Unit 1: Virtualization and Distributed Computing - Subjective Questions

Virtualization is the process of creating a logical or virtual version of a physical computing resource, such as a server, operating system, storage device, or network.

Its fundamental concepts include:

• Abstraction: Physical hardware details are hidden from users and applications.
• Resource pooling: CPU, memory, storage, and network resources are combined and allocated as required.
• Isolation: Each virtual machine operates independently, so a failure in one virtual machine generally does not affect others.
• Partitioning: A physical system can be divided into multiple virtual environments.
• Encapsulation: A virtual machine and its state can be represented as files, making it easier to copy, back up, or migrate.
• Hardware independence: Virtual machines can run on different compatible physical systems without major operating-system changes.

Virtualization improves hardware utilization, flexibility, scalability, disaster recovery, and centralized resource management.

A hypervisor, also called a virtual machine monitor, is software or firmware that creates, runs, and manages virtual machines. It allocates physical CPU, memory, storage, and network resources while maintaining isolation between virtual machines.

Type 1 hypervisor:

• Runs directly on the physical hardware.
• Does not require a host operating system.
• Usually provides better performance, security, and scalability.
• Commonly used in data centers and cloud platforms.
• Examples include VMware ESXi, Microsoft Hyper-V, and Xen.

Type 2 hypervisor:

• Runs as an application on a host operating system.
• Uses the host operating system's drivers and resource-management services.
• Is easier to install but introduces additional overhead.
• Commonly used for development, testing, and desktop virtualization.
• Examples include Oracle VirtualBox and VMware Workstation.

Thus, Type 1 hypervisors are preferred for production workloads, while Type 2 hypervisors are suitable for personal and development environments.

x86 virtualization allows multiple operating systems to execute concurrently on a physical computer based on the x86 instruction-set architecture.

Traditional x86 virtualization was difficult because some sensitive instructions did not automatically trap when executed without sufficient privilege. This made it difficult for a hypervisor to intercept and safely emulate every privileged operation.

Major challenges include:

• Privilege management: Guest operating systems expect to execute at the highest privilege level, which is normally controlled by the hypervisor.
• Sensitive instructions: Certain instructions behave differently depending on the processor privilege level.
• Memory virtualization: Guest virtual addresses must be translated into guest physical addresses and then into machine addresses.
• I/O virtualization: Multiple virtual machines must securely share physical devices.
• Interrupt handling: Hardware interrupts must be routed to the correct virtual machine.
• Performance overhead: Instruction translation, context switching, and device emulation can reduce performance.

Techniques such as binary translation, paravirtualization, and hardware-assisted virtualization using Intel VT-x or AMD-V address these challenges.

Full virtualization:

• Presents a complete virtual hardware environment to the guest.
• Allows an unmodified guest operating system to run.
• Uses techniques such as binary translation and trap-and-emulate.
• Provides strong compatibility but may introduce additional overhead.

Paravirtualization:

• Requires the guest operating system to be modified so that it is aware of virtualization.
• Replaces difficult privileged operations with calls called hypercalls to the hypervisor.
• Can provide high performance because fewer instructions need to be emulated.
• Cannot directly support an unmodified operating system.

Hardware-assisted virtualization:

• Uses processor features such as Intel VT-x and AMD-V.
• Allows the hypervisor to run guest operating systems in a dedicated execution mode.
• Supports unmodified guest operating systems.
• Reduces the complexity and overhead of software-based instruction translation.

In summary, full virtualization emphasizes compatibility, paravirtualization emphasizes efficient cooperation between the guest and hypervisor, and hardware-assisted virtualization uses processor support to achieve both compatibility and improved performance.

The major types of virtualization are:

• Server virtualization: Divides a physical server into multiple virtual machines. For example, a server may host separate Linux and Windows virtual machines.
• Operating-system-level virtualization: Runs isolated user-space environments that share the host kernel. Docker containers are a common example.
• Desktop virtualization: Hosts desktop environments centrally and delivers them to remote users. Virtual Desktop Infrastructure is an example.
• Application virtualization: Runs an application in an isolated environment without a conventional local installation.
• Storage virtualization: Combines multiple physical storage devices into a single logical storage pool.
• Network virtualization: Creates logical networks independent of the physical infrastructure. VLANs and software-defined networks are examples.
• Data virtualization: Presents data from multiple sources through a unified logical view without necessarily moving the data.

Each type separates a logical resource from its physical implementation, improving flexibility, utilization, isolation, or manageability.

Virtual machines:

• Virtualize physical hardware through a hypervisor.
• Each virtual machine includes a complete guest operating system and kernel.
• Provide strong isolation between workloads.
• Can run different operating systems on the same host.
• Usually require more memory, storage, and startup time.

Containers:

• Virtualize the operating system at the user-space level.
• Share the host operating system's kernel.
• Package an application with its libraries and dependencies.
• Start quickly and consume fewer resources.
• Require compatibility with the host kernel and generally provide weaker isolation than virtual machines.

Virtual machines are suitable when strong isolation or different operating systems are required. Containers are suitable for lightweight, portable, and rapidly scalable applications. They can also be combined by running containers inside virtual machines.

A Virtual Local Area Network, or VLAN, is a logical subdivision of a physical network that creates separate broadcast domains regardless of the physical location of connected devices.

VLAN operation is commonly based on IEEE 802.1Q tagging. A VLAN identifier is added to an Ethernet frame when the frame travels through a trunk link. Switches use this identifier to forward the frame only within the corresponding logical network. Communication between different VLANs requires a router or Layer 3 switch.

Benefits include:

• Isolation: Traffic belonging to different departments or applications is separated.
• Security: Unauthorized communication between logical groups can be restricted.
• Reduced broadcasts: Each VLAN forms a smaller broadcast domain.
• Flexibility: Virtual machines can be logically grouped without changing physical cabling.
• Simplified management: Administrators can apply policies to logical groups.
• Multi-tenancy: Cloud providers can separate network traffic belonging to different customers.

VLANs therefore improve network organization, security, and scalability in virtualized data centers.

A Virtual Storage Area Network, or VSAN, is a logical storage network created by grouping storage resources or partitioning a physical storage-area network into isolated virtual segments.

A VSAN supports storage virtualization by:

• Combining physical disks or storage devices into logical storage pools.
• Presenting virtual datastores or logical volumes to hosts and virtual machines.
• Separating storage traffic and administrative domains.
• Applying policies for capacity, availability, performance, and fault tolerance.
• Distributing or replicating data across multiple physical devices.
• Allowing storage capacity to be expanded without major changes to applications.

Its benefits include improved storage utilization, centralized management, simplified provisioning, scalability, workload isolation, high availability, and easier disaster recovery. Depending on the platform, VSAN may refer to logical segmentation of a Fibre Channel SAN or to software-defined storage built from disks attached to multiple servers.

VLAN and VSAN are both logical segmentation technologies, but they virtualize different resources.

Both technologies reduce dependence on physical layout and support centralized management. A VLAN controls how systems communicate over a network, whereas a VSAN controls how systems access and organize storage resources.

Distributed computing is a computing model in which multiple autonomous computers communicate through a network and coordinate their activities to solve a problem or provide a service.

Essential characteristics include:

• Multiple autonomous nodes: Each computer has its own processor, memory, and operating system.
• Communication through messages: Nodes exchange data using a network rather than shared physical memory.
• Resource sharing: Hardware, software, services, and data can be shared among users and applications.
• Concurrency: Multiple processes execute at the same time on different nodes.
• No global clock: Exact time synchronization across all nodes cannot be assumed.
• Independent failures: One node or communication link may fail while others continue operating.
• Transparency: The system attempts to appear as a single coherent platform.
• Scalability: Nodes and resources can be added to support increasing workloads.

Examples include cloud platforms, distributed databases, content-delivery networks, and large-scale web services.

The main goals of distributed computing are:

• Resource sharing: Make remote data, services, storage, and processors accessible.
• Scalability: Support growth in users, workload, data, and geographical coverage.
• Reliability and availability: Continue providing services despite component failures.
• Performance: Divide work among nodes and place services near users.
• Transparency: Hide distribution, location, replication, migration, and failures where practical.
• Openness: Support standard interfaces and interoperability.

Major challenges include:

• Network latency and limited bandwidth during communication.
• Partial failures, where some components fail while others remain operational.
• Concurrency control when several nodes access shared data.
• Consistency management across replicated data.
• Security across machines and untrusted networks.
• Clock synchronization and event ordering in the absence of a global clock.
• Heterogeneity of hardware, operating systems, and protocols.
• Load balancing and efficient resource allocation.

A successful distributed system must balance performance, consistency, availability, security, and implementation complexity.

Parallel computing uses multiple processing elements simultaneously to solve parts of a single problem. Distributed computing uses multiple autonomous networked computers to coordinate tasks or services.

Key differences are:

The models can overlap. For example, a computing cluster is a distributed system whose nodes may perform a parallel scientific computation.

Parallel computing is the simultaneous execution of multiple operations using more than one processing element to reduce the time required to solve a problem.

Task parallelism:

• Different processors execute different tasks or functions.
• The tasks may operate on the same data or on different data.
• For example, one thread may read input, another may process it, and a third may write results.

Data parallelism:

• The same operation is applied simultaneously to different portions of a dataset.
• For example, an image can be divided into blocks, with each processor applying the same filter to a separate block.

A program may combine both approaches. Parallel performance depends on workload division, synchronization, communication overhead, load balance, and the fraction of the program that can execute concurrently.

Flynn's classification organizes computer architectures according to the number of simultaneous instruction and data streams.

• SISD, or Single Instruction Single Data: One processor executes one instruction stream on one data stream. A traditional sequential computer is an example.
• SIMD, or Single Instruction Multiple Data: Many processing elements execute the same instruction on different data items. Vector processors and graphics processing units are common examples.
• MISD, or Multiple Instruction Single Data: Multiple processing units perform different operations on the same data stream. This model is uncommon but may appear in specialized fault-tolerant or pipeline systems.
• MIMD, or Multiple Instruction Multiple Data: Multiple processors execute different instruction streams on different data. Multicore systems, multiprocessors, and clusters are examples.

SIMD is effective for regular data-parallel workloads, while MIMD supports more general and independent parallel tasks. Modern systems may combine classifications, such as MIMD multicore processors whose individual cores contain SIMD execution units.

Shared-memory architecture:

• All processors access a common address space.
• Processes or threads communicate by reading and writing shared variables.
• Synchronization uses locks, semaphores, barriers, or atomic operations.
• Programming is often simpler, but memory contention and cache coherence can limit scalability.
• Multicore and symmetric multiprocessor systems are examples.

Distributed-memory architecture:

• Each processor or node has its own private memory.
• A processor cannot directly access another processor's memory.
• Data is exchanged explicitly using message passing.
• It scales to a larger number of nodes but requires careful data partitioning and communication.
• Clusters and massively parallel processing systems are examples.

Hybrid architectures combine both models. For example, cluster nodes may communicate through message passing, while cores inside each node communicate through shared memory.

Amdahl's law states that the theoretical speedup of a program is limited by the fraction of the program that must execute sequentially.

Let:

• be the execution time on one processor.
• be the parallelizable fraction of the program.
• be the sequential fraction.
• be the number of processors.

After normalization, the sequential part takes units of time. Ideally, the parallel part takes units. Therefore, the execution time using processors is:

Speedup is defined as:

Substitution gives:

As approaches infinity:

Thus, if , the maximum possible speedup is , regardless of how many processors are added. The law shows why reducing the sequential portion is essential for scalable parallel computing.

Major distributed-system architectural models include:

• Client-server model: Clients request services and servers process requests and return responses. Web applications commonly use this model.
• Multi-tier model: Presentation, application logic, and data storage are separated into tiers to improve scalability and maintainability.
• Peer-to-peer model: Each node can act as both a client and a server. Resources and responsibilities are distributed among peers.
• Cluster model: A group of connected computers works together to provide high performance or high availability.
• Service-oriented architecture: Applications are composed of reusable services accessed through defined interfaces.
• Microservices architecture: An application is divided into small, independently deployable services that communicate through lightweight protocols.
• Publish-subscribe model: Producers publish events to topics, and interested consumers receive them asynchronously.

The appropriate model depends on workload, scale, reliability, communication pattern, administrative control, and consistency requirements.

Transparency hides the distributed nature of a system from users and applications. It may hide resource location, replication, migration, concurrent access, or failures.

Scalability is the ability to maintain acceptable performance as users, nodes, data, or geographical coverage increase. It can be supported through decentralization, partitioning, caching, and replication.

Fault tolerance is the ability to continue operating despite hardware, software, or network failures. Techniques include redundancy, replication, checkpointing, retry mechanisms, and failover.

Consistency determines whether nodes and users observe compatible values for shared or replicated data. Strong consistency provides a single up-to-date view but may increase latency or reduce availability. Eventual consistency allows replicas to differ temporarily but requires them to converge when updates stop.

These properties are related and may involve tradeoffs. For example, replication can improve availability and read performance while making consistency management more difficult.

These paradigms overlap. A cloud data center may use virtualization for isolation, clusters for availability, distributed systems for service delivery, and parallel processors for workload acceleration.

Virtualization provides a flexible resource layer between physical infrastructure and distributed or cloud applications.

Contributions and benefits:

• Resource consolidation: Multiple workloads share the same physical server, increasing utilization.
• Isolation: Tenants and applications operate in separate virtual environments.
• Elastic provisioning: Virtual machines can be created, resized, cloned, or removed according to demand.
• Portability: Encapsulated virtual machines can be migrated between compatible hosts.
• Load balancing: Workloads can be distributed across hosts to improve performance and utilization.
• High availability: Failed workloads can be restarted on another physical host.
• Disaster recovery: Snapshots, replication, and virtual-machine images simplify recovery.
• Multi-tenancy: Providers can securely offer shared infrastructure to multiple customers.

Limitations:

• Hypervisor and device-emulation overhead may reduce performance.
• Resource contention can cause unpredictable performance.
• A compromised hypervisor can affect multiple workloads.
• Virtual-machine sprawl increases administrative complexity.
• Licensing, monitoring, storage, and network configuration remain operational challenges.

Therefore, virtualization is a key cloud-enabling technology, but distributed services must still handle failures, consistency, scalability, and network delays at the application level.