Virtualization and what the cloud means

A cloud server is a virtual machine. Running many of them on one physical machine is called virtualization.

A student team needs a server for its project app tonight. It opens a cloud provider's console, chooses a 2 GB server and presses Create. About a minute later the console shows a running server. The team logs in and installs its app.

The team's laptopSends the Create request
A large machine in a data centerHost
The 2 GB serverVirtual machine
Think first. A data center cannot install and wire a new computer in one minute. What did the team receive?
Students usually answer "a computer in a data center". Steer them to "a share of a computer, made by software". The next step shows how many shares fit on one machine.

A provider runs many guests on one large host. Move the sliders to set how many guests share the host and how much memory each gets. Then push the total past the host's memory and read the result.

One physical machine divided into guests
Choose the number of guests
More detail

One machine with one operating system, and one machine with guests

Bare metal: one machine, one operating system
Applications
Shell
Kernel
Hardware

Bare metal means a physical machine with no virtualization. One operating system controls all of it. The kernel is the core of the operating system, and it controls the hardware. The shell is the program you type commands into. If you buy a bigger machine, you own all of it, including the 80% you are not using at 3 a.m.

Virtualized: one machine, many guests
Hypervisor: divides the machine between the guests
Host operating system
Hardware: 256 GB RAM, 64 cores

Each guest has its own kernel, disk and network card as far as it can tell, but these are simulated. The host is the physical machine, and its hypervisor divides the hardware between the guests. The hypervisor can create or delete a guest in seconds.

Overselling memory

Providers often sell more memory than the host has, because guests are rarely all busy at the same time. The hypervisor keeps some memory for itself. Move the sliders until the total passes the host's memory to see what happens.

The server in Server memory is one of these guests.

Provisioning a server takes about a minute in the cloud. Press Provision a server and watch the clock and the bill. Then switch to the on-premise setup, which models buying a server in 2012, and press it again.

day 0, 00:00
Running
0
Wait for first server
—
Bill so far
₱0
Choose the setup
More detail

Try this

In the cloud setup, provision three servers and delete two. The bill stops growing for the two you deleted. Leave one running and forget it, and the bill keeps growing.

Elasticity: adding machines as demand arrives

In this simulation an autoscaler adds machines. An autoscaler is a program that watches the load and adds machines when they are needed. Load is measured in requests per second (rps). Choose a day, press Start the day and watch what the autoscaler can absorb, because a new machine takes time to boot.

demand capacity, booting running turned away
Machines
0
Served so far
—
Turned away
0
Machine-minutes
0
Choose a day

Lowering the boot time recovers some of the surge and never all of it, because a surge that arrives in one second ends before anything that has to react first. Three things help: capacity that is already running, a queue in front, and spreading out the arrival times. Spreading out the arrivals is the fix from Case study: enlistment.

The US standards agency NIST defines cloud computing by five characteristics. Click each card to read it. Four of the five get a tick from what you did in the earlier steps. The fifth, resource pooling, is the shared host.

Select one

Click any of the five.

More detail

Where you have met these before

On-demand self-service: in Case study: enlistment, each of the seven servers needed a quotation, a purchase order and a delivery. A cloud server needs a credit card and about a minute.

Broad network access: in One API, many clients, the same HTTP endpoint served a browser, an iOS app and an Android app.

Resource pooling: the host in "Share one machine" is shared by many customers.

Rapid elasticity: the autoscaling web tier in Case study: enlistment depends on it. That case showed that the web tier was not the limit and the database was.

Measured service: it produces the cost chart in CapEx and OpEx.

The definition is from Mell and Grance (2011).

Cloud service models differ in how many layers the provider runs for you. Read across a row to see who runs that layer. You pay the provider for the layers it runs. The three models are IaaS, PaaS and SaaS.

LayerOn-premiseIaaSPaaSSaaS
you run it the provider runs it

You can now say who runs each layer under IaaS, PaaS and SaaS.

More detail

Examples

An Amazon EC2 instance or a DigitalOcean Droplet is IaaS. You receive a virtual machine and still install the operating system updates yourself.

Railway and Supabase are PaaS. You push your code to them and they run everything under it.

The learning management system and a web email service are SaaS for the people who use them. Your project is a PaaS customer, and it is a SaaS provider to its own users.