Advanced Operating Systems

Virtualization and Cloud Computing

Lecture 8
Today’s outline

1. Cloud computing

2. Virtualization

3. What’s really cloud computing?

2
Cloud computing
• Computing at scale
– The need for scalability; scale of current services
– Scaling up: From PCs to data centers
– Problems with 'classical' scaling techniques

• Utility computing and cloud computing

– What are utility computing and cloud computing?
– What kinds of clouds exist today?
– What kinds of applications run on the cloud?

3
How many users and objects?
• Flickr has > 10 billion photos

• Facebook has 2 billion monthly users

• Google is serving > 3.5 billion queries/day on more than

130 trillion pages

• > 5 billion videos/day watched on YouTube

• 300 hours of video are uploaded to YouTube every minute!

4
How much data?
• Modern applications use massive data:
– Rendering 'Avatar' movie required >1 petabyte
of storage
– CERN's LHC will produce about 15 petabytes of
data per year

25,400 km
– In 2008, Google processed 20 petabytes per day
– Dropbox has > 500 petabytes of user data
– Google now designing for 1 exabyte of storage
– NSA Utah Data Center is said to have 5 zettabyte (!)
• How much is a zettabyte?
– 1,000,000,000,000,000,000,000 bytes (1021)
– A stack of 1TB hard disks that is 25,400 km high

5
How much computation?
• No single computer can
process that much data
– Need many computers!
• How many computers do
modern services need?
– Facebook is thought to have more than 60,000 servers
– 1&1 Internet has over 70,000 servers
– Akamai has > 95,000 servers in 71 countries
– Intel has ~100,000 servers in 97 data centers
– Microsoft has > 1 million servers in 2008
– Google is thought to have more than 1 million servers,
is planning for 10 million (according to Jeff Dean)
6
Scaling up

PC Server Cluster

• What if one computer is not enough?

– Buy a bigger (server-class) computer

• What if the biggest computer is not enough?

– Buy many computers

7
Clusters
Network switch
(connects nodes with
each other and
with other racks)

Rack Many nodes/blades

(often identical)

Storage device(s)

• Characteristics of a cluster:
– Many similar machines, close interconnection (same room?)
– Often special, standardized hardware (racks, blades)
– Usually owned and used by a single organization
8
Power and cooling
• Clusters need lots of power
– Example: 140 Watts per server
– Rack with 32 servers: 4.5kW (use special power
supply!)
– Most of this power is converted into heat

• Large clusters need massive cooling

– 4.5kW is about 3 space heaters
– And that's just one rack!
9
Scaling up

PC Server Cluster Data center

• What if your cluster is too big (hot, power hungry) to fit into your office
building?
– Build a separate building for the cluster
– Building can have lots of cooling and power
– Result: Data center

10
What does a data center look like?
Cooling
plant

Data centers
(size of a
football field)

Google data center in The Dalles, Oregon

• A warehouse-sized computer
– A single data center can easily contain 10,000 racks
with 100 cores in each rack (1,000,000 cores total)

11
What’s in a data center?

Source: 1&1

• Hundreds or thousands of racks

12
What’s in a data center?

Source: 1&1

• Massive networking

13
What’s in a data center?

Source: 1&1

• Emergency power supplies

14
What’s in a data center?

Source: 1&1

• Massive cooling

15
Energy matters!
Company Servers Electricity Cost
eBay 16K ~0.6*105 MWh ~$3.7M/yr
Akamai 40K ~1.7*105 MWh ~$10M/yr
Rackspace 50K ~2*105 MWh ~$12M/yr
Microsoft >200K >6*105 MWh >$36M/yr
Google >500K >6.3*105 MWh >$38M/yr
USA (2006) 10.9M 610*105 MWh $4.5B/yr
Source: Qureshi et al., SIGCOMM 2009

• Data centers consume a lot of energy

– Makes sense to build them near sources of cheap electricity
– Example: Price per KWh is 3.6ct in Idaho (near hydroelectric
power), 10ct in California (long distance transmission),
18ct in Hawaii (must ship fuel)
– Most of this is converted into heat  Cooling is a big issue!
16
Scaling up

PC Server Cluster Data center Network of data centers

• What if even a data center is not big enough?

– Build additional data centers
– Where? How many?

17
Global distribution

• Data centers are often globally distributed

– Example above: Google data center locations (inferred)
• Why?
– Need to be close to users (physics!)
– Cheaper resources
– Protection against failures
18
Cloud computing
• Computing at scale
– The need for scalability; scale of current services
– Scaling up: From PCs to data centers
– Problems with 'classical' scaling techniques

• Utility computing and cloud computing

– What are utility computing and cloud computing?
– What kinds of clouds exist today?
– What kinds of applications run on the cloud?

19
Problem #1: Difficult to dimension
Jobs cannot Dissatisfied
be completed customers
2x-10x leave

Provisioning for the peak load Provisioning below the peak

• Problem: Load can vary considerably

– Peak load can exceed average load by factor 2x-10x [Why?]
– But: Few users deliberately provision for less than the peak
– Result: Server utilization in existing data centers ~5%-20%!!
– Dilemma: Waste resources or lose customers!

20
Problem #2: Expensive
• Need to invest many $$$ in hardware
– Even a small cluster can easily cost $100,000
– Microsoft recently invested $499 million in a single
data center

• Need expertise
– Planning and setting up a large cluster is highly
nontrivial
– Cluster may require special software, etc.

• Need maintenance
– Someone needs to replace faulty hardware, install 21
Problem #3: Difficult to scale
• Scaling up is difficult
– Need to order new machines, install them, integrate with
existing cluster - can take months!
– Large scaling factors may require major redesign, e.g., new
storage system, new interconnect, new building (!)

• Scaling down is difficult

– What to do with superfluous hardware?
– Server idle power is about 60% of peak  Energy is
consumed even when no work is being done
– Many fixed costs, such as construction

22
Summary: Computing at scale
• Modern applications require huge amounts of processing
and data
– Measured in petabytes, millions of users, billions of objects
– Need special hardware, algorithms, tools to work at this scale

• Clusters and data centers can provide the resources we need

– Main difference: Scale (room-sized vs. building-sized)
– Special hardware; power and cooling are big concerns

• Clusters and data centers are not perfect

– Difficult to dimension; expensive; difficult to scale

23
Cloud computing
• Computing at scale
– The need for scalability; scale of current services
– Scaling up: From PCs to data centers
– Problems with 'classical' scaling techniques

• Utility computing and cloud computing

– What are utility computing and cloud computing?
– What kinds of clouds exist today?
– What kinds of applications run on the cloud?

24
The power plant analogy

Steam engine at Stott Park Bobbin Mill

Waterwheel at the Neuhausen ob Eck Open-Air Museum

• It used to be that everyone had their own power source

– Challenges are similar to the cluster: Needs large
up-front investment, expertise to operate, difficult to
scale up/down...
25
Scaling the power plant

• Then people started to build large, centralized

power plants with very large capacity...

26
Metered usage model

Power source Network Metering Customer

device

• Power plants are connected to customers by a network

• Usage is metered, and everyone (basically) pays only for
what they actually use

27
Why is this a good thing?
Electricity Computing
• Economies of scale • Cheaper to run one big data
– Cheaper to run one big center than many small ones
power plant than many
small ones • High utilization!
• Statistical multiplexing
– High utilization!
• No investment in data center; pay-
• No up-front commitment as-you-go model
– No investment in generator;
pay-as-you-go model
• Scalability
• Thousands of computers available
– Thousands of kilowatts on demand; add
available on demand; add more within seconds
more within seconds
28
 What is cloud computing?

29

[Link]
What is cloud computing?

The interesting thing about Cloud Computing is that we've

redefined Cloud Computing to include everything that we
already do.... I don't understand what we would do differently
in the light of Cloud Computing other than change the
wording of some of our ads.
Larry Ellison, quoted in the Wall Street Journal, September 26, 2008

A lot of people are jumping on the [cloud] bandwagon, but

I have not heard two people say the same thing about it.
There are multiple definitions out there of "the cloud".
Andy Isherwood, quoted in ZDnet News, December 11, 2008

30
So what is it, really?
• According to NIST:
Cloud computing is a model for enabling convenient, on-demand
network access to a shared pool of configurable computing
resources (e.g., networks, servers, storage, applications, and
services) that can be rapidly provisioned and released with
minimal management effort or service provider interaction

• Essential characteristics:
– On-demand self service
– Broad network access
– Resource pooling
– Rapid elasticity
31
Other terms you may have heard
• Utility computing
– The service being sold by a cloud
– Focuses on the business model (pay-as-you-go), similar to
classical utility companies

• The Web
– The Internet’s information sharing model
– Some web services run on clouds, but not all

• The Internet
– A network of networks
– Used by the web; connects (most) clouds to their customers
32
Everything as a Service
• What kind of service does the cloud provide?
– Does it offer an entire application, or just resources?
– If resources, what kind / level of abstraction?

Three types commonly distinguished:

• Software as a service (SaaS)
– Analogy: Restaurant. Prepares & serves entire meal, does the dishes, ...
• Platform as a service (PaaS)
– Analogy: Take-out food. Prepares meal, but does not serve it
• Infrastructure as a service (IaaS)
– Analogy: Grocery store. Provides raw ingredients
• Other *aaS types have been defined, but are less common
– Desktop, DB, Communication, Network, Monitoring, …
33
Software as a Service (SaaS)
User

Application
Middleware
Hardware
Cloud
provider

• Cloud provides an entire application

– Word processor, spreadsheet, CRM software, calendar...
– Customer pays cloud provider
– Example: Google Apps, [Link], Concur, Doodle
34
Platform as a Service (PaaS)
SaaS
provider User

Application
Middleware
Hardware
Cloud
provider

• Cloud provides middleware/infrastructure

– For example, Microsoft Common Language Runtime (CLR)
– Customer pays SaaS provider for the service; SaaS provider pays
the cloud for the infrastructure
– Example: Windows Azure, Google App Engine, Heroku
35
Infrastructure as a Service (IaaS)
SaaS
provider User

Application
Middleware
Hardware
Cloud
provider

• Cloud provides raw computing resources

– Virtual machine, blade server, storage, network, ...
– Customer pays SaaS provider for the service; SaaS provider pays
the cloud for the resources
– Examples: Amazon Web Services, DigitalOcean, Joyent
36
Private/hybrid/community clouds
CompanyCompany
A B

Public Community Private

• Who can become a customer of the cloud?

– Public cloud: Commercial service; open to (almost) anyone
Example: Amazon AWS, Microsoft Azure, Google App Engine
– Community cloud: Shared by several similar organizations.
Example: Google’s “Gov Cloud”
– Private cloud: Shared within a single organization.
Example: Internal datacenter of a large company.
37
Summary: Utility/cloud computing
• Why is cloud computing attractive?
– Analogy to 'classical' utilities (electricity, water, ...)
– No up-front investment (pay-as-you-go model)
– Low price due to economies of scale
– Elasticity: can quickly scale up/down as demand varies
• Different types of clouds
– SaaS, PaaS, IaaS; public/private/community clouds
• What runs on the cloud?
– Many potential applications: Application hosting,
backup/storage, scientific computing, content delivery, ...

38
Examples of cloud applications
• Application hosting
• Backup and Storage
• Content delivery
• E-commerce
• High-performance computing
• Media hosting
• On-demand workforce
• Search engines
• Web hosting
If interested, glance through
[Link]
for a list of many use cases
39
Case study:

[Link]
• Animoto: Lets users
create videos from
their own photos/music
– Auto-edits photos and aligns
them with the music, so it Source: Jeff Bezos' talk at Stanford on 4/19/08

"looks good"
• Built using Amazon EC2+S3+SQS
• Released a Facebook app in mid-April 2008
– More than 750,000 people signed up within 3 days
– EC2 usage went from 50 machines to 3,500 (x70
[Link]
scalability!)
40
Case study:
• Novartis Institutes for Biomedical Research is focused on
the drug discovery phase of the
~10 year / $1 billion drug development process
• In 2013, NIBR ran a project to screen 10 M compounds
against a common cancer target
• Compute requirements >> internal capacity / $
• The project ran across 10,500 EC2 Spot instances
(~87,000 cores) for $4,232 in 9 hours
• Equiv. of 39 years of computational chemistry

[Link]
41
Is the cloud good for everything?
• No.
• Sometimes it is problematic
– Auditability requirements
– Legislative frameworks
• Example: Personal data privacy
– EU Data Protection law
• Example: Processing medical records (US)
– HIPAA (Health Insurance Portability and
Accountability Act) privacy and security rule

42
Summary: Cloud applications
• Clouds are good for many things...
– Applications that involve large amounts of computation,
storage, bandwidth
– Especially when lots of resources are needed quickly
(Novartis example) or load varies rapidly (Animoto
example)

• ... but not for all things

– and might not be the cheapest solution either

43
Today’s outline

1. Cloud computing

2. Virtualization

3. What’s really cloud computing?

44
How clouds work?

Bob

Charlie
Alice
Physical machine
Daniel
• Suppose Alice has a machine with 4 CPUs and 8 GB of memory,
and three customers:
– Bob wants a machine with 1 CPU and 3GB of memory
– Charlie wants 2 CPUs and 1GB of memory
– Daniel wants 1 CPU and 4GB of memory
• What should Alice do?
45
Virtualization is key enabler

Bob
Virtual
machine
monitor
Charlie
Alice
Physical machine
Virtual machines Daniel

• Alice can sell each customer a virtual machine (VM) with the
requested resources
– From each customer’s perspective, it appears as if they had
a physical machine all by themselves (isolation)
46
Virtualization
“a technique for hiding the physical characteristics of
computing resources from the way in which other systems,
applications, or end users interact with those resources. This
includes making a single physical resource appear to function
as multiple logical resources; or it can include making multiple
physical resources appear as a single logical resource”

47
The idea of Virtualization: from 1960’s
• IBM VM/370 – A VMM for IBM mainframe
– Multiple OS environments on expensive hardware
– Desirable when few machine around
• Popular research idea in 1960s and 1970s
– Entire conferences on virtual machine monitors
– Hardware/VMM/OS designed together
– Allowed multiple users to share a batch oriented system
• Interest died out in the 1980s and 1990s
– Hardware got more cheaper
– Operating systems got more powerful (e.g. multi-user)

48
A Return to Virtual Machines
• Disco: Stanford research project (SOSP ’97)
– Run commodity OSes on scalable multiprocessors
– Focus on high-end: NUMA, MIPS, IRIX
• Commercial virtual machines for x86 architecture
– VMware Workstation (now EMC) (1999-)
– Connectix VirtualPC (now Microsoft)
• Research virtual machines for x86 architecture
– Xen (SOSP ’03)
– plex86
• OS-level virtualization
– FreeBSD Jails, User-mode-linux, UMLinux

49
Starting Point: A Physical Machine
• Physical Hardware
– Processors, memory,
chipset, I/O devices, etc.
– Resources often grossly
underutilized
• Software
– Tightly coupled to
physical hardware
– Single active OS instance
– OS controls hardware

50
What is a Virtual Machine?
• Software Abstraction
– Behaves like hardware
– Encapsulates all OS and
application state
• Virtualization Layer
– Extra level of indirection
– Decouples hardware, OS
– Enforces isolation
– Multiplexes physical
Virtual Machine hardware
Monitor across VMs
or Hypervisor

51
Virtualization Properties, Features
• Isolation
– Fault isolation
– Performance isolation (+ software isolation, …)
• Encapsulation
– Cleanly capture all VM state
– Enables VM snapshots, clones
• Portability
– Independent of physical hardware
– Enables migration of live, running VMs (freeze, suspend,…)
• Interposition
– Transformations on instructions, memory, I/O
– Enables transparent resource overcommitment,
encryption, compression, replication …
52
Types of Virtualization
Process Virtualization System Virtualization
• Language-level Java, .NET, • VMware Workstation,
Smalltalk Microsoft VPC, Parallels
• OS-level processes, Solaris • VMware ESX, Xen, Microsoft
Zones, BSD Jails, Docker Hyper-V
Containers
• Cross-ISA emulation Apple
68K-PPC-x86

53
Language Level Virtualization
• not-really-virtualization but using same techniques,
providing similar features
– Programming language is designed to run within
custom-built virtualized environment (e.g. Oracle JVM)
• Virtualization is defined as providing APIs that define a set
of features made available to a language and programs
written in that language to provide an improved execution
environment
– JVM compiled to run on many systems
– Programs written in Java run in the JVM no matter the
underlying system
– Similar to interpreted languages

54
Types of VMs – Emulation
• Another (older) way for running one OS on a different OS
– Virtualization requires underlying CPU to be same as guest was
compiled for while Emulation allows guest to run on different CPU
• Need to translate all guest instructions from guest CPU to native
CPU
– Emulation, not virtualization
• Useful when host and guest have different processor architectures
– Company replacing outdated servers with new servers containing
different CPU architecture, but still want to run old applications
• Performance challenge – orders of magnitude slower than native
code
– New machines faster than older machines so can reduce slowdown
• Where do you think it is used still?
– Very popular – especially in gaming to emulate old consoles
55
VMs – Application Containers
• Some goals of virtualization are segregation of apps, performance and
resource management, easy start, stop, move, and management of
them
• Can do those things without full-fledged virtualization
– If applications compiled for the host operating system, don’t need full virtualization
to meet these goals
• Containers / zones for example create virtual layer between OS and
apps
– Only one kernel running – host OS
– OS and devices are virtualized, providing resources within zone with impression
that they are only processes on system
– Each zone has its own applications; networking stack, addresses, and ports; user
accounts, etc
– CPU and memory resources divided between zones
• Zone can have its own scheduler to use those resources
56
Types of System Virtualization

• Native/Bare metal (Type 1)

– Higher performance
– ESX, Xen, HyperV

• Hosted (Type 2)
– Easier to install
– Leverage host’s device drivers
– VMware Workstation, Parallels
Attribution: [Link] 57
Types of Virtualization
Full virtualization (e.g. Para virtualization (e.g. XEN)
VMWare ESX) • OS modified to be
• Unmodified OS, virtualized,
virtualization is transparent • Better performance at cost
to OS of transparency
• VM looks exactly like a
physical machine

Attribution [Link] 58
What is a Virtual Machine Monitor?
• Classic Definition (Popek and Goldberg ’74)

• VMM Properties
– Equivalent execution: Programs running in the virtualized
environment run identically to running natively
– Performance: A statistically dominant subset of the
instructions must be executed directly on the CPU
– Safety and isolation: A VMM most completely control
access to system resources

59
VMM Implementation Goals
• Should efficiently virtualize the hardware
– Provide illusion of multiple machines
– Retain control of the physical machine

• Which subsystems should be virtualized?

– Processor => Processor Virtualization
– Memory => Memory Virtualization
– I/O Devices => I/O virtualization

60
Processor Virtualization

• An architecture is classically/strictly virtualizable if all its

sensitive instructions (those that violate safety and
encapsulation) are a subset of the privileged instructions.

• all instructions either trap or execute identically

• instructions that access privileged state trap
Attribution: [Link] 61
System Call Example

• Run guest operating system deprivileged

• All privileged instructions trap into VMM
• VMM emulates instructions against virtual state
e.g. disable virtual interrupts, not physical interrupts
• Resume direct execution from next guest instruction 62
x86 Virtualization challenges
• x86 not Classically Virtualizable
– x86 ISA has instructions that RD/WR privileged state
– …But which don’t trap in unprivileged mode
• Example: POPF instruction
– Pop top-of-stack into EFLAGS register
– [Link] bit privileged (interrupt enable flag)
– POPF silently ignores attempts to alter [Link] in
unprivileged mode! => no trap to return control to VMM

• Techniques to address inability to virtualize x86

– Replace non-virtualizable instructions with easily
Virtualized ones statically (Paravirtualization) 63
Virtualizing the CPU
• The VMM still need to multiplex VMs on CPUs

• How could this be done?

– # Physical CPUs more than #Virtual CPUs?
– # Virtual CPUs more than #Physical CPUs?

• Timeslice the VMs, each VM runs OS/Apps

• Use simple CPU scheduler

– Round robin, work-conserving (give extra to other VM)
– Can oversubscribe and give more #VCPUs that actual
64
Virtualizing Memory
• OS assumes that it has full control over memory
– Management: Assumes it owns it all
– Mapping: Assumes it can map any Virtual-> Physical
• However, VMM partitions memory among VMs
– VMM needs to assign hardware pages to VMs
– VMM needs to control mapping for isolation
• Cannot allow OS to map any Virtual => hardware page
• Hardware-managed TLBs make this difficult
– On TLB misses, the hardware walks page tables in
mem
– VMM needs to control access by OS to page tables

65
x86 Memory Management Primer
• The processor operates with virtual addresses
• Physical memory operates with physical addresses
• x86 includes a hardware translation lookaside buffer (TLB)
– Maps virtual to physical page addresses
• x86 handles TLB misses in HW
– HW walks the page tables => Inserts virtual to physical mapping

66
Shadow Page Tables
Three abstractions of memory
• Machine: actual hardware memory (e.g. 2GB of DRAM)
• Physical: abstraction of hardware memory, OS managed
– E.g. VMM allocates 512 MB to a VM, the OS thinks the
computer has 512 MB of contiguous physical memory
– (Underlying machine memory may be discontiguous)
• Virtual: virtual address space
– Standard 232 address space

• In each VM, OS creates and manages page tables for its

virtual address spaces without modification
– But these page tables are not used by the MMU 67
Three Abstractions of Memory
Virtual
Pages

• Native

Virtual
Pages

• Virtualized

68
Shadow Page Tables (continued)
• VMM creates and manages page tables that map virtual pages
directly to machine pages
– These tables are loaded into the MMU on a context switch
– VMM page tables are the shadow page tables
• VMM needs to keep its V => M tables consistent with changes
made by OS to its V=>P tables
– VMM maps OS page tables as read only
– When OS writes to page tables, trap to VMM
– VMM applies write to shadow table and OS table, returns
– Also known as memory tracing
– Again, more overhead...
69
Memory Management / Allocation
• VMMs tend to have simple memory management
– Static policy: VM gets 8GB at start
• Dynamic adjustment is hard since OS cannot handle
• No swapping to disk

• More sophistication: Overcommit with balooning

– Baloon driver runs inside OS => consume hardware
pages
– Baloon grows or shrinks (gives back mem to other VMs)

• Even more sophistication: memory de-duplication

– Identify pages that are shared across VMs! 70
Memory Ballooning

71
Page Sharing

72
Page Sharing

73
I/O Virtualization
• Challenge: Lots of I/O devices
– Problem: Writing device drivers for all I/O devices in the
VMM layer is not a feasible option
• Insight: Device driver already written for popular OSes
• Solution: Present virtual I/O devices to guest VMs and
channel I/O requests to a trusted host VM (popular OS)

74
Virtualizing I/O Devices
• However, overall I/O is complicated for VMMs
– Many short paths for I/O in OSes for performance
– Better if hypervisor needs to do less for I/O for guests,
– Possibilities include direct device access, DMA pass-
through, direct interrupt delivery (need H/W support!)
• Networking also complex as VMM and guests all need
network access
– VMM can bridge guest to network (direct access)
– VMM can provide network address translation (NAT)
• NAT address local to machine on which guest is running
• VMM provides address translation to guest to hide its address

75
VM Storage Management
• VMM provides both boot disk + other storage
– Type 1 VMM – storage guest root disks and config
information within file system provided by VMM as a
disk image
– Type 2 VMM – store as files in the host file system/OS
• Example of supported operations:
– Duplicate file (clone) -> create new guest
– Move file to another system -> move guest
– Convert formats: Physical-to-Virtual (P-to-V) and
Virtual-to-Physical (V-to-P)
• VMM also usually provides access to network attached
storage (just networking) => live migration
76
OS Component – Live Migration
• VMMs allow new functionality like live migration
– Running guest OS can be moved between systems,
without interrupting user access to the guest or its apps
– Very useful for resource management, no downtime
windows, etc
• How does it work?

77
Live Migration of VMs

78
OS Component – Live Migration
• VMMs allow new functionality like live migration
– Running guest OS can be moved between systems, without
interrupting user access to the guest or its apps
– Very useful for resource management, no downtime windows, etc
• How does it work?
1. Source VMM connects to the target VMM
2. Target VMM creates a new guest (e.g. create a new VCPU, etc)
3. Source sends all read-only guest memory pages to the target
4. Source sends all RD/WR pages to the target, marking them clean
5. Source repeats step 4, as some pages may be modified => dirty
6. When cycle of steps 4 and 5 becomes very short, source VMM freezes
guest, sends VCPU’s final state + other state details, sends final dirty
pages, and tells target to start running the guest
7. Target acknowledges that guest running => source terminates guest
79
Summary: Virtualization
Resources (CPU, memory, …) are virtualized
• VMM (Hypervisor) has translation tables that map requests for virtual
resources to physical resources
• Implementation must have low overheads
– Types: full system vs para virtualization vs containers
• different trade-offs

Properties: Isolation, Encapsulation, Portability, Interposition

• Multiple VMs can time-share resources -> high utilization
• VMs can be migrated -> balance load, handle planned maintenance,
cope with failures

Key enabler of Cloud computing!

80
Today’s outline

1. Cloud computing

2. Virtualization

3. What’s really cloud computing?

81
Serverless computing

• Enter post VM and container based clouds

• Cloud execution model in which cloud provider dynamically
manages the allocation of machine resources
• Developers don't have to worry about managing,
provisioning and maintaining servers
– Just write and deploy code on the cloud

• It’s a form of PaaS but pricing is based on the actual

amount of resources consumed by an application
– AWS charges for compute power in 100 ms increments
82
Next lecture topic:
Network file systems

83