by others through the
internet. So, where worldwide or internal tune organizational, so this grid computing can
be done through, I mean across the world or internal tunnel to organization.
(Refer Slide Time: 33:43)

So, IBM's grid computing has put forward that actually basic idea was use open
standards, and protocols to be an access to computing resources over the internet. So,
you are gaining a access of computing resources over the internet, and you are using
large number of small systems spread across a large geographical area region, and it
presents unified picture to the users. So, here also it is giving some kind of unified

picture to the user, so you are not really concisely accessing, each and every different
computer, so you are accessing a grid. So, perform you are getting is transparent to you.
And this availability of high speed networking surmounts the distance problem, you may
be asking how you are able to get high performance through wide area network. The
reason for that is, because of the advisement of wide area network technology, now you
can have high bandwidth, the high bandwidth of access through internet. And that how
this availability of high speed networking is possible, and it is surmounts the distance
problem.
(Refer Slide Time: 35:15)

So, the grids can be divided into several types, one is your data grid another is your
compute grid. So, data grids must focus on data location, data transfer, data access and
critical aspects of security, because you are handling data, that data collection, storage,
retrieval from a large number of bases. So, this is the function of a data grid similarly,
you can have compute grids, it provides user with compute power for solving jobs. So,
the ability to allow for independent management of computing resources, the ability to
provide mechanisms that can intelligently and transparently selecting computing,
resources.
Capable of running a users job. the understanding of the current and predicted loads on
grid resources, resource availability, dynamic resource configuration and provisioning.
So, these are the features than failure detection and failure mechanisms, so the essential

appropriate security mechanism for secure resource management access and integrity,
and example is seti@home, this is an example of a grid computer.
(Refer Slide Time: 36:35)

And you can have three different types of grid environments, one is global grids, second
is enterprise grids, third is cluster grids. So, global grids are collection of enterprise and
cluster grids available across the globe, an enterprise grids are multiple projects of
departments here, resources within a enterprise your campus, so this is enterprise grids.
So, it does not do not at the security and global policy management issues, because since
it is restricted within enterprise you do not really require, it is not necessary to address
security. And global policy management issues, which is required in place of global
grids.

(Refer Slide Time: 37:23)

And third type is cluster grids, it is a simplest form of grid, it provides compute service at
the project or department level. The key benefits of cluster grid architectures are number
one is maximize use of computer resources, such as Linux machines and increase
throughput for user job. So, these are the two key benefits, that you achieve with the help
of cluster grids and cluster grid is superset of compute resources, such as Linux clusters.
So, cluster grids can operate with a heterogeneous environment with mixed server types,
mixed operating systems and workloads, so this is another very important feature, you
can have heterogeneous environment.
(Refer Slide Time: 38:13)

So, this is a comparison between P2P computing and grid computing, all of you are
familiar with Peer 2 Peer computing, which involves a large number of nodes connected
loosely, and this P2P computing is primarily used for sharing files. So, you can distribute
files over a large number of nodes, then for efficient sharing you distributes, then you
access from different nodes, that is the basic idea of Peer 2 Peer computing. On the other
hand, grids provide unified view, so target solving some specific types of problems such
as scientific research, business logic support etcetera, so it is not restricted to all if you
have file sharing.
(Refer Slide Time: 39:05)

So, here is a comparison between cluster computing versus grid computing, so as I

mentioned cluster computing can be said to have a sub set of grid computing. Because,
cluster nodes are in close proximity, and interconnected by LAN as I have mentioned,
and grid nodes are geographically separate and connected through wide area network or
WAN. So, clusters provide guaranty of service, nodes are expected to the full resource
on the other hand, clusters are usually heterogeneous site of clusters.
So, since it is distributed throughout the network availability and performance of grid
resource are unpredictable; and request from within an administrative domine may gain
more priority of over request from outside. So, this is another feature that has led to I
mean, that differences in availability of different systems, availability and performance
of heat resources are unpredictable because of many reasons and this is one of them.

(Refer Slide Time: 40:22)

So, this is that example SETI at home that was develop for to detect, and alien signals
through Arecibo radio telescope, world is largest telescope, so this used the idles cycles
of computers to analyze the data, generated from the telescope. So, over 500000 active
participant most of whom run screensaver on home PC and so from these active
participants the computing power is used to analyze the data generated from the
telescope.
And performance on the average over 20 teraflop per second, so you can see, you can
have you can achieve massive performance, because of large number of participants,
although each of them is giving only affection of their computing power.

(Refer Slide Time: 41:27)

Now, the big question is at some level and all application share common needs, how it is
being done, so how to find resources, how acquire resources, how to locate and move
data, how to start and monitor computation, and how to manage it all securely and
conveniently. So, for that purpose the software that is being used is known as grid
middleware, so the single software infrastructure supports all the above features.
(Refer Slide Time: 42:06)

So, these are the different middleware components, number one is grid information
service GIS, and support this supports registering and querying grid resources, grid

resource broker end user submit their application requirements to this grid resource
broker. Then grid resource broker discovers resources by querying the GIS, GIS means
this grid information service, then you have got grid fabric which manages resources like
computers, storage devices, scientific instruments etcetera.
Then you have got core grid middleware, this offer services like process management,
allocation of resources, security and quality of service and so on. And user level grid
middleware, this so far services like programming tools, resource broker, scheduling
application tasks for execution on global resources. So, you can see this grid computing
is feasible, over large geographical area with the help of these grid middleware
components.
(Refer Slide Time: 43:23)

Now, we have come to the last topic that is your cloud computing.

(Refer Slide Time: 43:32)

Cloud computing is also a kind of distributed computing, I mean somewhat similar to

grid computing, in case of grid computing we have seen with grid computing, one can
provision computing resources as a utility that can be turned on and off. So, here in grid
computing it is done in this way, you can provision a resource and after the provisioning
is done, you may utilize the entire resource or may utilize it part partially. And then if
you do not need it, then you can turned it off that means a particular resource
provisioning can be done, and you turn it on or off.
But, whenever an utility is being used, it is not guaranteed that you are utilizing the
entire resource and this cloud computing goes one step further with on-demand resource,
provisioning. So, here we are not really I mean a particular resource is turned on and
then, it does not allow instead of turning it on and using it partially or not in this case. A
concept called resource provisioning is done, so resource provisioning means you need,
whenever you need and as much you need it you will use that.
So, this is the on-demand resource provisioning in contrast to on off type resource,
process provisioning is done in grid computing. So, cloud computing is internet based
computing, where by shared resources, software and information are provided to
computers and other devices on-demand. And one big advantage of this on-demand
resource provisioning is that, you pay for the bandwidth and server resources that one

need. And when the requirement is over, then the turn the whole thing off, so you have to
paying only for that part you are using.
(Refer Slide Time: 45:53)

And here are the benefits of cloud computing, customers avoids capital expenditure of
the company that means, here what you are doing you are not really deploying costly
infrastructure, so you are using third party infrastructure to perform you job, so this is
one very important concept. So, this reduces the cost of purchasing, physical
infrastructure by renting users from a third party provider, so you are only renting a part
of the resources provided by a third party.
So, instead of in contrast to gird computing this eliminates over provisioning, and used
with utility pricing that means, instead of as I mentioned whenever in grid computing it
may lead to over provisioning. But, in case of you cloud computing, over provisioning is
not done and as a consequence you pay only a very small part and when used with utility
pricing. That means, you are using at an utility and as much as you used, you pay for that
and it also removes the need to over provision, in order to meet the demands of millions
of user.
You see in not only you are paying less, but it also helping others this is what
understand, so you are taking or you are consuming only as much you can, so the
remaining part is available for others. So, in this way grid computing can offer to grid
service to a large number of users, and with grid computing companies can scale up to

massive capacities in on instants without having to invest in new infrastructure, train a

new personal or license new software. So, you can see, a company whenever you need
massive computing, it can scale it up by hiring resources from a third party.
And as soon as the this high performance computing is over to simply turn it off,
similarly you can use massive storage, whenever you need it and whenever you do not
need it you simply turn it off. And also you can use a special software, there is no need
for to license a new software, so you are not doing the licensing, the license is not your
name, license is in the name of the service provider, you are simply using it and you are
paying for your uses.
So, without having license of the software, without purchasing the software, without
purchasing the hardware you are able to use them, to your advantage, to your benefit this
is the benefit of cloud computing.
(Refer Slide Time: 49:21)

And there are three important segments of applications, first one is the known as
applications, so applications software has a service, so which is known as SaaS in short.
SaaS provides a complete turnkey application such as enterprise resource management
throughout through the internet, so what you are doing here, you are using an
application, complete application through internet provided by a third party. Second type
of second segment of cloud computing is platform, so platform has a service, so which is
known as PaaS.

So, PaaS offers full or partial application development that users can access, so it a
platform is provided to you, for full or partial application, and that can be used and with
the help of this segment, second segment platform segment. Third is your infrastructure,
so infrastructure has a service or IaaS, so a consumer can get service from a full
computer infrastructure through the internet. So, without procuring it, you are able to get
full service, full computer infrastructure through the internet provided by a third party.
So, these are the three main segments, which are mainly available through cloud
computing.
(Refer Slide Time: 51:13)

Now, let us look at the different benefits and advantages of cloud computing, so the
benefits are you pay per use that means, as much as you use you pay for only that. And
you can have instant scalability, as soon as you are requirement increases, you can scale
it very quickly. And you can scaled on very quickly in an instant, because you are not
deploying it, deployment it is deployed by others, so based on your requirement you can
scale it up and scale it down, and only you pay for your use.
Then you can have high security, high security is provided through cloud computing, and
you can have high reliability as well, because when it is in cloud computing as you have
seen, you have got large number of computers and systems, and it gives you high
reliability. And it also gives you high APIs, so these are the benefits of cloud
computing, and the advantages of cloud computing is listed here, with provides you

lower cost of ownership. Since, you are not deploying anything you are not buying any
infrastructure hardware or software, your cost of ownership is very small.
So, it provides you lower cost of ownership and reduce infrastructure management
responsibility, so infrastructure management responsibility is also not with you, since it
is not part of you, your not really performing the infrastructure management.
Infrastructure management is done by third party your simply using it, and this is one
advantage and it allow for unexpected resource loads. So, whenever you have got
unexpected resource load, you have got large number load, you can very easily scale it
up, to meet the a load and faster application roll out.
So, because of this advantages, you are working on an application by using the
infrastructure and resources of cloud computing, using cloud computing you can
complete it and you can roll out you application very quickly. So, these are the main
advantages and this cloud economics is best on multi-tenanted, so that means, you can
have multiple resources, your using as tenants and virtualization lower cost of cost by
increasing utilizations.
As I mentioned hardware and software or virtualized, and bit lowest cost by increasing
utilization of the resources, and economic subscale afforded by technology. So, since
large number people are using it that the economy as a result cost is divided among a
large number of users, so economic scale afforded by technology. So, technology itself is
providing this sharing of cost by a large number of users, and automated update policy.
So, that means that as, and when as the requirement keeps on increasing, the
infrastructure is also upgraded to satisfy the requirements of users, whether it is hardware
resource or software resource, the resources kept on increasing. So, these are the
different aspects of cloud computing, so to summarize in this lecture we have discuss
three important types of computing. First one is cluster computing that is primarily used
for high performance computing.
And we have seen computers are within a small geographic area that is LAN, on the
other hand grid computing and cloud, in case of grid computing and cloud computing
resources are distributed through internet. And of course, in grid computing it is done in
one way and in cloud computing it is done in one way, and because of many benefits and
advantages of cloud computing, it is becoming increasingly popular.

Thank you.