BCS601 Cloud Computing Module-1

Module-1
Distributed System Models and Enabling Technologies:
Scalable Computing Over the Internet, Technologies for Network Based Systems, System Models
for Distributed and Cloud Computing, Software Environments for Distributed Systems and
Clouds, Performance, Security and Energy Efficiency.
Textbook 1: Chapter 1: 1.1 to 1.5

EVOLUTION OF DISTRIBUTED COMPUTING

• Grids enable access to shared computing power and storage capacity from your desktop.
• Clouds enable access to leased computing power and storage capacity from your desktop.
• Grids are an open source technology. Resource users and providers alike can understand and
contribute to the management of their grid
• Clouds are a proprietary technology. Only the resource provider knows exactly how their cloud
manages data, job queues, security requirements and so on.
• The concept of grids was proposed in 1995. The Open science grid (OSG) started in 1995 The
EDG (European Data Grid) project began in 2001.
• In the late 1990`s Oracle and EMC offered early private cloud solutions . However the term cloud
computing didn't gain prominence until 2007. o high-performance computing (HPC) applications
is no longer optimal for measuring system performance
• The emergence of computing clouds instead demands high-throughput computing (HTC)
systems built with parallel and distributed computing technologies
• We have to upgrade data centers using fast servers, storage systems, and high-bandwidth
networks.
• From 1950 to 1970, a handful of mainframes, including the IBM 360 and CDC 6400
1.1 SCALABLE COMPUTING OVER THE INTERNET
Instead of using a centralized computer to solve computational problems, a parallel and distributed
computing system uses multiple computers to solve large-scale problems over the Internet. Thus,
distributed computing becomes data-intensive and network-centric.
The Age of Internet Computing
The Platform Evolution
o From 1960 to 1980, lower-cost minicomputers such as the DEC PDP 11 and VAX Series
o From 1970 to 1990, we saw widespread use of personal computers built with VLSI microprocessors.
o From 1980 to 2000, massive numbers of portable computers and pervasive devices appeared in both
wired and wireless applications
o Since 1990, the use of both HPC and HTC systems hidden in clusters, grids, or Internet clouds has
proliferated

Dr. Sampada K S, Associate professor CSE RNSIT pg. 1

BCS601 Cloud Computing Module-1

High-Performance Computing (HPC) and High-Throughput Computing (HTC) have evolved

significantly, driven by advances in clustering, P2P networks, and cloud computing.
• HPC Evolution:
o Traditional supercomputers (MPPs) are being replaced by clusters of cooperative
computers for better resource sharing.
o HPC has focused on raw speed performance, progressing from Gflops (1990s) to
Pflops (2010s).
• HTC and P2P Networks:
o HTC systems prioritize high-flux computing, emphasizing task throughput over
raw speed.
o P2P networks facilitate distributed file sharing and content delivery using globally
distributed client machines.
o HTC applications dominate areas like Internet searches and web services for
millions of users.
• Market Shift from HPC to HTC:
o HTC systems address challenges beyond speed, including cost, energy efficiency,
security, and reliability.
• Emerging Paradigms:
o Advances in virtualization have led to the rise of Internet clouds, enabling service-
oriented computing.
o Technologies like RFID, GPS, and sensors are fueling the growth of the Internet of
Things (IoT).
• Computing Model Overlaps:
o Distributed computing contrasts with centralized computing.
o Parallel computing shares concepts with distributed computing.
o Cloud computing integrates aspects of distributed, centralized, and parallel
computing.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 2

BCS601 Cloud Computing Module-1

The transition from HPC to HTC marks a strategic shift in computing paradigms, focusing on
scalability, efficiency, and real-world usability over pure processing power.
Computing Paradigm Distinctions
Centralized computing
A computing paradigm where all computer resources are centralized in a single physical
system. In this setup, processors, memory, and storage are fully shared and tightly integrated
within one operating system. Many data centers and supercomputers operate as centralized
systems, but they are also utilized in parallel, distributed, and cloud computing applications.
• Parallel computing
In parallel computing, processors are either tightly coupled with shared memory or loosely
coupled with distributed memory. Communication occurs through shared memory or message
passing. A system that performs parallel computing is a parallel computer, and the programs
running on it are called parallel programs. Writing these programs is referred to as parallel
programming.
• Distributed computing studies distributed systems, which consist of multiple autonomous
computers with private memory communicating through a network via message passing.
Programs running in such systems are called distributed programs, and writing them is known
as distributed programming.
Cloud computing refers to a system of Internet-based resources that can be either centralized
or distributed. It uses parallel, distributed computing, or both, and can be established with
physical or virtualized resources over large data centers. Some regard cloud computing as a
form of utility computing or service computing. Alternatively, terms such as concurrent
computing or concurrent programming are used within the high-tech community, typically
referring to the combination of parallel and distributed computing, although interpretations
may vary among practitioners.

• Ubiquitous computing refers to computing with pervasive devices at any place and time
using wired or wireless communication. The Internet of Things (IoT) is a networked connection
of everyday objects including computers, sensors, humans, etc. The IoT is supported by Internet
clouds to achieve ubiquitous computing with any object at any place and time. Finally, the term
Internet computing is even broader and covers all computing paradigms over the Internet. This
book covers all the aforementioned computing paradigms, placing more emphasis on
distributed and cloud computing and their working systems, including the clusters, grids, P2P,
and cloud systems.
Internet of Things The traditional Internet connects machines to machines or web pages to web
pages. The concept of the IoT was introduced in 1999 at MIT.
• The IoT refers to the networked interconnection of everyday objects, tools, devices, or computers.
One can view the IoT as a wireless network of sensors that interconnect all things in our daily life.
• It allows objects to be sensed and controlled remotely across existing network infrastructure

Dr. Sampada K S, Associate professor CSE RNSIT pg. 3

BCS601 Cloud Computing Module-1

Distributed System Families

Massively distributed systems, including grids, clouds, and P2P networks, focus on resource
sharing in hardware, software, and datasets. These systems emphasize parallelism and concurrency,
as demonstrated by large-scale infrastructures like the Tianhe-1A supercomputer (built in China in
2010 with over 3.2 million cores).
Future HPC (High-Performance Computing) and HTC (High-Throughput Computing) systems
will require multicore and many-core processors to support large-scale parallel computing. The
effectiveness of these systems is determined by the following key design objectives:
1. Efficiency – Maximizing resource utilization for HPC and optimizing job throughput, data
access, and power efficiency for HTC.
2. Dependability – Ensuring reliability, self-management, and Quality of Service (QoS), even
in failure conditions.
3. Adaptability – Supporting large-scale job requests and virtualized resources across different
workload and service models.
4. Flexibility – Enabling HPC applications (scientific and engineering) and HTC applications
(business and cloud services) to run efficiently in distributed environments.
The future of distributed computing depends on scalable, efficient, and flexible architectures that
can meet the growing demand for computational power, throughput, and energy efficiency.
Scalable Computing Trends and New Paradigms
Scalable computing is driven by technological advancements that enable high-performance
computing (HPC) and high-throughput computing (HTC). Several trends, such as Moore’s Law
(doubling of processor speed every 18 months) and Gilder’s Law (doubling of network bandwidth
each year), have shaped modern computing. The increasing affordability of commodity hardware
has also fueled the growth of large-scale distributed systems.
Degrees of Parallelism
Parallelism in computing has evolved from:
• Bit-Level Parallelism (BLP) – Transition from serial to word-level processing.
• Instruction-Level Parallelism (ILP) – Executing multiple instructions simultaneously
(pipelining, superscalar computing).
• Data-Level Parallelism (DLP) – SIMD (Single Instruction, Multiple Data) architectures.
• Task-Level Parallelism (TLP) – Parallel execution of independent tasks on multicore
processors.
• Job-Level Parallelism (JLP) – Large-scale distributed job execution in cloud computing.
Coarse-grained parallelism builds on fine-grained parallelism, ensuring scalability in HPC and HTC
systems.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 4

BCS601 Cloud Computing Module-1

Innovative Applications of Distributed Systems

Parallel and distributed systems support applications in various domains:
Domain Applications
Science & Engineering Weather forecasting, genomic analysis
Business, education, services industry, and E-commerce, banking, stock exchanges
health care
Internet and web services, and government Cybersecurity, digital governance, traffic
applications monitoring
Mission-Critical Systems Military, crisis management

HTC systems prioritize task throughput over raw speed, addressing challenges like cost, energy
efficiency, security, and reliability.
The Shift Toward Utility Computing
Utility computing follows a pay-per-use model where computing resources are delivered as a service.
Cloud computing extends this concept, allowing distributed applications to run on edge networks.

Challenges include:
• Efficient network processors
• Scalable storage and memory
• Virtualization middleware
• New programming models
The Hype Cycle of Emerging Technologies
New technologies follow a hype cycle, progressing through:
1. Technology Trigger – Early development and research.
2. Peak of Inflated Expectations – High expectations but unproven benefits.
3. Trough of Disillusionment – Realization of limitations.
4. Slope of Enlightenment – Gradual improvements.
5. Plateau of Productivity – Mainstream adoption.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 5

BCS601 Cloud Computing Module-1

For example, in 2010, cloud computing was moving toward mainstream adoption, while broadband
over power lines was expected to become obsolete.
The Internet of Things (IoT) and Cyber-Physical Systems (CPS)
• IoT: Interconnects everyday objects (sensors, RFID, GPS) to enable real-time tracking and
automation.
• CPS: Merges computation, communication, and control (3C) to create intelligent systems
for virtual and physical world interactions.
Both IoT and CPS will play a significant role in future cloud computing and smart infrastructure
development.
1.2 Technologies for Network-Based Systems
Advancements in multicore CPUs and multithreading technologies have played a crucial role in
the development of high-performance computing (HPC) and high-throughput computing (HTC).
Advances in CPU Processors

Dr. Sampada K S, Associate professor CSE RNSIT pg. 6

BCS601 Cloud Computing Module-1

• Modern multicore processors integrate dual, quad, six, or more processing cores to
enhance parallelism at the instruction level (ILP) and task level (TLP).
• Processor speed growth has followed Moore’s Law, increasing from 1 MIPS (VAX 780,
1978) to 22,000 MIPS (Sun Niagara 2, 2008) and 159,000 MIPS (Intel Core i7 990x, 2011).
• Clock rates have increased from 10 MHz (Intel 286) to 4 GHz (Pentium 4) but have
stabilized due to heat and power limitations.
Multicore CPU and Many-Core GPU Architectures

• Multicore processors house multiple processing units, each with private L1 cache and
shared L2/L3 cache for efficient data access.
• Many-core GPUs (e.g., NVIDIA and AMD architectures) leverage hundreds to thousands
of cores, excelling in data-level parallelism (DLP) and graphics processing.
• Example: Sun Niagara II – Built with eight cores, each supporting eight threads, achieving
a maximum parallelism of 64 threads.
Key Trends in Processor and Network Technology
• Multicore chips continue to evolve with improved caching mechanisms and increased
processing cores per chip.
• Network speeds have improved from Ethernet (10 Mbps) to Gigabit Ethernet (1 Gbps)
and beyond 100 Gbps to support high-speed data communication.
Modern distributed computing systems rely on scalable multicore architectures and high-speed
networks to handle massive parallelism, optimize efficiency, and enhance overall performance.
Multicore CPU and Many-Core GPU Architectures
Advancements in multicore CPUs and many-core GPUs have significantly influenced modern
high-performance computing (HPC) and high-throughput computing (HTC) systems. As CPUs
approach their parallelism limits, GPUs have emerged as powerful alternatives for massive
parallelism and high computational efficiency.
Multicore CPU and Many-Core GPU Trends

Dr. Sampada K S, Associate professor CSE RNSIT pg. 7

BCS601 Cloud Computing Module-1

• Multicore CPUs continue to evolve from tens to hundreds of cores, but they face challenges
like the memory wall problem, limiting data-level parallelism (DLP).
• Many-core GPUs, with hundreds to thousands of lightweight cores, excel in DLP and
task-level parallelism (TLP), making them ideal for massively parallel workloads.
• Hybrid architectures are emerging, combining fat CPU cores and thin GPU cores on a
single chip for optimal performance.
Multithreading Technologies in Modern CPUs
• Different microarchitectures exploit parallelism at instruction-level (ILP) and thread-
level (TLP):
o Superscalar Processors – Execute multiple instructions per cycle.
o Fine-Grained Multithreading – Switches between threads every cycle.
o Coarse-Grained Multithreading – Runs one thread for multiple cycles before
switching.
o Simultaneous Multithreading (SMT) – Executes multiple threads in the same cycle.

GPU Computing to Exascale and Beyond

• GPUs were initially designed for graphics acceleration but are now used for general-purpose
parallel computing (GPGPU).

Dr. Sampada K S, Associate professor CSE RNSIT pg. 8

BCS601 Cloud Computing Module-1

• Modern GPUs (e.g., NVIDIA CUDA, Tesla, and Fermi) feature hundreds of cores,
handling thousands of concurrent threads.
• Example: The NVIDIA Fermi GPU has 512 CUDA cores and delivers 82.4 teraflops,
contributing to the performance of top supercomputers like Tianhe-1A.

GPU vs. CPU Performance and Power Efficiency

• GPUs prioritize throughput, while CPUs optimize latency using cache hierarchies.
• Power efficiency is a key advantage of GPUs – GPUs consume 1/10th of the power per
instruction compared to CPUs.
• Future Exascale Systems will require 60 Gflops/W per core, making power efficiency a
major challenge in parallel and distributed computing.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 9

BCS601 Cloud Computing Module-1

Challenges in Future Parallel and Distributed Systems

1. Energy and Power Efficiency – Reducing power consumption while increasing
performance.
2. Memory and Storage Bottlenecks – Optimizing data movement to avoid bandwidth
limitations.
3. Concurrency and Locality – Improving software and compiler support for parallel
execution.
4. System Resiliency – Ensuring fault tolerance in large-scale computing environments.
The shift towards hybrid architectures (CPU + GPU) and the rise of power-aware computing
models will drive the next generation of HPC, HTC, and cloud computing systems.
1.2.3 Memory, Storage, and Wide-Area Networking
Memory Technology
• DRAM capacity has increased 4x every three years (from 16 KB in 1976 to 64 GB in 2011).
• Memory access speed has not kept pace, causing the memory wall problem, where CPUs
outpace memory access speeds.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 10

BCS601 Cloud Computing Module-1

Disks and Storage Technology

• Hard drive capacity has grown 10x every eight years, reaching 3 TB (Seagate Barracuda
XT, 2011).
• Solid-State Drives (SSDs) provide significant speed improvements and durability (300,000
to 1 million write cycles per block).
• Power and cooling challenges limit large-scale storage expansion.
System-Area Interconnects & Wide-Area Networking
• Local Area Networks (LANs) connect clients and servers.
• Storage Area Networks (SANs) & Network Attached Storage (NAS) support large-scale
data storage and retrieval.
• Ethernet speeds have evolved from 10 Mbps (1979) to 100 Gbps (2011), with 1 Tbps links
expected in the future.
• High-speed networking enhances distributed computing efficiency and scalability.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 11

BCS601 Cloud Computing Module-1

1.2.4 Virtual Machines and Virtualization Middleware

Virtualization in Distributed Systems
• Traditional computing tightly couples OS and hardware, reducing flexibility.
• Virtual Machines (VMs) abstract hardware resources, allowing multiple OS instances on a
single system.

Virtual Machine Architectures

1. Native VM (Hypervisor-based) – Direct hardware access via bare-metal hypervisors (e.g.,
VMware ESXi, Xen).
Native VMs, also known as bare-metal virtualization, directly run on physical hardware
without requiring a host operating system. These VMs rely on a hypervisor (or Virtual
Machine Monitor, VMM) to manage multiple virtual instances running on a single hardware
platform.
• Runs directly on the physical machine (bare-metal).
• The hypervisor is responsible for allocating resources (CPU, memory, I/O) to virtual
machines.
• Provides high performance and low overhead since it bypasses the host OS.
• Ensures strong isolation between VMs.
2. Host VM (Software-based) – Runs as an application on a host OS (e.g., VirtualBox, VMware
Workstation).
A hosted virtual machine runs as an application within an existing operating system, relying
on a host OS to provide access to hardware resources. These VMs are managed using
software-based virtualization platforms.
• Runs on top of a host operating system.
• Uses software-based virtualization techniques (binary translation, dynamic
recompilation).
• Has higher overhead compared to native VMs.
• Provides greater flexibility since it can run on general-purpose systems.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 12

BCS601 Cloud Computing Module-1

3. Hybrid VM – Uses a combination of user-mode and privileged-mode virtualization.

Hybrid VMs combine features of both native and hosted virtualization. They partially
virtualize hardware by running some components in user mode and others in privileged
mode. This architecture optimizes performance by reducing overhead while maintaining
flexibility and ease of management.
• Uses both hardware-assisted and software virtualization techniques.
• The hypervisor runs at the kernel level, but some functions rely on the host OS.
• Balances performance and flexibility for different workloads.

Virtual Machine Operations

• First, the VMs can be multiplexed between hardware machines, as shown in Figure 1.13(a).
• Second, a VM can be suspended and stored in stable storage, as shown in Figure 1.13(b).
• Third, a suspended VM can be resumed or provisioned to a new hardware platform, as shown in Figure 1.13(c).
• Finally, a VM can be migrated from one hardware platform to another, as shown in Figure 1.13(d).

• Multiplexing – Multiple VMs share physical resources.

• Suspension & Migration – VMs can be paused, saved, or migrated across different servers.
• Provisioning – VMs can be dynamically deployed based on workload demand.
Virtual Infrastructure
• Separates physical hardware from applications, enabling flexible resource management.
• Enhances server utilization from 5–15% to 60–80% (as claimed by VMware).

Dr. Sampada K S, Associate professor CSE RNSIT pg. 13

BCS601 Cloud Computing Module-1

1.2.5 Data Center Virtualization for Cloud Computing

Data Center Growth and Cost Breakdown
• 43 million servers worldwide (2010), with utilities (power & cooling) exceeding hardware
costs after three years.
• 60% of data center costs go toward maintenance and management, emphasizing energy
efficiency over raw performance.
Low-Cost Design Philosophy
• Commodity x86 servers & Ethernet replace expensive mainframes & proprietary
networking hardware.
• Software handles fault tolerance, load balancing, and scalability, reducing infrastructure
costs.
Convergence of Technologies Enabling Cloud Computing
1. Virtualization & Multi-core Processors – Enable scalable computing.
2. Utility & Grid Computing – Provide a foundation for cloud computing.
3. SOA, Web 2.0, and Mashups – Facilitate cloud-based service integration.
4. Autonomic Computing & Data Center Automation – Improve efficiency and fault
tolerance.
The Rise of Data-Intensive Computing
• Scientific research, business, and web applications generate vast amounts of data.
• Cloud computing & parallel computing address the data deluge challenge.
• MapReduce & Iterative MapReduce enable scalable data processing for big data and
machine learning applications.
• The convergence of data-intensive computing, cloud platforms, and multicore
architectures is shaping the next generation of distributed computing.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 14

BCS601 Cloud Computing Module-1

The integration of memory, storage, networking, virtualization, and cloud data centers is
transforming distributed systems. By leveraging virtualization, scalable networking, and cloud
computing, modern infrastructures achieve higher efficiency, flexibility, and cost-effectiveness,
paving the way for future exascale computing.

1.3 SYSTEM MODELS FOR DISTRIBUTED AND CLOUD COMPUTING

• Distributed and cloud computing systems are built using large-scale, interconnected autonomous
computer nodes. These nodes are linked through Storage Area Networks (SANs), Local Area
Networks (LANs), or Wide Area Networks (WANs) in a hierarchical manner.
• Clusters: Connected by LAN switches, forming tightly coupled systems with hundreds of
machines.
• Grids: Interconnect multiple clusters via WANs, allowing resource sharing across
thousands of computers.
• P2P Networks: Form decentralized, cooperative networks with millions of nodes, used in
file sharing and content distribution.
• Cloud Computing: Operates over massive data centers, delivering on-demand computing
resources at a global scale.
These systems exhibit high scalability, enabling web-scale computing with millions of
interconnected nodes. Their technical and application characteristics vary based on factors such as
resource sharing, control mechanisms, and workload distribution.
Functionality, Computer Clusters Peer-to-Peer Data/ Computational Cloud Platforms
Applications [10,28,38] Networks [34,46] Grids [ 6,18,51] [1,9,11,12,30]
Architecture,
Network of compute Flexible network Heterogeneous clusters Virtualized cluster of
Network
nodes interconnected by of client machines interconnected by high- servers over data centers
Connectivity,
SAN, LAN, or WAN logically speed network links over via SLA
and Size
hierarchically connected by an selected resource sites
overlay network
Control and Homogeneous nodes Autonomous client Centralized control, Dynamic resource
Resources with distributed control, nodes, free in and out, server- oriented with provisioning of servers,
Management running UNIX or Linux with self-organization authenticated security storage, and networks

Applications High-performance Most appealing to Distributed Upgraded web search,

and Network- computing, business file sharing, supercomputing, global utility computing, and
centric Services search e n g i n e s , and web content delivery, and problem solving, and outsourced Most
services, etc. social networking data center services appealing to business file
sharing, content delivery,
and
social networking
computing services
Representative Google search engine, Gnutella, eMule, TeraGrid, GriPhyN, UK Google App Engine, IBM
Operational SunBlade, IBM Road BitTorrent, Napster, EGEE, D-Grid, Bluecloud, AWS, and
Systems Runner, Cray Ka Za A, Skype, JXTA ChinaGrid, etc Microsoft Gnutella,
XT4, etc. eMule, BitTorrent,
Napster, KaZaA, Skype,
JXTA

Dr. Sampada K S, Associate professor CSE RNSIT pg. 15

BCS601 Cloud Computing Module-1

Clusters of Cooperative Computers

A computing cluster consists of interconnected stand-alone computers which work cooperatively as
a single integrated computing resource.
• In the past, clustered computer systems have demonstrated impressive results in handling heavy
workloads with large data sets.
Cluster Architecture
Server Clusters and System Models for Distributed Computing
1.3.1 Server Clusters and Interconnection Networks
Server clusters consist of multiple interconnected computers using high-bandwidth, low-latency
networks like Storage Area Networks (SANs), Local Area Networks (LANs), and InfiniBand.
These clusters are scalable, allowing thousands of nodes to be connected hierarchically.

• Clusters are connected to the Internet via a VPN gateway, which assigns an IP address to
locate the cluster.
• Each node operates independently, with its own OS, creating multiple system images
(MSI).
• The cluster manages shared I/O devices and disk arrays, providing efficient resource
utilization.
1.3.1.2 Single-System Image (SSI)
An ideal cluster should merge multiple system images into a single-system image (SSI), where all
nodes appear as a single powerful machine.
• SSI is achieved through middleware or specialized OS support, enabling CPU, memory,
and I/O sharing across all cluster nodes.
• Clusters without SSI function as a collection of independent computers rather than a unified
system.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 16

BCS601 Cloud Computing Module-1

1.3.1.3 Hardware, Software, and Middleware Support

• Cluster nodes consist of PCs, workstations, or servers, interconnected using Gigabit
Ethernet, Myrinet, or InfiniBand.
• Linux OS is commonly used for cluster management.
• Message-passing interfaces (MPI, PVM) enable parallel execution across nodes.
• Middleware supports features like high availability (HA), distributed memory sharing
(DSM), and job scheduling.
• Virtual clusters can be dynamically created using virtualization, optimizing resource
allocation on demand.
1.3.1.4 Major Cluster Design Issues

Features Functional Characterization Feasible Implementations

Availability and Support Hardware and software support f o r Failover, failback, check pointing,
sustained HA in cluster rollback recovery, nonstop OS, etc.
Hardware Fault Tolerance Automated failure management to Component redundancy, hot
eliminate all single points of failure swapping, RAID, multiple
power supplies, etc.
Single System Image (SSI) Achieving SSI at functional level with Hardware mechanisms or middleware
hardware a nd software support, support t o achieve DSM at coherent
middleware, or OS extensions c a c h e level
Efficient Communications To reduce me s sa ge -passing system Fast message passing, active
overhead and hide latencies messages, enhanced MPI library, etc.
Cluster-wide Job Using a global job management Application of single-job
system with better scheduling and management systems such as LSF,
Management monitoring Codine, etc.
Dynamic Load Balancing Balancing the workload of all Workload monitoring, process
processing nodes a l o n g with failure migration, job replication and gang
recovery scheduling, etc.
Scalability and Adding more servers to a cluster or Use of scalable interconnect,
adding more clusters to a grid as the performance monitoring, distributed
Programmability workload or data set increases execution environment, and better
software tools
• Lack of a cluster-wide OS limits full resource sharing.
• Middleware solutions provide necessary functionalities like scalability, fault tolerance, and
job management.
• Key challenges include efficient message passing, seamless fault tolerance, high
availability, and performance scalability.
Server clusters are scalable, high-performance computing systems that utilize networked
computing nodes for parallel and distributed processing. Achieving SSI and efficient
middleware support remains a key challenge in cluster computing. Virtual clusters and cloud
computing are evolving to enhance cluster flexibility and resource management.
1.3.2 Grid Computing, Peer-to-Peer (P2P) Networks, and System Models
Grid Computing Infrastructures

Dr. Sampada K S, Associate professor CSE RNSIT pg. 17

BCS601 Cloud Computing Module-1

Grid computing has evolved from Internet and web-based services to enable large-scale
distributed computing. It allows applications running on remote systems to interact in real-time.
1.3.2.1 Computational Grids
• A grid connects distributed computing resources (workstations, servers, clusters,
supercomputers) over LANs, WANs, and the Internet.

• Used for scientific and enterprise applications, including SETI@Home and astrophysics
simulations.
• Provides an integrated resource pool, enabling shared computing, data, and information
services.
1.3.2.2 Grid Families
Design Issues Computational and Data Grids P2P Grids

Grid Applications Reported Distributed supercomputing, Open grid with P2P flexibility, all
National Grid initiatives, etc. resources from client machines
Representative Systems TeraGrid built in US, ChinaGrid in JXTA, FightAid@home,
China, and the e-Science grid SETI@home
built in UK
Development Lessons Learned Restricted user groups, Unreliable user-contributed
middleware bugs, protocols to resources, limited to a few apps
acquire resources
• Computational and Data Grids – Used in national-scale supercomputing projects (e.g.,
TeraGrid, ChinaGrid, e-Science Grid).
• P2P Grids – Utilize client machines for open, distributed computing (e.g., SETI@Home,
JXTA, FightingAID@Home).
• Challenges include middleware bugs, security issues, and unreliable user-contributed
resources.
1.3.3 Peer-to-Peer (P2P) Network Families
P2P systems eliminate central coordination, allowing client machines to act as both servers and
clients.
1.3.3.1 P2P Systems

Dr. Sampada K S, Associate professor CSE RNSIT pg. 18

BCS601 Cloud Computing Module-1

Decentralized architecture with self-organizing peers.

No central authority; all nodes are independent.
Dynamic membership – peers can join and leave freely.

1.3.3.2 Overlay Networks

• Logical connections between peers, independent of the physical network.
• Two types:
o Unstructured overlays – Randomly connected peers, requiring flooding for data
retrieval (high traffic).
o Structured overlays – Use predefined rules for routing and data lookup, improving
efficiency.
1.3.3.3 P2P Application Families
P2P networks serve four main application categories:

Category Examples Challenges

File Sharing Napster, BitTorrent, Gnutella Copyright issues, security concerns

Collaboration PlatformsSkype, MSN, Multiplayer games Privacy risks, spam, lack of trust

Distributed Computing SETI@Home, Genome@Home Security vulnerabilities, selfish nodes

Open P2P Platforms JXTA, .NET, FightingAID@Home

Lack of standardization and security

1.3.3.4 P2P Computing Challenges

• Heterogeneity – Varying hardware, OS, and network configurations.
• Scalability – Must handle growing workloads and distributed resources efficiently.
• Data Location & Routing – Optimizing data placement for better performance.
• Fault Tolerance & Load Balancing – Peers can fail unpredictably.
• Security & Privacy – No central control means increased risk of data breaches and
malware.
P2P networks offer robust and decentralized computing, but lack security and reliability, making
them suitable only for low-security applications like file sharing and collaborative tools.
Both grid computing and P2P networks provide scalable, distributed computing models. While
grids are used for structured, high-performance computing, P2P networks enable decentralized,

Dr. Sampada K S, Associate professor CSE RNSIT pg. 19

BCS601 Cloud Computing Module-1

user-driven resource sharing. Future developments will focus on security, standardization, and
efficiency improvements.
Cloud Computing over the Internet
Cloud computing has emerged as a transformative on-demand computing paradigm, shifting
computation and data storage from desktops to large data centers. This approach enables the
virtualization of hardware, software, and data resources, allowing users to access scalable
services over the Internet.
1.3.4.1 Internet Clouds

• Cloud computing leverages virtualization to dynamically provision resources, reducing

costs and complexity.
• It offers elastic, scalable, and self-recovering computing power through server clusters and
large databases.
• The cloud can be perceived as either a centralized resource pool or a distributed computing
platform.
• Key benefits: Cost-effectiveness, flexibility, and multi-user application support.
1.3.4.2 The Cloud Landscape
Traditional computing systems suffer from high maintenance costs, poor resource utilization, and
expensive hardware upgrades. Cloud computing resolves these issues by providing on-demand
access to computing resources.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 20

BCS601 Cloud Computing Module-1

Three Major Cloud Service Models:

1. Infrastructure as a Service (IaaS)
o Provides computing infrastructure such as virtual machines (VMs), storage, and
networking.
o Users deploy and manage their applications but do not control the underlying
infrastructure.
o Examples: Amazon EC2, Google Compute Engine.
2. Platform as a Service (PaaS)
o Offers a development environment with middleware, databases, and
programming tools.
o Enables developers to build, test, and deploy applications without managing
infrastructure.
o Examples: Google App Engine, Microsoft Azure, AWS Lambda.
3. Software as a Service (SaaS)
o Delivers software applications via web browsers.
o Users pay for access instead of purchasing software licenses.
o Examples: Google Workspace, Microsoft 365, Salesforce.
Cloud Deployment Models:
• Private Cloud – Dedicated to a single organization (e.g., corporate data centers).
• Public Cloud – Hosted by third-party providers for general use (e.g., AWS, Google Cloud).
• Managed Cloud – Operated by a third-party service provider with customized
configurations.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 21

BCS601 Cloud Computing Module-1

• Hybrid Cloud – Combines public and private clouds, optimizing cost and security.
Advantages of Cloud Computing
Cloud computing provides several benefits over traditional computing paradigms, including:
1. Energy-efficient data centers in secure locations.
2. Resource sharing, optimizing utilization and handling peak loads.
3. Separation of infrastructure maintenance from application development.
4. Cost savings compared to traditional on-premise infrastructure.
5. Scalability for application development and cloud-based computing models.
6. Enhanced service and data discovery for content and service distribution.
7. Security and privacy improvements, though challenges remain.
8. Flexible service agreements and pricing models for cost-effective computing.
Cloud computing fundamentally changes how applications and services are developed, deployed,
and accessed. With virtualization, scalability, and cost efficiency, it has become the backbone of
modern Internet services and enterprise computing. Future advancements will focus on security,
resource optimization, and hybrid cloud solutions.
1.4 Software Environments for Distributed Systems and Clouds
This section introduces Service-Oriented Architecture (SOA) and other key software environments
that enable distributed and cloud computing systems. These environments define how
applications, services, and data interact within grids, clouds, and P2P networks.
1.4.1 Service-Oriented Architecture (SOA)
SOA enables modular, scalable, and reusable software components that communicate over a
network. It underpins web services, grids, and cloud computing environments.
1.4.1.1 Layered Architecture for Web Services and Grids
• Distributed computing builds on the OSI model, adding layers for service interfaces,
workflows, and management.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 22

BCS601 Cloud Computing Module-1

• Communication standards include:

o SOAP (Simple Object Access Protocol) – Used in web services.
o RMI (Remote Method Invocation) – Java-based communication.
o IIOP (Internet Inter-ORB Protocol) – Used in CORBA-based systems.
• Middleware tools (e.g., WebSphere MQ, Java Message Service) manage messaging,
security, and fault tolerance.
1.4.1.2 Web Services and Tools
SOA is implemented via two main approaches:
1. Web Services (SOAP-based) – Fully specified service definitions, enabling distributed OS-
like environments.
2. REST (Representational State Transfer) – Simpler, lightweight alternative for web
applications and APIs.
• Web Services provide structured, standardized communication but face challenges in
protocol agreement and efficiency.
• REST is flexible and scalable, better suited for fast-evolving environments.
• Integration of Services – Distributed systems use Remote Method Invocation (RMI) or
RPCs to link services into larger applications.
1.4.1.3 The Evolution of SOA

Dr. Sampada K S, Associate professor CSE RNSIT pg. 23

BCS601 Cloud Computing Module-1

SOA has expanded from basic web services to complex multi-layered ecosystems:
• Sensor Services (SS) – Devices like ZigBee, Bluetooth, GPS, and WiFi collect raw data.
• Filter Services (FS) – Process data before feeding into computing, storage, or discovery
clouds.
• Cloud Ecosystem – Integrates compute clouds, storage clouds, and discovery clouds for
managing large-scale applications.
SOA enables data transformation from raw data → useful information → knowledge → wisdom
→ intelligent decisions.
SOA defines the foundation for web services, distributed systems, and cloud computing. By
integrating sensors, processing layers, and cloud resources, SOA provides a scalable, flexible
approach for modern computing applications. The future of distributed computing will rely on
intelligent data processing, automation, and service-driven architectures.
1.4.1.4 Grids vs. Clouds
• Grids use static resources, whereas clouds provide elastic, on-demand resources via
virtualization.
• Clouds focus on automation and scalability, while grids are better for negotiated
resource allocation.
• Hybrid models exist, such as clouds of grids, grids of clouds, and inter-cloud architectures.

1.4.2 Trends toward Distributed Operating Systems

Traditional distributed systems run independent OS instances on each node, resulting in multiple
system images. A distributed OS manages all resources coherently and efficiently across nodes.
1.4.2.1 Distributed OS Approaches (Tanenbaum's Models)
1. Network OS – Basic resource sharing via file systems (low transparency).
2. Middleware-based OS – Limited resource sharing through middleware extensions (e.g.,
MOSIX for Linux clusters).
3. Truly Distributed OS – Provides single-system image (SSI) with full transparency across
resources.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 24

BCS601 Cloud Computing Module-1

1.4.2.2 Amoeba vs. DCE

• Amoeba (microkernel approach) offers a lightweight distributed OS model.
• DCE (middleware approach) extends UNIX for RPC-based distributed computing.
• MOSIX2 enables process migration across Linux-based clusters and clouds.
1.4.2.3 MOSIX2 for Linux Clusters
• Supports virtualization for seamless process migration across multiple clusters and clouds.
• Enhances parallel computing by dynamically balancing workloads across Linux nodes.
1.4.2.4 Transparency in Programming Environments
• Cloud computing separates user data, applications, OS, and hardware for flexible
computing.
• Users can switch between OS platforms and cloud services without being locked into
specific applications.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 25

BCS601 Cloud Computing Module-1

1.4.3 Parallel and Distributed Programming Models

Distributed computing requires efficient parallel execution models to process large-scale workloads.

Model Description Key Features

MPI (Message- Standard for writing parallel Explicit communication between

Passing Interface) applications on distributed systems processes via message-passing

Map function generates key-value

Web programming model for scalable
MapReduce pairs; Reduce function merges
data processing on large clusters
values

Open-source framework for

Distributed storage (HDFS) and
Hadoop processing vast datasets in business
MapReduce-based computing
and cloud applications

1.4.3.1 Message-Passing Interface (MPI)

• Used for high-performance computing (HPC).
• Programs explicitly send and receive messages for inter-process communication.
1.4.3.2 MapReduce
• Highly scalable parallel model, used in big data processing and search engines.
• Splits workloads into Map (processing) and Reduce (aggregation) tasks.
• Google executes thousands of MapReduce jobs daily for large-scale data analysis.
1.4.3.3 Hadoop
• Open-source alternative to MapReduce, used for processing petabytes of data.
• Scalable, cost-effective, and fault-tolerant, making it ideal for cloud services.
1.4.3.4 Grid Standards and Toolkits
Grids use standardized middleware to manage resource sharing and security.

Standard Function Key Features

Supports heterogeneous computing,

OGSA (Open Grid Defines common grid
security policies, and resource
Services Architecture) services
allocation

Middleware for resource Uses PKI authentication, Kerberos,

Globus Toolkit (GT4)
discovery and security SSL, and delegation policies

Grid computing framework Supports autonomic computing and

IBM Grid Toolbox
for AIX/Linux clusters security management

Dr. Sampada K S, Associate professor CSE RNSIT pg. 26

BCS601 Cloud Computing Module-1

• Distributed OS models are evolving, with MOSIX2 enabling process migration and
resource sharing across Linux clusters.
• Parallel programming models like MPI and MapReduce optimize large-scale computing.
• Cloud computing and grid computing continue to merge, leveraging virtualization and
elastic resource management.
• Standardized middleware (OGSA, Globus) enhances grid security, interoperability, and
automation.
1.5 Performance, Security, and Energy Efficiency
This section discusses key design principles for distributed computing systems, covering
performance metrics, scalability, system availability, fault tolerance, and energy efficiency.
1.5.1 Performance Metrics and Scalability Analysis
Performance is measured using MIPS, Tflops, TPS, and network latency. Scalability is crucial in
distributed systems and has multiple dimensions:
1. Size Scalability – Expanding system resources (e.g., processors, memory, storage) to improve
performance.
2. Software Scalability – Upgrading OS, compilers, and libraries to accommodate larger
systems.
3. Application Scalability – Increasing problem size to match system capacity for cost-
effectiveness.
4. Technology Scalability – Adapting to new hardware and networking technologies while
ensuring compatibility.
1.5.1.3 Scalability vs. OS Image Count
• SMP systems scale up to a few hundred processors due to hardware constraints.
• NUMA systems use multiple OS images to scale to thousands of processors.
• Clusters and clouds scale further by using virtualization.
• Grids integrate multiple clusters, supporting hundreds of OS images.
• P2P networks scale to millions of nodes with independent OS images.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 27

BCS601 Cloud Computing Module-1

1.5.1.4 Amdahl’s Law (Fixed Workload Scaling)

• Speedup in parallel computing is limited by the sequential portion of a program.

• Speedup Formula:
where α is the fraction of the workload that is sequential.
• Even with hundreds of processors, speedup is limited if sequential execution (α) is high.
Problem with Fixed Workload

• In Amdahl’s law, we have assumed the same amount of workload for both sequential and parallel
execution of the program with a fixed problem size or data set. This was called fixed-workload speedup
by Hwang and Xu [14]. To execute a fixed workload on n processors, parallel processing may lead to
a system efficiency defined as follows:

1.5.1.6 Gustafson’s Law (Scaled Workload Scaling)

• Instead of fixing workload size, this model scales the problem to match available
processors.

• Speedup Formula:
• This speedup is known as Gustafson’s law. By fixing the parallel execution time at
level W, the following efficiency expression is obtained:

• More efficient for large clusters, as workload scales dynamically with system size.
1.5.2 Fault Tolerance and System Availability
• High availability (HA) is essential in clusters, grids, P2P networks, and clouds.
• System availability depends on Mean Time to Failure (MTTF) and Mean Time to Repair
(MTTR): Availability=MTTF/(MTTF+MTTR)

• Eliminating single points of failure (e.g., hardware redundancy, fault isolation) improves
availability.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 28

BCS601 Cloud Computing Module-1

• P2P networks are highly scalable but have low availability due to frequent peer failures.
• Grids and clouds offer better fault isolation and thus higher availability than traditional
clusters.
• Scalability and performance depend on resource expansion, workload distribution, and
parallelization.
• Amdahl’s Law limits speedup for fixed workloads, while Gustafson’s Law optimizes
large-scale computing.
• High availability requires redundancy, fault tolerance, and system design improvements.
• Clouds and grids balance scalability and availability better than traditional SMP or
NUMA systems.
Network Threats, Data Integrity, and Energy Efficiency
This section highlights security challenges, energy efficiency concerns, and mitigation strategies
in distributed computing systems, including clusters, grids, clouds, and P2P networks.
1.5.3 Network Threats and Data Integrity
Distributed systems require security measures to prevent cyberattacks, data breaches, and
unauthorized access.
1.5.3.1 Threats to Systems and Networks

• Loss of Confidentiality – Due to eavesdropping, traffic analysis, and media scavenging.

• Loss of Integrity – Caused by penetration attacks, Trojan horses, and unauthorized
access.
• Loss of Availability – Denial of Service (DoS) and resource exhaustion disrupt system
operation.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 29

BCS601 Cloud Computing Module-1

• Improper Authentication – Allows attackers to steal resources, modify data, and conduct
replay attacks.
1.5.3.2 Security Responsibilities
Security in cloud computing is divided among different stakeholders based on the cloud service
model:
• SaaS: Cloud provider handles security, availability, and integrity.
• PaaS: Provider manages integrity and availability, while users control confidentiality.
• IaaS: Users are responsible for most security aspects, while providers ensure availability.
1.5.3.3 Copyright Protection
• Collusive piracy in P2P networks allows unauthorized file sharing.
• Content poisoning and timestamped tokens help detect piracy and protect digital rights.
1.5.3.4 System Defense Technologies
Three generations of network security have evolved:
1. Prevention-based – Access control, cryptography.
2. Detection-based – Firewalls, intrusion detection systems (IDS), Public Key Infrastructure
(PKI).
3. Intelligent response systems – AI-driven threat detection and response.
1.5.3.5 Data Protection Infrastructure
• Trust negotiation ensures secure data sharing.
• Worm containment & intrusion detection protect against cyberattacks.
• Cloud security responsibilities vary based on the service model (SaaS, PaaS, IaaS).

1.5.4 Energy Efficiency in Distributed Computing

Distributed systems must balance high performance with energy efficiency due to increasing power
costs and environmental impact.
1.5.4.1 Energy Consumption of Unused Servers
• Many servers are left powered on but idle, leading to huge energy waste.
• Global energy cost of idle servers: $3.8 billion annually, with 11.8 million tons of CO₂
emissions.
• IT departments must identify underutilized servers to reduce waste.
1.5.4.2 Reducing Energy in Active Servers

Dr. Sampada K S, Associate professor CSE RNSIT pg. 30

BCS601 Cloud Computing Module-1

Energy consumption can be managed across four layers (Figure 1.26):

1. Application Layer – Optimize software to balance performance and energy consumption.

2. Middleware Layer – Smart task scheduling to reduce unnecessary computations.
3. Resource Layer – Use Dynamic Power Management (DPM) and Dynamic Voltage-
Frequency Scaling (DVFS).
4. Network Layer – Develop energy-efficient routing algorithms and optimize bandwidth
usage.
1.5.4.3 Dynamic Voltage-Frequency Scaling (DVFS)
• Reduces CPU voltage and frequency during idle times to save power.
o Formula for Energy Consumption in CMOS Circuits:

o Lowering voltage and frequency significantly reduces energy usage.

• Potential savings: DVFS can cut power consumption while maintaining performance.
• Energy efficiency is critical due to high costs and environmental impact.
• Techniques like DPM and DVFS can significantly reduce power consumption without
compromising performance.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 31

BCS601 Cloud Computing Module-1

Dr. Sampada K S, Associate professor CSE RNSIT pg. 32

 BCS601 Cloud Computing Module-3

Module 2
Virtual Machines and Virtualization of Clusters and Data Centers: Implementation Levels of
Virtualization, Virtualization Structure/Tools and Mechanisms, Virtualization of CPU/Memory
and I/O devices, Virtual Clusters and Resource Management, Virtualization for Data Center
Automation.

Textbook 1: Chapter 3: 3.1 to 3.5

3.1 Implementation Levels of Virtualization

Virtualization: Virtualization is the process of creating virtual instances of computing

resources—such as servers, storage, networks, or operating systems—on a single physical
machine. It allows multiple virtual environments to run independently on the same hardware,
improving resource utilization, scalability, and flexibility.

Difference Between Traditional Computers and Virtualized Computers

Feature Traditional Computer Virtualized Computer

Hardware Dedicated to a single OS and Multiple virtual instances share the same
Usage applications. physical hardware.
Operating Runs one OS per physical Can run multiple OSes (Windows,
System machine. Linux, etc.) on the same machine.
Maximizes hardware efficiency by
Resource Often underutilized, as
sharing CPU, RAM, and storage among
Utilization resources are fixed.
VMs.
All applications share the same Each virtual instance (VM or container)
Isolation
OS. is isolated from others.
Adding resources requires new Can dynamically allocate more resources
Scalability
physical machines. without new hardware.
If one application crashes, it can Virtual instances are independent,
Security
affect the whole system. reducing the impact of failures.
Manual setup and configuration Can quickly deploy and clone
Deployment
per machine. VMs/containers.
Requires individual system Easier to maintain, as snapshots and
Maintenance
updates and backups. templates allow fast recovery.
Higher hardware costs, as each Lower costs by running multiple virtual
Cost
OS requires a separate machine. instances on fewer physical machines.
Limited flexibility; one system Can run different OSes and applications
Flexibility
per hardware. simultaneously.

Example Scenarios

Traditional Computer: A single physical server running Windows, dedicated to one

task, such as hosting a database.
Virtualized Computer: A physical server running multiple VMs—one with Windows
for the database, another with Linux for a web server, and another for testing.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 1

BCS601 Cloud Computing Module-3

Benefits of Virtualization

1. Better Resource Utilization – Maximizes hardware efficiency by running multiple

VMs on a single machine.
2. Cost Savings – Reduces hardware, power, and maintenance costs.
3. Scalability & Flexibility – Easily scale resources as needed.
4. Improved Security & Isolation – VMs are isolated from each other, reducing
security risks.
5. Disaster Recovery – Snapshots and backups allow quick recovery in case of failures.

Virtualization Layer

The virtualization layer is a software layer that abstracts physical hardware resources (CPU,
memory, storage, network, etc.) and presents them as virtual resources to applications and
operating systems. It acts as a bridge between the physical hardware and virtual instances,
ensuring proper allocation, isolation, and management of resources.

Role of the Virtualization Layer

• Abstraction: Hides the complexity of hardware, allowing virtual instances to operate

independently.
• Isolation: Ensures that virtual instances (VMs or containers) don’t interfere with one
another.
• Resource Allocation: Dynamically allocates resources based on demand, ensuring
efficient usage.
• Compatibility: Enables different operating systems and applications to coexist on the
same physical machine.
• Portability: Allows virtual machines or containers to be easily moved across physical
hosts.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 2

BCS601 Cloud Computing Module-3

Benefits of the Virtualization Layer

• Efficient Use of Resources: Consolidates multiple workloads onto fewer physical
machines.
• Scalability: Enables dynamic allocation of resources to meet changing demands
• Improved Security: Isolates virtual instances to minimize security risks.
• Flexibility: Supports diverse operating systems and applications on the same
hardware.
• Disaster Recovery: Facilitates backups, snapshots, and migrations for rapid recovery.

What is a Virtual Machine Monitor (VMM)?

• A Virtual Machine Monitor (VMM), also called a hypervisor, is a software layer or

hardware component responsible for managing and running virtual machines (VMs)
on a host system. It provides the necessary abstraction and resource allocation to
allow multiple virtual machines to share the same physical hardware.

Difference between Virtualization layer and VMM

Aspects VMM Virtualization Layer

Definition . A component (software or hardware) The broader abstraction layer that
that creates, manages, and runs virtual virtualizes resources (e.g., hardware,
machines by abstracting hardware. OS, applications, or networks).

Directly responsible for managing Encompasses the VMM along with

Role
VMs and their access to hardware other components like APIs, drivers,
resources. and management tools.
Scope Focused specifically on running and Includes hardware, VMM, virtual
managing virtual machines. networking, storage, and resource
abstraction tools.

Includes hypervisors (like VMM),

Examples VMware ESXi, Microsoft Hyper-V, SDN (software-defined networking),
KVM. and SDS (software-defined storage).
Primary - Creates and runs virtual machines. - Abstracts all physical resources.
Function - Provides resource isolation and - Enables dynamic management of
sharing for VMs. hardware, networks, and applications.

Focuses on hardware abstraction for Operates at multiple levels, including

Abstraction virtual machines. hardware, operating systems, storage,
Level and networks.
Key - Virtual CPUs. - VMM.
Components - Virtual memory management. - Management tools (e.g., VMware
- Device emulation. vCenter).
- Virtual storage. - Virtual switches (e.g., Open
vSwitch).
- APIs and drivers.
Example - Running multiple virtual machines on - Managing data center resources using
Scenarios a single physical server. platforms like VMware vSphere.
- Isolating workloads using hypervisors - Supporting cloud platforms (e.g.,
like Xen or KVM. AWS or Azure).

Dr. Sampada K S, Associate professor CSE RNSIT pg. 3

BCS601 Cloud Computing Module-3

3.1.1 Levels of Virtualization Implementation

Virtualization is done in five abstract levels ranging from hardware to applications

3.1.1.1 Instruction Set Architecture Level:

At the ISA level, virtualization primarily focuses on enabling the execution of
instructions written for one architecture on a host machine with a different architecture.
The process is achieved through instruction set emulation, which can either interpret
instructions one-by-one or dynamically translate blocks of instructions to improve
efficiency.
Note: Emulation refers to the process of mimicking or imitating the behavior of one
system, device, or software by another system. In the context of computing, emulation
allows one machine or software to replicate the functions of another, enabling
programs, applications, or operating systems designed for one platform to run on a
different platform.

How ISA Virtualization Works

1. Instruction Emulation:
o The source ISA (e.g., MIPS) is emulated on the target ISA (e.g., x86) through
a software layer.
o The software layer interprets or translates the source instructions into target
machine instructions.
2. Virtual ISA (V-ISA):

Dr. Sampada K S, Associate professor CSE RNSIT pg. 4

BCS601 Cloud Computing Module-3

o A virtual instruction set architecture acts as an abstraction, making it

possible for various source ISAs to execute on the same host machine by
translating and optimizing the instructions.
o A software layer, added to the compiler, facilitates this translation and
manages differences between ISAs.

Key Techniques in ISA Virtualization

1. Code Interpretation:
o Process: An interpreter program translates source instructions into host
(native) instructions one-by-one during execution.
o Characteristics:
▪ Simple to implement.
▪ High overhead due to the need to process each instruction individually.
o Performance: Slow, as each source instruction may require tens or even
hundreds of native instructions to execute.
2. Dynamic Binary Translation:
o Process:
▪ Instead of interpreting instructions one-by-one, this method translates
blocks of source instructions (basic blocks, traces, or superblocks)
into target instructions.
▪ The translated blocks are cached, so subsequent executions do not need
re-translation.
o Characteristics:
▪ Faster than interpretation due to caching and reuse of translated
instructions.
▪ Optimization opportunities arise from analyzing multiple instructions
in a block.
o Performance: Significantly better than interpretation but requires more
complex implementation.
3. Binary Translation and Optimization:
o Purpose: Enhance performance and reduce the overhead of translation.
o Methods:
▪ Static Binary Translation: Translates the entire binary code before
execution, which avoids runtime translation but can miss opportunities
for runtime optimizations.
▪ Dynamic Binary Translation: Translates instructions at runtime,
enabling better adaptability to runtime conditions.
▪ Dynamic Optimizations: Includes reordering, inlining, and loop
unrolling to improve the efficiency of translated code.

ISA-level virtualization via instruction set emulation opens immense possibilities for
running diverse workloads across platforms, supporting legacy systems, and enabling
hardware independence. The shift from simple interpretation to more advanced
techniques like dynamic binary translation and optimizations has significantly
improved its performance and applicability, making it a key enabler for cross-platform
software execution.

Advantages of ISA Virtualization

1. Legacy Code Support:

Dr. Sampada K S, Associate professor CSE RNSIT pg. 5

BCS601 Cloud Computing Module-3

o Allows legacy binary code (e.g., for MIPS or PowerPC) to run on newer
hardware (e.g., x86 or ARM).
o Extends the lifespan of legacy software without needing hardware redesign.
2. Cross-Architecture Compatibility:
o Applications can run on hardware with different ISAs, enhancing portability
and flexibility.
3. Facilitates Hardware Upgrades:
o Software compiled for older processors can run on modern processors, easing
transitions to new architectures.
4. Enables Platform Independence:
o Virtual ISAs abstract the underlying hardware, enabling software to operate
across heterogeneous platforms.

Challenges and Limitations

1. Performance Overhead:
o Emulating an ISA on another is inherently slower due to instruction-by-
instruction interpretation or translation.
o Dynamic binary translation improves performance but still adds runtime
overhead.
2. Complexity:
o Implementing dynamic binary translation and optimizations requires advanced
techniques and significant development effort.
3. Scalability:
o Supporting highly diverse ISAs can become challenging, especially when
optimizing performance for multiple architectures.

3.1.1.2 Hardware abstraction level:

Hardware-level virtualization is a method of virtualization implemented directly on
the physical hardware to enable the creation and management of virtual machines
(VMs). It abstracts and virtualizes a computer's physical resources—such as CPU,
memory, and input/output (I/O) devices—allowing multiple operating systems to run
concurrently on the same physical machine.

Key Features of Hardware-Level Virtualization

1. Bare-Metal Hypervisors:
o A hypervisor (Type 1) operates directly on the hardware without requiring an
underlying host operating system.
o It creates and manages virtual hardware environments for virtual machines.
2. Resource Virtualization:
o Virtualizes hardware components such as CPUs, memory, network interfaces,
and storage.
o VMs appear to have dedicated hardware, even though they share the
underlying physical resources.
3. Improved Hardware Utilization:
o Allows multiple users or workloads to share the same hardware, increasing
resource utilization and efficiency.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 6

BCS601 Cloud Computing Module-3

4. Isolation:
o Each VM operates in isolation, meaning that the failure or compromise of one
VM does not affect others.

Advantages of Hardware-Level Virtualization

1. High Performance:
o Since the hypervisor runs directly on hardware, it minimizes overhead and
provides near-native performance for VMs.
2. Scalability:
o Easily supports multiple VMs, enabling efficient use of physical server
resources.
3. Fault Isolation:
o Problems in one VM (e.g., OS crashes or software bugs) do not impact other
VMs or the host system.
4. Versatility:
o Supports running different operating systems or environments on the same
physical hardware.

3.1.1.3 Operating System Level

Operating System (OS) level virtualization is a type of virtualization that operates at the OS
kernel layer, creating isolated environments called containers or virtual environments within
a single instance of an operating system. This approach allows multiple isolated user spaces to
run on the same physical hardware while sharing the same operating system kernel.

Key Features of OS-Level Virtualization

1. Single OS Kernel:
o All containers share the same underlying OS kernel, eliminating the need for
separate kernels for each environment.
o More lightweight compared to traditional hardware-level virtualization since
there's no need to emulate hardware.
2. Isolated Environments (Containers):
o Containers behave like independent servers, with their own libraries, binaries,
and configuration files.
o Processes running inside one container are isolated from processes in other
containers.
3. Efficient Resource Utilization:
o OS-level virtualization efficiently shares hardware resources like CPU,
memory, and storage across containers.
o Reduces overhead compared to full virtualization, as there is no need for a
hypervisor or virtual hardware.

Advantages of OS-Level Virtualization

1. Lightweight:Containers consume fewer resources compared to traditional VMs

because they do not require a full guest operating system.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 7

BCS601 Cloud Computing Module-3

2. High Performance:Since all containers share the same OS kernel, there is minimal
overhead, resulting in near-native performance.
3. Scalability:Containers can be created, started, stopped, and destroyed quickly, making
them ideal for dynamic environments.
4. Isolation:Although containers share the same kernel, they provide process and file
system isolation, ensuring that one container does not interfere with another.
5. Ease of Deployment:Containers package applications with their dependencies, making
them portable across different environments.

Disadvantages of OS-Level Virtualization

1. Single OS Limitation:Since all containers share the same kernel, they must use the
same operating system. For example, you cannot run a Windows container on a Linux
host.
2. Weaker Isolation:Compared to hardware-level virtualization, OS-level virtualization
provides less isolation. If the kernel is compromised, all containers are at risk.
3. Compatibility Issues:Applications that require specific kernel modules or features not
supported by the shared kernel may face compatibility challenges.

3.1.1.4 Library Support Level:

Library-level virtualization focuses on virtualizing the interface between an application
and the operating system by intercepting or emulating the API (Application
Programming Interface) calls. This type of virtualization is lightweight and primarily
targets specific application needs, rather than virtualizing entire operating systems or
hardware.

Key Features of Library-Level Virtualization

1. API Hooks:
o Applications typically interact with the operating system via APIs exported by
user-level libraries.
o Library-level virtualization works by intercepting API calls and redirecting
them to virtualized implementations.
2. Controlled Communication:
o Virtualization happens by managing the communication link between the
application and the underlying system.
o This approach avoids direct interaction with the operating system and replaces
it with controlled, virtualized responses.
3. Application-Specific Virtualization:
o Focused on enabling specific features or compatibility, such as supporting
applications from one environment on another.

How Library-Level Virtualization Works

• Applications are written to use standard library calls for their functionality, such as file
access, networking, or graphics.
• Library-level virtualization intercepts these calls (using API hooks) and replaces the
original functionality with emulated or redirected behavior.

Library-level virtualization is a targeted and lightweight form of virtualization that

focuses on emulating user-level APIs to enable application compatibility and

Dr. Sampada K S, Associate professor CSE RNSIT pg. 8

BCS601 Cloud Computing Module-3

portability. It plays a critical role in scenarios like running software across platforms,
leveraging hardware features in virtualized environments, and extending the life of
legacy applications. While it may not provide the full isolation or flexibility of OS- or
hardware-level virtualization, its efficiency and simplicity make it invaluable for
specific use cases.

3.1.1.5 User-Application Level:

User-Application-level virtualization refers to the virtualization of individual

applications, treating each application as a virtualized entity, often running as an
isolated process. It is also known as process-level virtualization since the
virtualization focuses on isolating and managing application processes rather than
entire systems or hardware.

Key Features of User-Application-Level Virtualization

1. High-Level Language (HLL) Virtual Machines:

o A common form of application-level virtualization involves High-Level
Language (HLL) Virtual Machines, which provide an abstraction layer for
running applications written in a specific language.
o The virtualization layer operates as an application program on top of the OS,
exporting a virtual machine abstraction.
o Examples:
▪ Java Virtual Machine (JVM): Executes Java bytecode, enabling Java
applications to run on any platform with a JVM implementation.
▪ Microsoft .NET Common Language Runtime (CLR): Runs programs
written in .NET-supported languages like C# and VB.NET.
2. Application Isolation and Sandboxing:
o Another form of user-application level virtualization involves isolating
applications from the host OS and other applications.
o This approach wraps the application in a layer that prevents it from interfering
with system resources or other applications.
3. Application Streaming:
o Applications are deployed as self-contained packages that can be executed in a
virtualized environment without traditional installation processes.
o The application runs in an isolated sandbox, reducing dependency on the host
OS.

Approaches to User-Application Level Virtualization

1. High-Level Language (HLL) Virtualization:

2. Application Isolation and Sandboxing:
3. Application Streaming

Advantages of User-Application Level Virtualization

1. Cross-Platform Compatibility:
o Applications written for an abstract VM (e.g., JVM, CLR) can run on any
system with the corresponding VM implementation.
2. Improved Security:

Dr. Sampada K S, Associate professor CSE RNSIT pg. 9

BCS601 Cloud Computing Module-3

o Applications are isolated from the host OS and other applications, reducing the
risk of system compromise or interference.
3. Simplified Deployment:
o Applications can be distributed as self-contained packages, eliminating the need
for complex installation procedures or OS-level dependencies.
4. Resource Efficiency:
o Compared to hardware- or OS-level virtualization, application-level
virtualization has lower overhead as it focuses only on individual processes.
5. Portability:
o Virtualized applications can be easily moved between systems or platforms.

Disadvantages of User-Application Level Virtualization

1. Performance Overhead:
o Running applications in a virtualized environment may introduce some latency
compared to native execution.
2. Limited Scope:
o Unlike OS- or hardware-level virtualization, application-level virtualization
cannot provide a full OS environment or support multiple users.
3. Compatibility Challenges:
o Not all applications can be easily virtualized, especially those with tight
integration with the underlying OS or hardware.

3.1.1.6 Relative Merits of Different Approaches:

In the above table, the column headings correspond to four technical merits. “Higher
Performance” and “Application Flexibility” are self-explanatory. “Implementation
Complexity” implies the cost to implement that particular virtualization level. “Application
Isolation” refers to the effort required to isolate resources committed to different VMs.

The number of X’s in the table cells reflects the advantage points of each implementation level.
Five X’s implies the best case and one X implies the worst case. Overall, hardware and OS
support will yield the highest performance. However, the hardware and application levels are
also the most expensive to implement. User isolation is the most difficult to achieve. ISA
implementation offers the best application flexibility.

3.1.2 VMM Design Requirements and providers

Dr. Sampada K S, Associate professor CSE RNSIT pg. 10

BCS601 Cloud Computing Module-3

Hardware-level virtualization adds a layer, the Virtual Machine Monitor (VMM), between
the hardware and operating systems. The VMM manages hardware resources and allows
multiple operating systems to run simultaneously on a single hardware setup by virtualizing
components like the CPU. A VMM must meet three key requirements:

1. Provide an environment essentially identical to the original machine.

2. Minimize performance overhead.
3. Maintain complete control over system resources.

While VMMs ensure functional equivalence to physical hardware, two performance-related

exceptions are allowed: resource availability limitations (when multiple VMs share resources)
and timing dependencies (due to software layers and concurrent VMs).

Efficiency is crucial for VMMs, as slow emulators or interpreters are unsuitable for real
machines. To ensure performance, most virtual processor instructions should execute directly
on physical hardware without VMM intervention.

The VMM manages resources by allocating them to programs, restricting unauthorized access,
and regaining control when needed. However, implementing VMMs can be challenging on
certain processor architectures (e.g., x86), where privileged instructions cannot always be
trapped. Processors not designed for virtualization may require hardware modifications to meet
VMM requirements, a method known as hardware-assisted virtualization.

Key Observations:

• VMware Workstation supports a wide range of guest operating systems and uses full
virtualization.
• VMware ESX Server eliminates a host OS, running directly on hardware with para-
virtualization.
• Xen supports diverse host OSs and uses a hypervisor-based architecture.
• KVM runs exclusively on Linux hosts and supports para-virtualization for multiple
architectures.

3.1.3 Virtualization Support at the OS level:

Dr. Sampada K S, Associate professor CSE RNSIT pg. 11

BCS601 Cloud Computing Module-3

Cloud computing, enabled by VM technology, shifts the cost and responsibility of managing
computational centers to third parties, resembling the role of banks. While transformative, it
faces two significant challenges:

1. Dynamic Resource Scaling: Adapting to varying computational demands, where tasks

may require a single CPU at times and hundreds of CPUs at others.
2. Slow VM Instantiation: Current methods for creating VMs, such as fresh boots or
replicating templates, are slow and do not account for the current application state.

To address these challenges and enhance cloud computing efficiency, significant research and
development are needed.

3.1.3.1 Why OS-Level Virtualization?

OS-level virtualization addresses the challenges of hardware-level virtualization, such as slow

initialization, storage issues due to repeated content in VM images, slow performance, and the
need for para-virtualization or hardware modifications. It introduces a virtualization layer
within the operating system to partition physical resources and enable multiple isolated virtual
environments (VEs), also called containers or Virtual Private Systems (VPS).

VEs share the same OS kernel but appear as independent servers to users, each with its own
processes, file system, user accounts, network settings, and more. This approach, known as
single-OS image virtualization, is an efficient alternative to hardware-level virtualization.
Figure 3.3 illustrates operating systems virtualization from the point of view of a machine
stack.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 12

BCS601 Cloud Computing Module-3

3.1.3.2 Advantages of OS Extensions

OS-level virtualization offers two key advantages over hardware-level virtualization:

1. Efficiency and Scalability: OS-level VMs have low startup/shutdown costs, minimal
resource requirements, and high scalability.
2. State Synchronization: VMs can synchronize state changes with the host environment
when needed.

These benefits are enabled by two mechanisms:

• All VMs share a single OS kernel, reducing overhead.

• The virtualization layer allows VMs to access host resources without modifying them.

In cloud computing, these features address the slow initialization of hardware-level VMs and
their inability to account for the current application state.

3.1.3.3 Disadvantages of OS Extensions

The primary disadvantage of OS-level virtualization is that all VMs on a single container must
belong to the same operating system family. For example, a Linux-based container cannot run
a Windows OS. This limitation challenges its usability in cloud computing, where users may
prefer different operating systems.

To implement OS-level virtualization:

• A virtualization layer is added to the OS to partition hardware resources for isolated

virtual environments (VMs).
• VMs share the same OS kernel, and access requests are redirected to their respective
resource partitions.
• Virtual root directories can be created using methods like the chroot command in UNIX
systems.

Two methods for managing resource partitions are:

1. Duplicating resources for each VM: This incur high resource costs and overhead.
2. Sharing most resources with the host and creating private copies on demand: This
is more efficient and commonly used.

Due to its limitations and overhead in some scenarios, OS-level virtualization is often
considered a secondary choice compared to hardware-assisted virtualization.

3.1.3.4 Virtualization on Linux or Windows Platforms

OS-level virtualization systems are predominantly Linux-based, while Windows-based

platforms are still in the research phase. The Linux kernel provides an abstraction layer that
allows software to interact with resources without needing hardware-specific details. However,
new hardware may require updates or patches to the Linux kernel for extended functionality.

Key points about Linux OS-level virtualization:

Dr. Sampada K S, Associate professor CSE RNSIT pg. 13

BCS601 Cloud Computing Module-3

• Most Linux platforms are not tied to a specific kernel, enabling a host to run multiple
VMs simultaneously on the same hardware.
• Linux-based tools, such as Linux vServer and OpenVZ, support running applications
from other platforms through virtualization.
• On Windows, FVM is a specific tool developed for OS-level virtualization on the
Windows NT platform.

Example 3.1: Virtualization Support for the Linux Platform

OpenVZ is an open-source, container-based virtualization tool for Linux platforms. It enables

the creation of virtual private servers (VPSes) that operate like independent Linux servers.
OpenVZ modifies the Linux kernel to provide features like virtualization, resource
management, isolation, and checkpointing.

Key Features:

1. Isolation:
o Each VPS has its own files, user accounts, process tree, virtual network, virtual
devices, and interprocess communication (IPC) mechanisms.
2. Resource Management:
o Disk Allocation: Two levels:
▪ First level: The OpenVZ server administrator assigns disk space limits
to each VM.
▪ Second level: VM administrators manage disk quotas for users and
groups.
o CPU Scheduling:
▪ First level: OpenVZ's scheduler allocates time slices based on virtual
CPU priority and limits.
▪ Second level: Uses the standard Linux CPU scheduler.
o Resource Control: OpenVZ has ~20 parameters to control VM resource usage.
3. Checkpointing and Live Migration:
o Allows saving the complete state of a VM to a disk file, transferring it to another
machine, and restoring it there.
o The process takes only a few seconds, although network connection re-
establishment causes minor delays.

Advantages:

• Efficient resource management and isolation.

• Quick VM migration with minimal downtime.

Challenges:

• Delay in processing due to migrating active network connections.

Table 3.3 provides a summary of OS-level virtualization tools

Dr. Sampada K S, Associate professor CSE RNSIT pg. 14

BCS601 Cloud Computing Module-3

3.1.4 Middleware Support for Virtualization

Library-level virtualization, also referred to as user-level Application Binary Interface (ABI)

or API emulation, enables execution environments to run specific programs on a platform
without the need to virtualize the entire operating system. The key functions include API call
interception and remapping.

Overview of Library-Level Virtualization Systems/ Middleware and Library Support for

Virtualization:

1. WABI:
o Middleware that translates Windows system calls into Solaris system calls,
allowing Windows applications to run on Solaris systems.
2. Lxrun:
o A system call emulator enabling Linux applications designed for x86 hosts to
run on UNIX systems.
3. WINE:
o Provides library support to virtualize x86 processors, enabling Windows
applications to run on UNIX-based systems.
4. Visual MainWin:
o A compiler support system that allows developers to use Visual Studio to
create Windows applications capable of running on some UNIX hosts.
5. vCUDA:
o A virtualization solution for CUDA, enabling applications requiring GPU
acceleration to utilize GPU resources remotely. (Discussed in detail in Example
3.2.)

Key Benefits:

• Enables cross-platform application execution without the overhead of a full virtualized

operating system.
• Useful for running alien programs efficiently on different platforms.

Challenges:

• Performance depends on the accuracy and efficiency of API call remapping.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 15

BCS601 Cloud Computing Module-3

• Limited to specific program compatibility based on system call emulation.

Example 3.2 vCUDA for Virtualization of General-Purpose GPUs

vCUDA is a virtualization solution designed to enable CUDA applications to run on guest

operating systems (OSes) by virtualizing the CUDA library. It allows compute-intensive
applications to leverage GPU acceleration in a virtualized environment.

Key Features of vCUDA:

1. Purpose:
o Virtualizes the CUDA library for guest OSes, enabling CUDA applications to
execute GPU-based tasks indirectly through the host OS.
2. Architecture:
o Follows a client-server model with three main components:
▪ vCUDA Library:
▪ Resides in the guest OS as a substitute for the standard CUDA
library.
▪ Intercepts and redirects API calls to the host OS.
▪ Manages virtual GPUs (vGPUs).
▪ Virtual GPU (vGPU):
▪ Abstracts GPU hardware, provides a uniform interface, and
manages device memory allocation.
▪ Tracks and stores CUDA API flow.
▪ vCUDA Stub:
▪ Resides in the host OS.
▪ Receives and interprets requests from the guest OS.
▪ Creates execution contexts for CUDA API calls and manages the
physical GPU resources.
3. Functionality of vGPU:
o Abstracts the GPU structure, giving applications a consistent view of hardware.
o Handles memory allocation by mapping virtual addresses in the guest OS to real
device memory in the host OS.
o Stores the flow of CUDA API calls for proper execution.
4. Workflow:
o CUDA applications on the guest OS send API calls to the vCUDA library.
o The vCUDA library redirects these calls to the vCUDA stub on the host OS.
o The vCUDA stub processes the requests, executes them on the physical GPU,
and returns results to the guest OS.

Benefits of vCUDA:

• Enables GPU acceleration in virtualized environments without directly running

CUDA on hardware-level VMs.
• Efficiently handles resource allocation and execution across guest and host OSes.

Challenges:

• Relies heavily on the client-server architecture and the efficiency of API call
redirection.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 16

BCS601 Cloud Computing Module-3

• Performance may depend on the complexity of GPU tasks and overhead of

virtualization.

3.2 Virtualization Structures / Tools And Mechanisms

There are three typical classes of VM architectures, differentiated by the placement of the
virtualization layer in the system stack. Virtualization transforms a machine’s architecture by
inserting a virtualization layer between the hardware and the operating system. This layer
converts real hardware into virtual hardware, enabling different operating systems (e.g., Linux
and Windows) to run simultaneously on the same physical machine.

Classes of VM Architectures:

1. Hypervisor Architecture (or VMM):

o The hypervisor operates directly on the hardware layer, acting as the
virtualization layer.
o It manages the hardware resources and virtualizes them for guest operating
systems.
o Supports multiple OS instances on the same hardware efficiently.
2. Para-Virtualization:
o A modified version of the guest OS collaborates with the virtualization layer.
o Offers better performance compared to traditional hypervisors by reducing the
overhead of virtualization.
3. Host-Based Virtualization:
o The virtualization layer runs as an application on top of an existing host
operating system.
o Easier to implement but less efficient due to the additional layer of the host OS.

Key Points:

• The virtualization layer is crucial for translating real hardware into virtual hardware.
• These architectures enable flexibility in running multiple operating systems on the same
machine.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 17

BCS601 Cloud Computing Module-3

• Hypervisors (or VMMs) and other approaches vary in performance, complexity, and
implementation.

3.2.1 Hypervisor and Xen Architecture

The hypervisor (or Virtual Machine Monitor, VMM) enables hardware-level virtualization by
acting as an intermediate layer between physical hardware (e.g., CPU, memory, disk, network
interfaces) and the operating systems (OS). It facilitates the creation of virtual resources that
guest OSes and applications can utilize.

Types of Hypervisor Architectures:

1. Micro-Kernel Hypervisor:
o Only includes essential and unchanging functionalities, such as physical
memory management and processor scheduling.
o Device drivers and other changeable components are kept outside the
hypervisor.
o Examples: Microsoft Hyper-V.
o Advantages: Smaller code size, reduced complexity, and easier maintainability.
2. Monolithic Hypervisor:
o Integrates all functionalities, including device drivers, within the hypervisor
itself.
o Examples: VMware ESX for server virtualization.
o Advantages: Comprehensive functionality but with a larger codebase and
potential complexity.

Key Features of a Hypervisor:

• Supports virtualized access to physical hardware through hypercalls for guest OSes
and applications.
• Converts physical devices into virtual resources for use by virtual machines (VMs).
• Plays a critical role in resource management and scheduling for multiple VMs.

These architectures allow efficient use of physical hardware while enabling multiple OSes to
run simultaneously.

3.2.1.1 The Xen Architecture

Xen is an open-source microkernel hypervisor developed at Cambridge University. It acts as a

virtualization layer between hardware and the operating system, enabling multiple guest OS
instances to run simultaneously. The core components of a Xen system include the hypervisor,
kernel, and applications.

A key feature of Xen is Domain 0 (Dom0), a privileged guest OS that manages hardware
access and resource allocation for other guest domains (Domain U). Since Dom0 controls the
entire system, its security is critical. If compromised, an attacker can control all virtual
machines.

Xen allows users to manage VMs flexibly creating, copying, migrating, and rolling back
instances. However, this flexibility also introduces security risks, as VMs can revert to previous

Dr. Sampada K S, Associate professor CSE RNSIT pg. 18

BCS601 Cloud Computing Module-3

states, potentially reintroducing vulnerabilities. Unlike traditional machines with a linear

execution timeline, Xen VMs form a tree-like execution state, enabling multiple instances
and rollbacks, which benefits system management but also creates security challenges.

Note: A key feature of Xen is Domain 0 (Dom0), a privileged virtual machine responsible for
managing hardware, I/O operations, and other guest VMs (Domain U). Dom0 is the first OS
to load and has direct hardware access, allowing it to allocate resources and manage devices
for unprivileged guest domains.

Control, I/O (Domain 0) Guest domain Guest domain

Application

Application

Application

Application

Application

Application

Application

Application

Application

Application
Domain0 XenoLinux XenoWindows

XEN (Hypervisor)

Hardware devices

FIGURE 3.5
The Xen architecture’s special domain 0 for control and I/O, and several guest domains for user applications.

3.2.2 Binary Translation with Full Virtualization

Hardware virtualization can be categorized into full virtualization and host-based

virtualization based on implementation technologies.

Full Virtualization

• Full virtualization allows a guest OS to run without modification by using binary

translation to handle critical instructions.
• Noncritical instructions execute directly on hardware for efficiency, while critical
instructions are trapped and emulated by the Virtual Machine Monitor (VMM).
• VMware and similar solutions place the VMM at Ring 0 and the guest OS at Ring 1,
ensuring complete decoupling from hardware.
• Although binary translation enables compatibility, it introduces performance
overhead, particularly for I/O-intensive applications. Performance on x86 systems
typically reaches 80% to 97% of native execution speed.

Host-Based Virtualization

• In this approach, a host OS manages hardware, with a virtualization layer placed on

top to run guest OSes.
• Unlike full virtualization, the host OS remains in control, providing drivers and low-
level services, simplifying deployment.
• However, performance suffers due to multiple layers of mapping—every hardware
request must pass through four layers, leading to significant slowdowns.
• If the guest OS and hardware have different ISAs, binary translation is required,
further reducing performance.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 19

BCS601 Cloud Computing Module-3

While host-based virtualization offers flexibility, it is generally less efficient than full
virtualization with a VMM.

3.2.3 Para-Virtualization with Compiler Support

Para-virtualization reduces virtualization overhead by modifying the guest OS to replace

privileged instructions with hypercalls that communicate directly with the hypervisor. Unlike
full virtualization, which relies on binary translation, para-virtualization improves
performance but introduces compatibility and maintenance challenges.

Challenges of Para-Virtualization

1. Compatibility & Portability Issues: Since para-virtualization modifies the guest OS,
supporting unmodified OSes becomes difficult.
2. High Maintenance Costs: OS kernel modifications require ongoing updates and
support.
3. Variable Performance Gains: The performance improvement depends on the
workload and system architecture.

Para-Virtualization Architecture

Dr. Sampada K S, Associate professor CSE RNSIT pg. 20

BCS601 Cloud Computing Module-3

Para-Virtualization with Compiler Support

• Guest OS Modification: The OS kernel is modified, but user applications may also
need changes.
• Hypercalls: Privileged instructions that would normally run at Ring 0 are replaced
with hypercalls to the hypervisor.
• Intelligent Compiler: A specialized compiler assists in identifying and replacing
nonvirtualizable instructions with hypercalls, optimizing performance.
• Improved Efficiency: Compared to full virtualization, para-virtualization
significantly reduces overhead, making VM execution closer to native
performance.
• Limitation: Since the guest OS is modified, it cannot run directly on physical
hardware without a hypervisor.

Para-Virtualization on x86 Architecture

• Traditional x86 processors have four privilege levels (Rings 0–3):

o Ring 0: Kernel (full control of hardware)
o Ring 3: User applications (restricted access)
• In para-virtualization, the guest OS cannot execute at Ring 0 directly. Instead, it
operates at a lower ring, using hypercalls to interact with the hypervisor at Ring 0.

Due to the inefficiency of binary translation, many virtualization solutions, including Xen,
KVM, and VMware ESX, use para-virtualization.

KVM (Kernel-Based Virtual Machine)

• A Linux-based para-virtualization system, integrated into the Linux 2.6.20 kernel.

• Leverages Linux's existing memory management and scheduling, making it more
lightweight than a full hypervisor.
• Supports unmodified guest OSes, including Windows, Linux, Solaris, and UNIX
variants, with hardware-assisted virtualization.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 21

BCS601 Cloud Computing Module-3

Example 3.3 VMware ESX Server for Para-Virtualization

VMware pioneered the virtualization market, providing solutions for desktops, servers, and
data centers. VMware ESX is a bare-metal hypervisor designed for x86 symmetric
multiprocessing (SMP) servers, enabling efficient virtualization by directly managing
hardware resources.

Key Components of VMware ESX Server

1. Virtualization Layer (VMM Layer)

o The Virtual Machine Monitor (VMM) virtualizes CPU, memory, network, disk
controllers, and human interface devices.
o Each VM has its own set of virtual hardware resources, isolated from others.
2. Resource Manager
o Allocates CPU, memory, storage, and network bandwidth to VMs.
o Maps physical resources to virtual hardware resources for efficient distribution.
3. Hardware Interface Components
o Includes device drivers that facilitate direct access to I/O devices.
o Uses a VMkernel-based para-virtualization architecture for improved
performance.
4. Service Console (Legacy in ESX, removed in ESXi)
o Handles system booting, initialization of VMM and Resource Manager, and
administrative functions.

Para-Virtualization in ESX

• The VMkernel interacts directly with the hardware, bypassing the need for a host OS.
• Para-virtualized drivers (e.g., VMXNET for networking, PVSCSI for disk I/O) improve
performance.
• Provides better efficiency than full virtualization while supporting unmodified guest
OSes via hardware-assisted virtualization (Intel VT, AMD-V).

By leveraging para-virtualization, VMware ESX optimizes performance, enhances resource

management, and ensures scalability, making it a preferred choice for enterprise data center
virtualization.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 22

BCS601 Cloud Computing Module-3

3.3 Virtualization of CPU, Memory, And I/O Devices

3.3.1 Hardware Support for Virtualization

Modern processors support multiple processes running simultaneously, but they require
protection mechanisms to prevent system crashes. This is achieved by dividing execution into
user mode and supervisor mode:

• User Mode: Runs unprivileged instructions, restricting direct hardware access.

• Supervisor Mode: Runs privileged instructions, allowing direct control over system
hardware.

In virtualized environments, ensuring OS and application correctness is more complex due to

additional abstraction layers.

Hardware Virtualization Products

Several solutions leverage hardware support for virtualization, including:

1. VMware Workstation
o A host-based virtualization software suite for x86 and x86-64 systems.
o Runs multiple VMs simultaneously on a host OS.
2. Xen Hypervisor
o Works on IA-32, x86-64, Itanium, and PowerPC 970 architectures.
o Modifies Linux to function as a hypervisor, controlling guest OSes.
3. KVM (Kernel-Based Virtual Machine)
o Integrated into the Linux kernel as a virtualization infrastructure.
o Supports hardware-assisted virtualization (Intel VT-x, AMD V) and
paravirtualization via the VirtIO framework.
o VirtIO components include:
▪ Paravirtual Ethernet Card (for networking).
▪ Disk I/O Controller (optimized storage access).
▪ Balloon Device (dynamically adjusts VM memory allocation).

Dr. Sampada K S, Associate professor CSE RNSIT pg. 23

BCS601 Cloud Computing Module-3

▪ VGA Graphics Interface (enhanced graphics performance using

VMware drivers).

By leveraging hardware-assisted virtualization, modern hypervisors like VMware, Xen, and

KVM achieve higher efficiency, reduced overhead, and improved performance for virtual
machines.

Example 3.4: Hardware Support for Virtualization in the Intel x86 Processor

Software-based virtualization methods are complex and introduce performance overhead. To

address this, Intel provides hardware-assisted virtualization techniques that simplify
implementation and improve efficiency.

Figure 3.10 provides an overview of Intel’s full virtualization techniques. For processor
virtualization, Intel offers the VT-x or VT-i technique. VT-x adds a privileged mode (VMX
Root Mode) and some instructions to processors. This enhancement traps all sensitive
instructions in the VMM automatically. For memory virtualization, Intel offers the EPT, which
translates the virtual address to the machine’s physical addresses to improve performance. For
I/O virtualization, Intel implements VT-d and VT-c to support this.

3.3.2 CPU Virtualization

A Virtual Machine (VM) replicates a real computer system, executing most instructions on the
host processor in native mode for efficiency. However, critical instructions must be carefully
managed to ensure stability and correctness.

Types of Critical Instructions

1. Privileged Instructions
o Execute only in privileged mode.
o If executed in user mode, they trigger a trap.
2. Control-Sensitive Instructions
o Modify system resources (e.g., changing memory configuration).
3. Behavior-Sensitive Instructions

Dr. Sampada K S, Associate professor CSE RNSIT pg. 24

BCS601 Cloud Computing Module-3

o Change behavior based on system configuration (e.g., memory load/store

operations).

Virtualizability of CPU Architectures

• RISC CPUs (e.g., PowerPC, SPARC) are naturally virtualizable since all sensitive
instructions are privileged.
• x86 CPUs were not originally designed for virtualization, as some sensitive instructions
(e.g., SGDT, SMSW) are not privileged.
o These instructions bypass the VMM, making virtualization difficult without
software-based techniques like binary translation.

Paravirtualization in CPU Virtualization

• In a UNIX-like system, a system call triggers an 80h interrupt, transferring control to

the OS kernel.
• In a paravirtualized system (e.g., Xen), the system call:
1. Triggers the 80h interrupt in the guest OS.
2. Also triggers the 82h interrupt in the hypervisor.
3. The hypervisor processes the call before returning control to the guest OS
kernel.
• Benefit: Runs unmodified applications in a VM.
• Drawback: Small performance overhead due to additional hypervisor intervention.

Hardware-Assisted CPU Virtualization

To simplify virtualization, Intel and AMD introduced hardware extensions:

• Intel VT-x / AMD-V: Adds a new privilege mode (Ring -1).

o Hypervisor runs at Ring -1.
o Guest OS runs at Ring 0 (normal OS mode).
o All privileged instructions are automatically trapped into the hypervisor.

Advantages of Hardware-Assisted Virtualization:

i. Removes the need for binary translation (improving performance).
ii. Allows unmodified operating systems to run in virtual machines.
iii. Simplifies virtualization implementation.

Example 3.5: Intel Hardware-Assisted CPU Virtualization

Virtualizing x86 Processors

• Unlike RISC processors, x86 processors were not originally virtualizable.

• Since x86-based legacy systems are widely used, virtualization is necessary.

Intel VT-x Technology

• Introduces VMX Root Mode, an additional privilege level.

• Adds special instructions to:
o Start/stop VMs.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 25

BCS601 Cloud Computing Module-3

o Allocate memory pages to store CPU states.

• Used in hypervisors like Xen, VMware, and Microsoft Virtual PC.

Performance Considerations

• High efficiency expected, but switching between hypervisor and guest OS causes
overhead.
• Hybrid Approach (used by VMware):
o Offloads some tasks to hardware while keeping others in software.
• Combining Para-Virtualization with Hardware-Assisted Virtualization further boosts
performance.

3.3.3 Memory Virtualization

1. Virtual Memory Virtualization Basics

• Similar to traditional virtual memory in modern OSes.

• Standard OSes use page tables for one-stage mapping (Virtual Memory → Physical
Memory).
• Virtualized environments require a two-stage mapping:
o Guest OS: Virtual Memory → Physical Memory (Guest View).
o VMM: Physical Memory → Machine Memory (Actual Hardware).

2. Memory Management in Virtualized Systems

• Memory Management Unit (MMU) and Translation Lookaside Buffer (TLB) help
optimize performance.
• Guest OS controls virtual-to-physical mapping but cannot directly access machine
memory.
• VMM (Hypervisor) handles actual memory allocation to prevent conflicts.

3. Shadow Page Tables & Nested Paging

Dr. Sampada K S, Associate professor CSE RNSIT pg. 26

BCS601 Cloud Computing Module-3

• Shadow Page Table (SPT):

o Maintained by the VMM to map guest physical memory → machine memory.
o Reduces guest OS overhead but increases memory usage and performance
costs.
• Nested Page Tables (NPT):
o Adds another layer of address translation.
o First introduced in AMD Barcelona (2007) as hardware-assisted memory
virtualization.
o Reduces VMM overhead by offloading translation to hardware.

4. VMware's Approach

• Uses shadow page tables to map virtual memory to machine memory.

• TLB caching avoids repeated translations, improving performance.
• AMD and Intel now support hardware-assisted memory virtualization to reduce
overhead.

Example 3.6 Extended Page Table by Intel for Memory Virtualization

1.The Problem with Shadow Page Tables

• Software-based shadow page tables were inefficient and caused high performance
overhead.
• Frequent memory lookups and context switches slowed down virtualized environments.

2. Intel’s EPT Solution

• Hardware-assisted memory virtualization that eliminates the need for shadow page
tables.
• Works with Virtual Processor ID (VPID) to optimize Translation Lookaside Buffer
(TLB) usage.
• Reduces memory lookup time and improves performance significantly.

3. How EPT Works

• Uses a four-level page table hierarchy (same as guest OS page tables).

Dr. Sampada K S, Associate professor CSE RNSIT pg. 27

BCS601 Cloud Computing Module-3

• Translation Process:
1. Guest OS uses Guest CR3 (Control Register 3) to point to L4 page table.
2. CPU must translate Guest Physical Address (GPA) to Host Physical Address
(HPA) using EPT.
3. The CPU first checks the EPT TLB for an existing translation.
4. If not found, it searches the EPT page tables (up to 5 times in the worst case).
5. If still not found, an EPT violation exception is triggered.
6. The CPU will access memory multiple times to resolve the mapping (up to 20
memory accesses).

4. Optimization: EPT TLB Expansion

• Intel increased the size of EPT TLB to store more translations and reduce memory
accesses.
• This dramatically improves memory access speed and virtualization efficiency.

5. Impact of EPT on Virtualization

• Removes the need for shadow page tables, reducing overhead.

• Works seamlessly with Intel VT-x for faster CPU virtualization.
• Improves performance for workloads running in virtualized environments like
VMware, Xen, and KVM.

3.3.4 I/O Virtualization

Dr. Sampada K S, Associate professor CSE RNSIT pg. 28

BCS601 Cloud Computing Module-3

I/O virtualization allows virtual machines (VMs) to share and access physical I/O devices
efficiently. There are three main methods to implement I/O virtualization:

1. Full Device Emulation

• The VMM (Virtual Machine Monitor) emulates well-known hardware devices.

• Guest OS interacts with a virtual device, which is mapped to a real physical device by
the virtualization layer.
• Functions performed by the virtualization layer:
o Emulates device features like interrupts, DMA, and device enumeration.
o Remaps guest OS I/O addresses to real device addresses.
o Multiplexes and routes I/O requests from multiple VMs.
o Supports advanced features like Copy-on-Write (COW) disks.
• Advantages:
o Allows compatibility with unmodified guest OSes.
• Disadvantages:
o High overhead and slow performance due to software emulation.

2. Para-Virtualization (Split Driver Model)

• Used in Xen and other virtualization platforms.

• Uses a frontend driver (inside the guest OS) and a backend driver (inside the hypervisor
or privileged VM).
• How it works:
o Frontend driver (Domain U): Handles I/O requests from guest OS.
o Backend driver (Domain 0): Manages real I/O devices and routes data between
VMs.
o Communication occurs via a shared memory block.
• Advantages:
o Better performance than full device emulation.
• Disadvantages:
o Higher CPU overhead due to the need for additional processing.

3. Direct I/O Virtualization

• Allows direct access between the VM and the physical device.

• Benefits:
o Near-native performance because I/O operations bypass software emulation.
o Lower CPU overhead compared to para-virtualization.
• Challenges:
o Current implementations focus mainly on networking for mainframes.
o Difficulties in reclaiming and reassigning physical devices after VM migration.
o Risk of arbitrary device states (e.g., DMA errors leading to system crashes).

4. Hardware-Assisted I/O Virtualization

• Intel VT-d technology supports I/O DMA remapping and device interrupt remapping.
• Helps unmodified, specialized, or virtualization-aware guest OSes run efficiently.

5. Self-Virtualized I/O (SV-IO)

Dr. Sampada K S, Associate professor CSE RNSIT pg. 29

BCS601 Cloud Computing Module-3

• Uses multicore processors to manage I/O virtualization tasks.

• Virtual Interface (VIF) Model:
o Provides virtual devices with an access API for VMs.
o Provides a management API for the VMM.
o Each virtual device (VIF) has a unique ID and uses message queues for
communication.
• Applications:
o Virtual network interfaces
o Virtual block devices (disks)
o Virtual camera devices

Summary

Method Performance CPU Overhead Hardware Dependency

Full Device Emulation Low High No
Para-Virtualization Moderate Moderate Yes (Modified OS)
Direct I/O High Low Yes (Special Hardware)
Hardware-Assisted I/O (VT-d) High Low Yes
Self-Virtualized I/O (SV-IO) High Low-Moderate Yes

I/O virtualization continues to evolve, with hardware-assisted methods like VT-d and SV-
IO improving efficiency and reducing overhead.

Example 3.7 VMware Workstation for I/O Virtualization

VMware Workstation is a hosted hypervisor that runs as an application on a host operating

system (OS). It relies on guest OS, host OS, and VMM (Virtual Machine Monitor) to
implement I/O virtualization efficiently.

How VMware Workstation Implements I/O Virtualization

1. VMApp (Application Layer):

o Runs as an application inside the host OS.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 30

BCS601 Cloud Computing Module-3

o Manages VM operations and user interactions.

2. VMDriver (Host OS Driver):
o Loaded into the host OS kernel.
o Acts as a bridge between VMApp and VMM.
o Facilitates control transfers between the host OS and the VMM world.
3. Virtual Machine Monitor (VMM):
o Runs at a privileged level directly on the hardware.
o Manages the execution of guest OSes and handles virtualization tasks.
4. Processor Execution:
o A physical processor operates in two distinct states:
▪ Host World: Executes tasks for the host OS.
▪ VMM World: Runs the guest OS and virtualized workloads.
o VMDriver facilitates transitions between these two execution states.

Advantages of VMware Workstation’s I/O Virtualization

1.Broad hardware compatibility (supports various guest OSes without modification).

2. Runs as a regular application on host OS, making installation and management simple.
3. Leverages host OS drivers for better I/O support.

Disadvantages

1. Performance overhead due to full device emulation.

2. Slower than direct I/O virtualization approaches.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 31

BCS601 Cloud Computing Module-3

3.3.5 Virtualization in multicore Processors

As multi-core processors become more prevalent, virtualizing them presents unique challenges
compared to single-core processors. While multi-core CPUs offer higher performance by
integrating multiple cores on a single chip, virtualization introduces complexities in task
scheduling, parallel execution, and resource management.

Challenges in Multi-Core Virtualization

1. Parallelism & Programming Models

• Applications must be parallelized to fully utilize multiple cores.

• New programming models, languages, and libraries are required to simplify parallel
programming.

2. Task Scheduling & Resource Management

• Efficient scheduling algorithms are needed to distribute tasks across cores.

• Resource allocation policies must balance performance, complexity, and power
efficiency.

3. Dynamic Heterogeneity

• The integration of different types of cores (fat CPU cores & thin GPU cores) on the
same chip makes resource management more complex.
• As transistor reliability decreases and complexity increases, system designers must
adapt scheduling techniques dynamically.

Multi-core virtualization presents challenges in parallel execution, task scheduling, and

resource management. Applications need to be parallelized, and efficient scheduling is crucial
for performance. Dynamic heterogeneity, where CPU and GPU cores coexist, further
complicates resource allocation.

To address these challenges:

• Virtual Processor Cores (VCPUs) abstract low-level hardware management, allowing

dynamic core migration and suspension.
• Virtual Hierarchies adapt cache and coherence structures to optimize workload
distribution, reducing cache misses and improving isolation between VMs.

These advancements enhance performance, flexibility, and efficiency, making multi-core

virtualization ideal for server consolidation and cloud computing.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 32

BCS601 Cloud Computing Module-3

3.4 VIRTUAL CLUSTERS AND RESOURCE MANAGEMENT

What are Virtual Clusters?

A physical cluster is a group of physical servers connected through a network. In contrast, a

virtual cluster consists of virtual machines (VMs) spread across multiple physical machines.
These VMs are connected via a virtual network and can be managed flexibly.

Key Characteristics of Virtual Clusters:

Dr. Sampada K S, Associate professor CSE RNSIT pg. 33

BCS601 Cloud Computing Module-3

• VMs can run on physical machines and be assigned dynamically.

• Each VM runs its own operating system, which can differ from the host system.
• VMs improve server utilization by running multiple applications on the same physical
hardware.
• VMs can be replicated across multiple servers to enhance performance, fault tolerance,
and disaster recovery.
• Virtual clusters can expand or shrink dynamically, depending on demand.
• If a physical machine fails, only the VMs on that machine are affected, while other
VMs and the host system continue running.

Challenges in Virtual Cluster Management:

1. Fast Deployment & Scheduling

o Virtual clusters must be quickly set up, shut down, and switched to optimize
resource use.
o Green computing techniques aim to save energy, but live VM migration can
cause performance overhead.
o Load balancing strategies help distribute workloads efficiently across VMs.
2. High-Performance Virtual Storage
o VM images (software templates) must be stored efficiently to reduce space and
improve deployment speed.
o Copy-on-Write (COW) technology allows creating new VMs quickly by
modifying existing templates instead of duplicating entire disk images.
o Automating VM configurations can save time when managing large numbers of
VMs.

Conclusion:

Virtual clusters provide flexibility, efficient resource usage, and better fault tolerance.
However, they require careful management for fast deployment, effective load balancing, and
optimized storage. Strategies like automated configuration and optimized migration help
improve performance while reducing overhead.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 34

BCS601 Cloud Computing Module-3

3.4.2 Live VM Migration Steps and Performance Effects

In a mixed-node cluster, virtual machines (VMs) typically run on physical hosts, but if a host
fails, its VM role can be taken over by another VM on a different host. This enables flexible
failover compared to traditional physical-to-physical failover. However, if a host fails, its VMs
also fail, which can be mitigated through live VM migration.

Virtual Cluster Management Approaches:

1. Guest-Based Manager: The cluster manager runs inside the guest OS (e.g., OpenMosix,
Sun’s Oasis).
2. Host-Based Manager: The cluster manager runs on the host OS, supervising VMs (e.g.,
VMware HA).
3. Independent Manager: Both guest and host have separate cluster managers, increasing
complexity.
4. Integrated Cluster Management: A unified manager controls both virtual and physical
resources.

Live VM Migration Process (6 Steps):

1. Start Migration: Identify the VM and destination host, often triggered by load balancing
or server consolidation strategies.
2. Memory Transfer: The VM’s memory is copied to the destination host in multiple
rounds, ensuring minimal disruption.
3. Suspend and Final Copy: The VM pauses briefly to transfer the last memory portion,
CPU, and network states.
4. Commit and Activate: The destination host loads the VM state and resumes execution.
5. Redirect Network & Cleanup: The network redirects to the new VM, and the old VM
is removed.

Performance Effects:

Dr. Sampada K S, Associate professor CSE RNSIT pg. 35

BCS601 Cloud Computing Module-3

• The first memory copy takes 63 seconds, reducing network speed from 870 MB/s to
765 MB/s.
• Additional memory copy rounds further reduce speed to 694 MB/s in 9.8 seconds.
• The total downtime is only 165 milliseconds, ensuring minimal service disruption.

Key Benefits of Live Migration:

• Ensures continuous service availability in cloud computing, HPC, and computational

grids.
• Enables dynamic resource allocation on demand.
• Supports disaster recovery with minimal overhead.
• Prevents resource contention with careful network and CPU usage planning.

Live VM migration enhances cloud computing by enabling seamless workload balancing and
minimizing downtime during host failures. Platforms like VMware and Xen support these
migrations, allowing multiple VMs to run efficiently on a shared physical infrastructure.

VM running normally on Host A Stage 0: Pre-Migration

Active VM on Host A

Alternate physical host may be preselected for migration Block devices

mirrored and free resources maintained

Stage 1: Reservation

Initialize a container on the target host

Overhead due to copying Stage 2: Iterative pre-copy

Enable shadow paging

Copy dirty pages in successive rounds.

Downtime Stage 3: Stop and copy

(VM out of service) Suspend VM on host A

Generate ARP to redirect traffic to Host B Synchronize all

remaining VM state to Host B

Stage 4: Commitment

VM state on Host A is released

VM running normally on Host B Stage 5: Activation

VM starts on Host B Connects to local devices

Resumes normal operation

Figure. 3.20 Live migration process of a VM from one host to another.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 36

BCS601 Cloud Computing Module-3

3.4.3 Migration of Memory, Files, and Network Resources

Shared clusters reduce costs and improve resource utilization. When migrating a system to a
new physical node, key factors include memory migration, file system migration, and
network migration.

1. Memory Migration

• VM memory migration involves moving memory states efficiently.

• The Internet Suspend-Resume (ISR) technique uses temporal locality to transfer only
changed data.
• A tree-based file structure minimizes data transfer by copying only modified parts.
• ISR results in high downtime, making it unsuitable for live migrations.

2. File System Migration

• VM migration requires a consistent, location-independent file system accessible on all

hosts.
• Two approaches:
1. Virtual Disk Transport (copies entire disk but is slow for large data).
2. Global Distributed File System (avoids copying, supports direct network
access).
• Smart Copying uses spatial locality, transferring only differences between the old and
new locations.
• Proactive State Transfer predicts the destination and pre-transfers critical data.

3. Network Migration

• Migrating VMs must maintain network connections without disruption.

• Each VM has a virtual IP and MAC address, separate from the host machine.
• Unsolicited ARP replies notify network peers of the VM’s new location.
• Switched networks detect new VM locations and reroute traffic automatically.

4. Live VM Migration Techniques

• Precopy Approach:
o Transfers all memory pages first, then iteratively copies only modified pages.
o Reduces downtime but increases total migration time.
• Checkpoint/Recovery & Trace/Replay (CR/TR-Motion):
o Transfers execution logs instead of dirty pages, minimizing migration time.
o Limited by differences in source and target system performance.
• Postcopy Approach:
o Transfers memory pages once but has higher downtime due to fetch delays.
• Memory Compression:
o Uses spare CPU resources to compress memory pages before transfer, reducing
data size.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 37

BCS601 Cloud Computing Module-3

5. Live Migration Using Xen

• Xen Hypervisor allows multiple OSes to share hardware.

• Domain 0 (Dom0) manages VM creation, termination, and migration.
• Xen uses a send/receive model to transfer VM states efficiently.

Key Takeaways

• Live migration ensures minimal downtime while keeping services running.

• Optimized memory, file, and network transfer techniques improve efficiency.
• Different migration strategies (Precopy, Postcopy, CR/TR-Motion, and Compression)
balance speed and downtime based on system needs.
• Xen provides a reliable VM migration framework for enterprise environments.

Example 3.8 Live Migration of VMs between Two Xen-Enabled Hosts

What is Live Migration?
Live migration is the process of moving a running Virtual Machine (VM) from one physical
machine to another without stopping its operations. This means users experience little to no
downtime while the VM is being transferred.

How Does Xen Handle Live Migration?

Xen uses a method called Remote Direct Memory Access (RDMA) to speed up migration by
bypassing traditional network communication protocols like TCP/IP. This avoids unnecessary
processing by the CPU, making the transfer faster and more efficient.

Key Steps in Xen’s Live Migration Process:

1. Migration Daemon (a background process) manages the migration.
2. Shadow Page Tables track memory changes during the migration process.
3. Dirty Bitmaps record modified memory pages. These pages are updated and sent to the
new machine in precopy rounds.
4. Memory pages are compressed before transfer to reduce data size and improve speed.
5. The new host decompresses the memory pages and resumes the VM.

Trade-offs in Migration
• The compression algorithm must be fast and effective for different types of memory
data.
• Using a single compression method for all memory pages is not efficient because
different memory types require different strategies.

Conclusion
Live migration in Xen, enhanced by RDMA, allows seamless VM transfer with minimal impact
on performance. Techniques like precopying, dirty bitmaps, and compression improve
efficiency while ensuring smooth operation.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 38

BCS601 Cloud Computing Module-3

3.4.4 Dynamic Deployment of Virtual Clusters

What is Dynamic Deployment of Virtual Clusters?
Dynamic deployment allows virtual clusters (vClusters) to change in size, move, or adapt to
resource demands. This helps in efficient resource management, improved performance, and
cost savings in cloud computing.

Research Projects on Virtual Clusters:

1. Cluster-on-Demand (COD) – Duke University
o Goal: Dynamically allocate servers to multiple vClusters.
o Results: Efficient sharing of VMs using Sun GridEngine.
2. Cellular Disco – Stanford University
o Goal: Deploy a virtual cluster on a shared-memory multiprocessor system.
o Results: Multiple VMs successfully managed under Cellular Disco Virtual
Machine Monitor (VMM).
3. VIOLIN – Purdue University
o Goal: Improve performance through dynamic adaptation of VMs.
o Results: Reduced execution time for parallel applications by adapting to
changing workloads.
4. GRAAL – INRIA, France
o Goal: Evaluate parallel algorithms in Xen-enabled virtual clusters.
o Results: Achieved 75% of maximum performance while using only 30% of the
total resources.

Example 3.9: COD (Cluster-on-Demand) – Duke University

COD is a virtual cluster management system that allows automatic resizing and reallocation of
clusters.

• Uses Sun GridEngine for scheduling workloads.

• Users can request resources dynamically via a web interface.
• Figure 3.24 shows how the number of servers in different vClusters changed over eight
days based on demand.
• Helps with load balancing, resource reservation, and automatic provisioning.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 39

BCS601 Cloud Computing Module-3

Example 3.10: VIOLIN – Purdue University

VIOLIN focuses on live VM migration to dynamically adjust virtual clusters for better resource
use.

• Supports multiple virtual environments (VIOLIN 1–5) running on shared physical

clusters.
• Can move VMs between clusters as needed without disrupting applications.
• Key result: Increased resource utilization with only a 1% increase in execution time.

Note:
• Virtual clusters help efficiently manage and allocate resources.
• COD and VIOLIN show that dynamic adaptation can significantly improve resource
utilization.
• Live migration allows VMs to be moved with minimal downtime.
• These techniques enable scalable, flexible, and cost-effective cloud computing
solutions.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 40

BCS601 Cloud Computing Module-3

3.5 VIRTUALIZATION FOR DATA-CENTER AUTOMATION

Data centers have expanded rapidly, with major IT companies like Google, Amazon, and
Microsoft investing heavily in automation. This automation dynamically allocates hardware,
software, and database resources to millions of users while ensuring cost-effectiveness and
Quality of Service (QoS). The rise of virtualization and cloud computing has driven this
transformation, with market growth from $1.04 billion in 2006 to a projected $3.2 billion by
2011.

Server Consolidation in Data Centers

Data centers handle heterogeneous workloads, categorized as:

• Chatty workloads (e.g., web video services) that have fluctuating demand.
• Noninteractive workloads (e.g., high-performance computing) that require consistent
resource allocation.

To meet peak demand, resources are often statically allocated, leading to underutilized servers
and wasted costs in hardware, space, and power. Server consolidation—particularly
virtualization-based consolidation—optimizes resource management by reducing physical
servers and improving hardware utilization.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 41

BCS601 Cloud Computing Module-3

Benefits of Server Virtualization

• Enhances hardware utilization by consolidating underutilized servers.

• Improves resource provisioning through agile VM deployment.
• Reduces costs, including server purchases, maintenance, power, cooling, and cabling.
• Enhances availability and business continuity, as VM crashes do not impact other VMs
or the host system.

Challenges and Optimization Strategies

• Resource Scheduling: Efficient, multi-level schedulers improve utilization and QoS.

• Dynamic CPU Allocation: Adjusts resources based on VM utilization and workload
demand.
• Two-Level Resource Management: Uses local VM controllers and global server
controllers for optimized allocation.
• Power Management: A VM-aware power budgeting scheme balances power savings
with performance while addressing hardware heterogeneity.

By leveraging virtualization and multicore processing (CMP), data centers can enhance
efficiency, but optimization in memory access, VM reassignment, and power management
remains a challenge.

Virtual Storage Management

• Storage virtualization in system environments differs from traditional storage

aggregation, focusing on VM image management and application data handling.
• A key challenge is storage bottlenecks, as VMs compete for disk resources.
• Parallax, a distributed storage system, addresses these challenges by offering scalable
virtual disks across a common physical storage device. It enhances storage flexibility,
reduces storage footprint, and supports advanced features like snapshots.

Overall, virtualization significantly improves data-center efficiency by optimizing server and

storage management, reducing costs, and enabling scalable, automated resource allocation.

Example 3.11 – Parallax Virtual Storage System

Parallax is a scalable virtual storage system designed for cluster-based environments. It enables
efficient storage management by using a set of per-host storage appliances that share access to
a common block device.

Key Features of Parallax:

• Cluster-Wide Storage Management: A centralized administrative domain manages all

storage appliance VMs.
• Pushes Storage Functionality to Hosts: Features like snapshots are implemented in
software instead of dedicated hardware.
• Virtual Disk Images (VDIs): Provides single-writer virtual disks, accessible from any
physical host in the cluster.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 42

BCS601 Cloud Computing Module-3

• Efficient Block Virtualization: Uses Xen’s block tap driver and tapdisk library for
handling block storage requests across VMs.
• Storage Appliance VM: Acts as an intermediary between client VMs and physical
hardware, facilitating live upgrades of block device drivers.

Parallax enhances flexibility, scalability, and ease of storage management in virtualized data
centers by integrating advanced block storage virtualization techniques.

3.5.3 Cloud OS for Virtualized Data Centers

To function as cloud providers, data centers must be virtualized using Virtual Infrastructure
(VI) managers and Cloud OSes. Table 3.6 outlines four such platforms:

1. Nimbus (Open-source)
2. Eucalyptus (Open-source)
3. OpenNebula (Open-source)
4. vSphere 4 (Proprietary, VMware)

Key Features of VI Managers & Cloud OSes:

• VM Creation & Management: All platforms support virtual machines and virtual
clusters for elastic cloud resources.
• Virtual Networking: Nimbus, Eucalyptus, and OpenNebula offer virtual network
support, enabling flexible communication between VMs.
• Dynamic Resource Provisioning: OpenNebula stands out by allowing advance
reservations of cloud resources.
• Hypervisor Support:
o Nimbus, Eucalyptus, and OpenNebula use Xen & KVM for virtualization.
o vSphere 4 utilizes VMware ESX & ESXi hypervisors.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 43

BCS601 Cloud Computing Module-3

• Virtual Storage & Data Protection: Only vSphere 4 supports virtual storage along with
networking and data protection.

Example 3.12 Eucalyptus for Virtual Networking of Private Cloud

Eucalyptus is an open-source software system designed for private cloud infrastructure and
Infrastructure as a Service (IaaS). It enables virtual networking and VM management, but does
not support virtual storage.

Key Features of Eucalyptus:

• Private Cloud Deployment:

o Supports Ethernet and Internet-based networking to connect VMs.
o Can interact with public and private clouds.
• Component-based Web Services Architecture:
o Uses WS-Security policies for secure communication.
o Web services expose language-agnostic APIs via WSDL documents.
• Resource Managers in Eucalyptus:
o Instance Manager: Manages VM execution, inspection, and termination.
o Group Manager: Handles scheduling and virtual network management.
o Cloud Manager: Central user entry-point, queries nodes, makes scheduling
decisions.
• AWS Compatibility:
o Works like Amazon EC2 APIs and supports S3 storage emulation.
o Compatible with SOAP, REST, and CLI-based management.
• Platform Support:
o Installed on Linux-based platforms.

Eucalyptus provides a flexible and scalable solution for private cloud networking but lacks
some security and general-purpose cloud features.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 44

BCS601 Cloud Computing Module-3

Example 3.13 VMware vSphere 4 as a Commercial Cloud OS

vSphere 4, released by VMware in April 2009, is a virtualization platform designed for private
cloud management. It extends earlier VMware products like Workstation, ESX, and Virtual
Infrastructure. The system interacts with applications through vCenter and provides
infrastructure and application services.

The infrastructure services include three components:

• vCompute (ESX, ESXi, DRS)

• vStorage (VMS, thin provisioning)
• vNetwork (distributed switching and networking)

The application services focus on:

• Availability (VMotion, Storage VMotion, HA, Fault Tolerance, Data Recovery)

• Security (vShield Zones, VMsafe)
• Scalability (DRS, Hot Add)

Users must understand vCenter interfaces to manage applications effectively. More details are
available on the vSphere 4 website.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 45

BCS601 Cloud Computing Module-3

3.5.4 Trust Management in Virtualized Data Centers

A Virtual Machine Monitor (VMM) creates and manages Virtual Machines (VMs) by acting
as a software layer between the operating system and hardware. It provides secure isolation
and manages access to hardware resources, making it the foundation of security in virtualized
environments. However, if a hacker compromises the VMM or management VM, the entire
system is at risk. Security issues also arise from random number reuse, which can lead to
encryption vulnerabilities and TCP hijacking attacks.

VM-Based Intrusion Detection

Intrusion Detection Systems (IDS) help identify unauthorized access. IDS can be:

• Host-based IDS (HIDS) – Runs on a monitored system but is vulnerable to attacks.

• Network-based IDS (NIDS) – Monitors network traffic but struggles to detect fake
actions.

A VM-based IDS leverages virtualization to isolate VMs, preventing compromised VMs from
affecting others. The Virtual Machine Monitor (VMM) can audit access requests, combining
the strengths of HIDS and NIDS. There are two methods for implementation:

1. IDS as an independent process inside a high-privileged VM on the VMM.

2. IDS integrated into the VMM with direct hardware access.

Garfinkel and Rosenblum proposed a VMM-based IDS that monitors guest VMs using a policy
framework and trace-based security enforcement. However, logs used for analysis can be
compromised if the operating system is attacked.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 46

BCS601 Cloud Computing Module-3

Honeypots and Honeynets

Besides IDS, honeypots and honeynets are used to detect attacks by tricking attackers into
interacting with fake systems. Honeypots can be physical or virtual, and in virtual honeypots,
the host OS and VMM must be protected to prevent attacks from guest VMs.

Example 3.14 EMC Establishment of Trusted Zones for Protection of Virtual

Clusters Provided to Multiple Tenants

EMC and VMware collaborated to develop security middleware for trust management in
distributed systems and private clouds. The concept of trusted zones was introduced to enhance
security in virtual clusters, where multiple applications and OS instances for different tenants
operate in separate virtual environments.

Trusted Zones Architecture

• Physical Infrastructure (Cloud Provider) – Forms the foundation at the bottom.

• Virtual Clusters (Tenants) – Separate virtual environments for different tenants.
• Public Cloud (Global Users) – Represents broader user communities.
• Security Measures – Includes anti-virus, worm containment, intrusion detection, and
encryption.

The trusted zones ensure secure isolation of VMs while allowing controlled interactions among
tenants, providers, and global communities. This approach strengthens security in private cloud
environments.

Dr. Sampada K S, Associate professor CSE RNSIT pg. 47

BCS601 Cloud Computing Module-3

Dr. Sampada K S, Associate professor CSE RNSIT pg. 48