CloudComputingandSecurity-BIS613D

Module – 01 [Chapter 1: 1.1 to 1.5 TB 01]

DistributedSystemModelsandEnablingTechnologies
1.1ScalableComputingovertheInternet:
 Computing has evolved over 60 years in architecture, operating systems, network
connectivity, andworkloads.
 Modern systems use parallel and distributed computing across multiple computers over
the Internet to solve large-scale problems.
 Distributed computing is data-intensive and network-centric.
 These advancements have significantly enhanced the quality of life and information
services.
1.1.1 TheAgeofInternetComputing
 Billions use the Internet daily, requiring supercomputers and data centers to handle
high-performance computingforlarge-scale users.
 The Linpack Benchmark isno longerideal formeasuring performance inthis context.

 Computing clouds now demand high-throughput computing (HTC) systems using

parallel and distributed computing.
 Data centers must be upgraded withfast servers, storage, andhigh-bandwidthnetworks
to support emerging technologies in network-based computingand webservices.
[Link] ThePlatformEvolution
 Computer technology has evolved through five generations, each lasting 10-20 years,
overlapping byabout 10 years.
 1950-1970: Mainframes like IBM 360 and CDC 6400 met the needs of large businesses
and governments.
 1960-1980: Mini- computers (e.g., DEC PDP 11, VAX Series) became popular with small
businesses and colleges.
 1970-1990: Personal computers usingVLSI microprocessors became widespread.

 1980-2000: Portable computers and pervasive devices emerged in both wired and
wireless applications.
 Since 1990, HPC and HTCsystems have grown in use, hidden in clusters, grids, or
Internet clouds, serving consumers and web-scale services.
 The trendis toleverage shared web resources andvast data overthe Internet.

BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D
 Supercomputers are being replaced by clusters of cooperative, homogeneous compute
nodes.
 HTC systems like P2P networks are used fordistributed file sharing and content delivery,
withgloballydistributed machines.
 P2P, cloud computing, and web services focus more on HTC than HPC, leading to the
development of computational and data grids.

[Link] High-PerformanceComputing
 HPC systems have focused on raw speed, increasing from Gflops in the 1990s to Pflops
by2010.
 This growth was driven by demands from scientific, engineering, and manufacturing
sectors.
 The Top 500 most powerful supercomputers are ranked by floating-point speed in the
Linpackbenchmark.
 However, supercomputers are used byless than10% of all computerusers.

 Most users now rely on desktop computers or large servers for Internet searches and
market-driven tasks.
[Link] High-ThroughputComputing

BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D
 High-end computing is shifting from the HPC to HTC paradigm, focusing on high-flux
computing for tasks like Internet searches and web services.
 The performance goal now measures throughput—the number of tasks completed per
unit oftime.
 HTC needs improvements in batch processing speed, cost, energy efficiency, security,
and reliability at datacenters.
[Link] ThreeNewComputingParadigms
 In 1969, Leonard Kleinrock predicted that computer networks would evolve into utilities
servicinghomes and offices.
 Overtime, the definitionof "computer" has evolved:

o 1984: JohnGage said, “The networkis the computer.”

o 2008: DavidPattersonstated, “The data center is the computer.”
o Recently, Rajkumar Buyyanoted, “The cloud is the computer.”
[Link] ComputingParadigmDistinctions
Thehigh-techcommunityhasdebatedthedefinitionsofcentralized,parallel, distributed,andcloud
computing.
 Centralized computing is the opposite of distributed computing.

 Parallel computing overlaps significantlywithdistributedcomputing.

 Cloud computing overlaps with all three: distributed, centralized, and parallel
computing.
Centralized Computing:
 All resources (processors, memory, storage) are centralizedinone physical system.

 Resources are sharedandtightlycoupled within asingle integrated OS.

 Manydata centers and supercomputers use centralized systems.

 These systems are used inparallel, distributed, andcloud computingapplications.

Parallel Computing
 Processors are either tightly coupled with centralized shared memory or loosely coupled
withdistributedmemory.
 Interprocessorcommunicationoccurs through sharedmemory ormessage passing.

 Systemscapable of parallel computing are knownas parallel computers.

BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D
 Programs running on parallel computers are called parallel programs, and the process is
knownas parallel programming.
Distributed Computing
 Involves multiple autonomous computers, each with its own private memory,
communicatingvia anetwork.
 Informationexchange occurs throughmessage passing.

 Programs running in a distributed system are called distributed programs, and writing
themis knownas distributed programming.
CloudComputing
 AnInternet cloud canbe centralized ordistributed.

 It appliesparallel or distributed computing, orboth.

 Clouds canuse physical orvirtualized resources overlarge data centers.

 Some view cloudcomputing as a form of utilityorservice computing.

[Link] DistributedSystemFamilies
 Since the mid-1990s, P2P networks and clusters have formed computational grids and
data gridsforwide-area computing.
 Interest in Internet cloud resources for data-intensive applications has increased, with
clouds shifting desktop computing to service-oriented computing using server clusters
and databases.
 Grids and clouds focusonresource sharing in hardware, software, and data.
 These systems aim toexploit parallelism and concurrencyacross manymachines.
 P2P networks can involve millions of client machines, and cloud computing clusters may
have thousands of nodes.
 Future HPC and HTC systems will require multicore or many-core processors to handle
large numbers of computing threads, emphasizing parallelism, scalability, efficiency, and
reliability.
 Efficiency depends on speed, programming, and energy factors (throughput per watt),
withdesign goals centered onthroughput, scalability, andenergyefficiency:
o Efficiency: Resource utilizationinHPC and job throughput, data access, and
power efficiencyinHTC.
o Dependability: ReliabilityandQoS under failure conditions.
o Adaptation: Supportingbillions of job requests across large datasetsandcloud
resources.
BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D
o Flexibility: Running efficiently in both HPC (science) and HTC (business)
applications.
1.1.2 ScalableComputing Trendsand NewParadigms
 Predictable technology trends drive computing applications, with designers aiming to
forecast future system capabilities.
 Moore’ s law (processor speed doubles every 18 months) and Gilder’ s law (network
bandwidthdoubles annually) have shaped technology, thoughfuture validityis uncertain.
 The price/performance ratio of commodity hardware, driven by desktop, notebook, and
tablet markets, hasboosted large-scale computing adoption.
 Distributed systems focus on resource distribution, concurrency, and high degrees of
parallelism(DoP).
[Link] DegreesofParallelism
 Bit-Level Parallelism (BLP): Transition from bit-serial to word-level processing over
time (e.g., 4-bit to 64-bit CPUs).
 Instruction-LevelParallelism (ILP): Executes multiple instructions simultaneouslyusing
pipelining, superscalar computing, and multithreading. ILP requires branch prediction,
dynamic scheduling, and compilersupport.
 Data-Level Parallelism (DLP): Uses SIMD and vector machines for parallel data
processing, requiring more hardware andcompiler support.
 Task-Level Parallelism (TLP): Explores parallelismwithmulticore processors, but faces
challenges inprogramming and efficient execution.
 Job-Level Parallelism (JLP): As we move to distributed processing, parallelism
increases to job-level granularity, built on fine-grainparallelism.

[Link] InnovativeApplications
 HPC and HTC systems aim for transparency in data access, resource allocation, process
location, concurrency, job replication, and failure recovery.
 Key applications drive the development of parallel and distributed systems across
domains like science, engineering, business, healthcare, education, military, and
government.
BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D
 All applications require computing economics, web-scale data, reliability, and scalable
performance.
 Example: Distributed transaction processing in banking demands consistency in
replicated transaction records, despite challenges like lack of software support, network
saturation, and security threats.

1.1.2.3TheTrendtowardUtilityComputing

1. Computing Paradigms
 Ubiquitous, reliable, scalable, autonomic, and composable withQoS/SLAs.

 Alignwiththe computerutilityvision.

2. Utility Computing
 Business model where customers pay forcomputing resources.

 Grid/cloud platforms are utilityservice providers.

3. Cloud vsUtility Computing

 Cloud is broader, withdistributed applications onedge servers.

4. Technological Challenges
BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D
 Advancesneededinprocessors, storage, OS, virtualization, programmingmodels,
and resource management.
1.1.2.4TheHypeCycleofNewTechnologies
 HypeCyclePhases:
1. Innovation Trigger: Initial breakthroughorinnovation.
2. Peak of Inflated Expectations: Hype and exaggerated expectations.
3. Trough of Disillusionment: Disappointment as technology fails to meet
expectations.
4. Slope of Enlightenment: Understandingandacceptance grow.
5. Plateau of Productivity: Mainstream adoption and maturity.
 Symbols forAdoptionTimeline:
o Hollow Circles: Mainstream adoptionin<2 years.
o Gray Circles: Mainstream adoption in2-5 years.
o Solid Circles: Mainstreamadoptionin5-10 years.
o Triangles: Adoption>10 years.
o Crossed Circles: Technologies becomingobsolete before plateau.
 2010 Example:
o Consumer-generated media: In disillusionment, expected to reach adoption in<2
years.
o Internet micropayments: Moving fromenlightenment tomaturity in 2-5 years.
o 3Dprinting: 5-10 years from rising expectations to mainstream adoptionJG.
o Meshnetworksensors:>10 years toplateau.
 Cloud Technology: Expected to move from peak expectations to productivity in 2-5
years.
 Obsolete Technologies:
o Broadband over power line: Expected to become obsolete before leaving the
valley.
 Promising Technologies (2010):
o Cloud, biometric authentication, interactive TV, speech recognition, predictive
analytics, and media tablets.

BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D

1.1.3TheInternet of ThingsandCyber-Physical Systems

1.1.3.1The Internet ofThings (IoT) -
 Definition: The IoT refers to the networked interconnection of everyday objects, devices,
and computers through wireless sensors, like RFID orGPS.
 Introduction: The concept of IoT wasintroduced in1999 at MIT.

 IP Addresses: With IPv6, 2128 IP addresses are available, enabling identification of all
objects on Earth, including computers and devices.
 Scale of IoT:
o Estimatedthat each personwill be surrounded by1,000 to 5,000 objects.
o The IoT needs totrack 100 trillionobjects (static ormoving)simultaneously.
 Addressability: Universal addressability is required for all IoT objects to ensure they are
identifiable, searchable, and manageable.
 Geographical Development: The IoT is more heavilydeveloped inAsia and Europe.

 CommunicationPatterns:

1. H2H(Human-to-Human)
2. H2T(Human-to-Thing)
BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D
3. T2T(Thing-to-Thing)
 Dynamic Connectivity: Connections are made at any time (day, night, indoors, outdoors,
onthe move)and anyplace (PC, mobile phones, etc.)withlow cost.
 IoTGrowth: The IoT is still initsearlystages, with many prototypes in limited areas.
 Cloud Computing: Cloud technologies are expected to support fast, efficient, and
intelligent interactions amonghumans, machines, and objects.
 Smart Earth Vision: The goal is to create intelligent cities, efficient resources (water,
power, transport), green IT, better schools, healthcare, and more, though achieving this
globallywill take time.
1.1.3.2Cyber-Physical Systems(CPS)-
 Definition: A CPS integrates computational processes withthe physical world, combining
cyber (computational)and physical (real-world)components.
 Technologies: Merges computation, communication, and control into a feedback system
betweenthe physical and informationworlds.
 Focus: IoTconnects physical objects, while CPS explores virtual realityapplicationsinthe
physical world.
 Impact: CPS may transform interactions with the physical world, similar to how the
Internet transformed virtual interactions.
1.2TechnologiesforNetwork-BasedSystems-
 Focus on hardware, software, and network technologies for distributed computing
systems.
 Emphasis on designing distributed operating systems to manage massive parallelism in
distributed environments.
1.2.1Multicore CPUs andMultithreadingTechnologies
1.2.1.1Advances inCPU Processors – infig:1.4
 Multicore Architecture: Modern CPUs feature multiple cores (dual, quad, etc.) to
exploit ILP andTLP.
 Performance Growth: CPU speed has grown from 1 MIPS in 1978 (VAX 780) to 22,000
MIPS in2008(Sun Niagara 2), followingMoore's law.
 Clock Rate Limitation: The clock rate, once increasingto 4 GHz, nowfaceslimitsdue to
power and heat constraints.
BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D
 ILP Exploitation: Techniques like superscalar architecture, branch prediction, and
speculative executionare usedto enhance ILP.
 GPU Advances: GPUs utilize many-core architecture to exploit DLP and TLP, with
hundreds tothousands of simple cores.

1.2.1.2MulticoreCPUandMany-CoreGPUArchitectures-
Multi-Core andMany-CoreProcessors –
 Architecture: Multi-core CPUs have multiple cores, eachwithits ownL1cache, and
share anL2 cache.
 Future chipsmayalso include anL3 cache.
 High-End Processors: Examples include Intel i7, Xeon, AMDOpteron, and Sun Niagara.
 Multithreading: Cores canbe multithreaded, suchas NiagaraII with8coresand8
threads per core (64 threads total).
 Performance: The Intel Core i7990x (2011)reported 159,000 MIPS executionrate.

BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D
 Future of Multicore CPUs: CPUs may scale to hundreds of cores, but are limited in
exploiting massive DLP due to memorywall issues.
 Many-Core GPUs: GPUs with hundreds of thin cores are developed to handle massive
parallelism.
 Processor Trends: x-86 processors are replacing RISC processors in high-performance
systems, including data centers andsupercomputers.
 Heterogeneous Chips: Future processors may combine fat CPU cores and thin GPU
cores onthe same chip.
1.2.1.3MultithreadingTechnology-
 ProcessorTypes:
1. SuperscalarProcessor: Single-threaded withmultiple functional units.
2. Fine-Grain MultithreadedProcessor: Switches threads everycycle.
3. Coarse-Grain Multithreaded Processor: Executes multiple instructions from the
same thread forseveral cyclesbefore switching.
4. Dual-Core CMP: Twocores, eachsingle-threaded andtwo-way superscalar.
5. SMT Processor: Simultaneously schedules instructions from multiple threads in
one cycle.
 Execution Patterns: These processors handle instructions from different threads in
distinct ways, withsome more efficient thanothers inutilizingfunctional units.
1.2.2GPUComputingtoExascaleandBeyond-
 GPU Definition: A graphics coprocessor or accelerator used to offload graphics tasks
fromthe CPU, first introduced byNVIDIAin1999 (GeForce 256).
 GPU vs CPU: GPUs have hundreds of cores, while CPUs like Xeon X5670 have only six
cores.
 Parallelism: GPUs utilize a throughput architecture to execute many concurrent threads
slowly, unlike CPUs that focus onfast executionofa single thread.
 GPGPU: General-purpose computing on GPUs (GPGPUs) has gained traction in HPC,
withNVIDIA’ sCUDA model supporting thisshift.
1.2.2.1HowGPUsWork-
 GPU Evolution: Early GPUs were CPU coprocessors; modern NVIDIA GPUs have 128
cores, eachhandling8threads, allowing 1,024 concurrent threads.
 Parallelism: GPUs excel at massive parallelism, executing many threads compared to
CPUs withfewerthreads.

BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D
 Optimization: CPUs focus on latency, while GPUs prioritize high throughput and manage
on-chip memory.
 Applications: GPUs power HPC systems and supercomputers for parallel processing
beyond graphics, handlingdata-intensive calculations.
 Usage: GPUs are common in mobile phones, game consoles, PCs, and servers, with
NVIDIA CUDA Tesla/Fermi used inHPC systems.
1.2.2.2GPUProgrammingModel -
CPU-GPU Interaction: The CPU offloads floating-point computations to the GPU, which
has manysimple processing cores (multiprocessors)for massive parallelism.

CUDA Programming: NVIDIA’ s CUDA is used for large-scale data processing in GPUs
like the GeForce 8800 or Tesla Fermi.

BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D

Fermi GPU Architecture: The Fermi GPU (2011) has 16 streaming multiprocessors (SMs),
eachwith512CUDA cores, totaling 3 billiontransistors.
GPU Capabilities: Fermi GPUs can handle up to 82.4 Tflops, and future GPUs may reach
Exascale systems (1Eflop).
Challenges for Exascale Computing: Key challenges include energy, memory,
concurrency, and system resiliency.
1.2.2.3PowerEfficiency of theGPU
Bill Dally's View: Power efficiency and parallelism are keyadvantagesof GPUs overCPUs.
Exaflops System Requirement: 60 Gflops/watt per core is needed to run an exaflops
system.
Power Consumption:
 CPU: 2 nJ/instruction.
 GPU: 200 pJ/instruction (10 timesmore efficient thanCPU).
 Optimization Differences:
 CPU: Optimizedfor latencyin caches and memory.
 GPU: Optimized forthroughput withexplicit memorymanagement.

BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D
 2010 Performance/PowerRatio:
 GPU: 5 Gflops/watt per core.
 CPU: Less than1Gflop/watt percore.
 Challenges forFuture Supercomputers:
 Data movement dominates powerconsumption.
 Need foroptimized storage hierarchiesand memory tailored to applications.

 KeyFocus Areas:
 Development of self-aware OS and runtimes.
 Creationof locality-aware compilers and auto-tunersfor GPU-based systems.
 Conclusion: Both power efficiencyand software development are crucial for the future of
parallel and distributedcomputing systems.
1.2.3Memory,Storage,andWide-AreaNetworking
1.2.3.1Memory Technology(Fig 1.10)
 DRAM Growth: DRAM capacity grew from 16KB in 1976 to 64GB in 2011, increasing 4x
everythree years.
 Memory Access: Memory access times have improved little, and the memory wall issue
worsensas processors get faster.

BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D

 Hard Drive Growth: Hard drive capacitygrewfrom260MB in 1981to 250GB in2004, and
Seagate's Barracuda XT reached 3TB in 2011(10x increase everyeight years).
 Future Trends: Disk array capacity will continue to grow. The gap between processor
speed and memory capacity may worsen, further limiting CPU performance due to the
memory wall.

1.2.3.2DisksandStorageTechnology (Fig 1.10)

 Disk Growth: Disk arrays surpassed 3TB capacity beyond 2011, growing 7 orders of
magnitude in 33 years.
 Flash and SSD Impact: Flash memory and SSDs are shaping the future of HPC and HTC
systems. SSDs can handle 300,000 to 1 million write cycles per block, lasting several
years evenwithheavy usage.
 Power and Cooling Limitations: Power consumption and cooling will limit large system
development. Power increases linearly with clock frequency and quadratically with
voltage.
 Future Trends: Clock rates cannot increase indefinitely, and lower voltage supplies are in
demand.
 Jim Gray's Quote: "Tape units are dead, disks are tape units, flashes are disks, and
memory are caches now,"highlighting the future of storage technology.
 2011 SSDs: Despite advances, SSDs remain too expensive to replace traditional disk
arraysin the storage market.
1.2.3.3System-AreaInterconnects (Fig 1.11)
BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D
 Small Cluster Connections: Small clusters are typically interconnected via Ethernet
switches or local areanetworks (LAN).
 NetworkTypes:
o LAN: Connects client hosts tobigservers.
o Storage Area Network (SAN): Connects servers to network storage (e.g., disk
arrays).
o Network-Attached Storage (NAS): Connectsclient hosts directlyto disk arrays.
 Cluster Setup: In clusters without large distributed storage, a multiport Gigabit Ethernet
switchwithcoppercablescan link machines.
 Availability: All three network types are commerciallyavailable.

1.2.3.4Wide-AreaNetworking (Fig : 1.10 and 1.12)

 Ethernet Growth: Ethernet bandwidth grew from 10 Mbps in 1979 to 1 Gbps in 1999, and
40– 100 GE in2011, with1Tbpslinks expected by 2013.
 Bandwidth by Year: In 2006, international, national, and desktop connections had
bandwidths of 1,000, 100, 10, and 1Gbps, respectively.

 Performance Growth: Network performance increased by a factor of two annually,

outpacingMoore's law.
BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D
 Future Trends: High-bandwidth networking will enable more concurrent computers and
the development of massivelydistributedsystems.
 HPC Interconnects: IDC 2010 report predicts InfiniBand and Ethernet as the main
interconnects inHPC. Gigabit Ethernet is widely used indata center serverclusters.
1.2.4Virtual MachinesandVirtualizationMiddleware
1.2.4.1VirtualMachines (Fig 1.12)
 Host Machine and VM: A host machine has physical hardware running an OS (e.g.,
Windows on x86 architecture). Virtual machines (VMs) run specific applications on
virtual resourcesmanaged bya guest OS.
 Virtual Machine Monitor (VMM): A middleware layer between the VM and the host,
called a hypervisor, manages the VM. It can run in privileged (bare-metal VM) or
non-privilegedmode (host VM).
 VM Types:
o Bare-metal VM: The hypervisor directly handles hardware (e.g., XEN on x86
running Linux).
o Host VM: The VMM runsin non-privileged mode, with the host OS unchanged.
o Dual Mode VM: The VMM has parts running at user and supervisor levels, with
possible hostOS modifications.
 Benefits: VMs offer hardware independence, allowing applications and OS to be bundled
as virtual appliances and runondifferent hardware platforms.
1.2.4.2VMPrimitiveOperations (Fig 1.13)
 VMM and VMAbstraction: The VMMprovides aVM abstraction, allowing standardOSes
like Windows or Linux to runasif onphysical hardware.
 First, the VMs can be multiplexed between hardware machines, as shown in Figure
1.13(a).
 Second, aVM canbe suspended and stored in stable storage, as showninFigure 1.13(b).
 Third, a suspended VM can be resumed or provisioned to a new hardware platform, as
showninFigure 1.13(c).
 Finally, aVM canbe migrated from one hardware platform to another, as showninFigure
1.13(d).
 VM Operations (as described by Mendel Rosenblum):
o Multiplexing: VMscan be shared between hardware machines.
o SuspensionandStorage: VMs canbe suspended and stored instable storage.
o Resumption: Suspended VMs canbe resumed ormoved tonew hardware.
o Migration: VMs canbe migratedbetweenhardware platforms.
 Benefits: These operations enable flexibility, enhance server resource utilization, and help
eliminate server sprawl by consolidating multiple functions on the same hardware
platform.

BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D

1.2.4.3VirtualInfrastructures (Fig 1.14)

 Virtual Infrastructure: It connects physical resources (compute, storage, networking)to
applications inVMs, separating hardware and software.
 Benefits: It maps system resources dynamically to applications, reducing costs and
increasingefficiencyandresponsiveness.
 Examples: Serverconsolidation and containment are keyexamplesof virtualization
benefits.
 FurtherDiscussion: VMs and virtualizationsupport for clusters, clouds, and grids

1.2.5DataCenterVirtualization forCloudComputing (Fig 1.13)

 Cloud Architecture: Built withcommodityhardware, x86 processors, low-cost terabyte
disks, andGigabit Ethernet.
BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D
 DesignFocus: Prioritizes performance/price ratio, emphasizing storage and energy
efficiencyover sheer speed.
 Data Center Growth: As of 2010, about 43 millionservers were in use. Utilitycosts
exceed hardware costs afterthree years.
1.2.5.1DataCenterGrowthand Cost Breakdown(Fig 1.14)
 DataCenterSizes: Large data centers use thousands of servers, while smallerones
have hundreds.
 Cost Breakdown (2009IDC Report):
o 30%: IT equipment(servers, disks)
o 33%: Chiller
o 18%: UPS
o 9%: CRAC
o 7%: Power distribution, lighting, transformers
 Management and Maintenance: 60% of data center costsgo towardmanagement and
maintenance.
 Cost Trends: Serverpurchase costsremained stable, but electricityand cooling costs
rose from5% to 14% over 15 years.
1.2.5.2Low-Cost DesignPhilosophy
 Cost Considerations: High-endswitches and routers maybe too expensive fordata
centers. Commodityswitches, networks, and x86serversare preferredforcloud
computing.
 Software Layer: Manages network trafficbalancing, fault tolerance, andexpandability.
 NetworkTechnology: Ethernet is the primarynetwork technology used in cloud
computing datacenters.
1.2.5.3ConvergenceofTechnologies
Cloud computing results fromthe convergence of four keytechnologies:
1. Hardware virtualizationandmulti-core chips enable dynamic cloud configurations.
2. Utility and grid computing formthe foundationforcloudcomputing.
3. AdvancesinSOA, Web 2.0, andmashups drive cloudevolution.
4. Autonomic computing and data center automationsupport cloudgrowth.
Cloud computing addresses the "datadeluge" challenge, whichscience and societyface in
managingand analyzingvast amounts of data from various sources. This requires scalable file
systems, databases, algorithms, workflows, andvisualizationtools. As science becomes
data-centric, cloud computing plays acrucial role in data-intensive scientific discovery.
The CSTB recommended tools fordata capture, creation, and analysis, recognizing the
interactionbetween cloud computing and e-science, which relies onmulticore and parallel
computing. Cloud computing's flexible services and technologies like MapReduce offerdata
parallelismandfault tolerance, supporting scientific applications.

BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D
Data-intensive science, cloud computing, and multicore computing are converging,
transforming the next generation of computing in architecture, design, and programming
challenges. These advancements turndata intoknowledge, drivingmachine wisdominSOA.
1.3SystemModelsforDistributedandCloudComputing (Table 1.2)
Distributed and cloudcomputing systems consist of many autonomous nodes connectedby
SANs, LANs, or WANs. LAN switchescan link hundreds of machines, while WANs connect
multiple local clusters toformlarge-scale systems withmillions of computers.
These systems are highlyscalable and canachieve web-scale connectivity. Theyare classified
into fourtypes: clusters, P2P networks, computing grids, and cloud datacenters. These
systems mayinvolve hundreds tomillions of nodes working cooperativelyat various levels,
each serving different technical and applicationneeds.

Clusters dominate supercomputing, with 417of the Top 500 in 2009 using thisarchitecture,
formingthe base forgrids andclouds. P2P networks are preferred forbusiness but face
copyright issues in content industries. National gridswere underused due to unreliable
middleware. Cloud computing offers low cost and simplicityfor bothusers and providers.
1.3.1ClustersofCooperativeComputers
A computing clusteris a group of interconnected stand-alone computersworking together as a
single resource, achievingimpressive resultsinhandling heavy workloads and large data sets.
1.3.1.1ClusterArchitecture (Fig: 1.15)
BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D
A typical serverclusteruses alow-latency, high-bandwidthinterconnectionnetworklike SAN
(Myrinet) orLAN (Ethernet). For largerclusters, networks canbe built with multiple levels of
Gigabit Ethernet, Myrinet, or InfiniBand switches. Clustersare connected tothe Internet via a
VPN gateway, whichlocates the cluster. Clusters typically have looselycoupled nodes, each
withits ownOS, leadingto multiple system imagesacross autonomous nodes.

1.3.1.2Single-SystemImage
 Greg Pfistersuggests an ideal clustershould merge multiple systemimages intoa
single-systemimage (SSI).
 SSI is supportedbya cluster OS ormiddleware toshare CPUs, memory, andI/O across
all nodes.
 SSI creates the illusionofa unified, powerful resource, making the cluster appearasa
single machine to the user.
 Without SSI, a clusteris just a collection of independent computers.
1.3.1.3Hardware,Software,andMiddlewareSupport
 Cluster designprinciples forsmall andlarge clusters will be discussedinChapter2.
 Massive parallel processing (MPP)clusters dominate the Top 500 HPC list.
 Building blocks: computer nodes (PCs, workstations, servers, SMP), communication
software (PVM, MPI), and network interface cards.
 Most clusters use Linux OS and are interconnected via high-bandwidthnetworks
(Gigabit Ethernet, Myrinet, InfiniBand).
 Special middleware supports featureslike SSI and highavailability (HA).
 Both sequential and parallel applicationscan run; special parallel environments are
neededforcluster resource use.
 Distributed shared memory(DSM)maybe used toshare memoryacross servers.
 Achieving full SSI can be expensive ordifficult; manyclusters remainloosely coupled.
 Virtual clusters canbe dynamicallycreated via virtualization.

BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D
1.3.1.4MajorClusterDesignIssues (Table 1.3)
 Cluster-wide OS for full resource sharing is not yet available.
 Middleware or OS extensionsare used to achieve SSI at selected levels.
 Without middleware, clusternodes cannot collaborate effectively.
 Middleware is essential forhigh performance insoftware environments and applications.
 Keybenefitsof clusters: scalable performance, efficient message passing, high
availability, fault tolerance, andjobmanagement.

1.3.2GridComputingInfrastructures
 Overthe past 30 years, usershave experienced growth fromInternet toweb andgrid
computingservices.
 Internet services like Telnet allowlocal-to-remote computerconnections.
 Web services like HTTP provide remote access to web pages.
 Gridcomputing enablesclose interaction betweenapplications ondistant computers.
 Forbes projectsIT-based economy growth from$1 trillionin2001 to $20 trillion by2015,
drivenbythese services.
1.3.2.1Computational Grids (Fig 1.16)
 A computing grid combines computers, software, middleware, instruments, people, and
sensors, similar to an electric utilitygrid.
 Grids are built overLAN, WAN, or Internet backbones at regional, national, or global
scales.
 Theyserve as integratedcomputing resources and virtual platforms for virtual
organizations.
 Grids use workstations, servers, clusters, and supercomputers, withPCs, laptops, and
PDAs asaccess devices.
BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D

 Example: Acomputational grid spans multiple resource sites owned bydifferent

organizations, integrating various computing resources (e.g., workstations, servers,
clusters).
 The grid is built overIP broadband networks (LANs andWANs)and presentedas an
integratedresource pool.
 Special instruments (e.g., radio telescopes inSETI@Home)maybe involved.
 The grid integrates computing, communication, content, and transactions asservices for
enterprises and consumers.
 National andinternational gridslike NSF TeraGrid, EGEE, and ChinaGrid support
distributed scientific applications.
1.3.2.2GridFamilies (Table 1.4)
 Gridtechnologyrequires newdistributed computingmodels, software/middleware,
network protocols, and hardware.
 National gridprojects are followed by industrial platforms from companieslike IBM,
Microsoft, Sun, HP, Dell, Cisco, and others.
 The rise of Grid Service Providers (GSPs)and grid applications mirrors the growthof
Internet and web services.
 Grids are categorizedinto computational/data grids(national level) and P2P grids.

1.3.3Peer-to-PeerNetworkFamilies
BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D
 A client-serverarchitecture is a common distributed system where client machines
connect to a central serverfor tasks like computing, email, file access, and databases.
 The P2P architecture offers a client-oriented distributed model, as opposedto
server-based systems.
 Thissectioncovers P2P systemsat the physical and overlaynetwork(logical)levels.
1.3.3.1P2PSystems (Fig 1.17)
 InP2P systems, eachnode acts as both aclient anda server, contributing resources.
 Peer machines autonomouslyjoin or leave the system, withno central coordinationor
database.
 The systemis self-organizing withdistributed control, meaning no peer has a global view
of the entire network.
 The P2P network is ad hoc, formed using the TCP/IP and NAI protocols, and its size and
topologychange dynamically.

1.3.3.2OverlayNetworks (Fig1.17)
 Data items/filesare distributedacrosspeers in the network, forminganoverlay network
at the logical level.
 New peers joinbyadding their peerIDto the overlay, while leaving peershave theirID
removedautomatically.
 There are two typesof overlaynetworks:
o Unstructured: Randomgraph, where messages are sent via flooding, leading to
hightraffic andunpredictable searchresults.
o Structured: Follows a set topology, with rules for adding/removing nodes and
optimized routing mechanisms.
1.3.3.3P2PApplicationFamilies (Table 1.5)
P2P networks are classified into fourapplication-basedgroups:
BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D
1. Distributed File Sharing: Includes networks like Gnutella, Napster, andBitTorrent for
sharing digital content.
2. CollaborationNetworks: Forinstant messaging and collaborative tools like MSN, Skype.
3. Distributed P2P Computing: Used inapplications like SETI@home for collective
computing power.
4. Platform-Specific Networks: Platforms like [Link] fortasks suchas naming,
communication, andresource aggregation.

1.3.4CloudComputingovertheInternet
 Computational science is becoming data-intensive, requiring balancedsystems with
petascale I/O and networking.
 Future computingwill involve sending programs tolarge data sets instead of moving
data to workstations.
 Thistrend leadsto the rise of cloud computing, where software, hardware, anddata are
provided on-demand.
 IBMdefines cloudcomputing asa pool of virtualizedresources that canhost various
workloads.
 Cloud computing supports rapid scaling, self-recoveryfromfailures, and real-time
monitoring forresource management.
“A cloud is apool ofvirtualized computer resources. Acloudcan host a variety of
differentworkloads, including batch-style backend jobs andinteractive and
user-facing applications.”
1.3.4.1Internet Clouds (Fig 1.18)
 Cloud computing uses virtualizedplatforms withelastic resources, provisioning
hardware, software, and data dynamically.
 It shifts desktop computing to service-oriented platforms using serverclusters and large
databases.
 Cloud computing offerslowcost and simplicity forusersand providers, enabledby
machine virtualization.

BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D

 It is designed to support multiple user applications while ensuringsecurity, trust, and

reliability.
 Some view the cloud asa centralized resource pool, while others see it as adistributed
server cluster.
1.3.4.2TheCloudLandscape
 Traditional distributed systemsface performance bottlenecks like system maintenance,
poor utilization, and highupgrade costs.
 Cloud computing resolves these issues byproviding on-demandcomputing, offering
flexibilityand scalability.
 Cloud service modelsinclude:
o IaaS: Provides infrastructure like servers, storage, and networks; users manage
applications on virtual machines. Theuserdoesn't controlthecloudinfrastructurebut
canrequestandreleaseresourcesasneeded.
o PaaS: Offers platforms fordeploying user-built applications withintegrated tools,
APIs, and middleware.(e.g., Java, Python, Web 2.0, .NET).
o SaaS: Delivers browser-initiated software for business processes, consumer
relationship management (CRM), enterprise resources planning (ERP), human
resources (HR), andcollaborative applications., etc., withno upfront investment
required.
 Cloud deployment modes: private, public, managed, and hybrid, eachwithvarying
securitylevels and SLAs.

BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D

 Keybenefitsof cloud computing:

1. Energyefficiencyandsecure locations.
2. Improved resource utilization through peak-load sharing.
3. Separationof infrastructure andapplication management.
4. Cost reductioncompared to traditional computing.
5. Enhanced programming and development.
6. Efficient service anddata discovery.
7. Addressingprivacy, security, and reliabilityissues.
8. Clear service agreements, business models, andpricing policies.
1.4 SOFTWAREENVIRONMENTSFOR DISTRIBUTEDSYSTEMS ANDCLOUDS
1.4.1Service-OrientedArchitecture(SOA) (Fig: 1.20)
 Ingrids/web services, Java, and CORBA, an entityis aservice, Java object, or
CORBA distributed object.

 These architecturesbuild onthe OSI layers fornetworking abstractions.

 Higher-level environments are built on topof this base.
 These environments focusonentity interfacesand inter-entity communication.
BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D
 Theyrebuild the top fourOSI layersat the entitylevel, not the bit level.
1.4.1.1LayeredArchitectureforWebServicesandGrids
 Entity interfaces in distributed systems correspond to WSDL, Java methods, and CORBA
IDL specifications.
 These interfacesare linked withhigh-level communication systems like SOAP, RMI, and
IIOP.
 These communication systems support message patterns, fault recovery, and
specialized routing.
 Middleware like WebSphere MQ orJMS providesrichfunctionalityand virtualized
routing.
 Fault tolerance inWebServices Reliable Messaging (WSRM)mimics OSI layer fault
tolerance.
 Security often uses or reimplements concepts like IPsec orsecure sockets.
 Higher-level services support entitycommunication, suchas registries, metadata, and
management.
 Examplesof discoveryservices include JNDI, CORBATrading Service, UDDI, LDAP, and
ebXML.
 Management services include state andlifetime support, suchas CORBA Life Cycle and
Persistent states.
 Distributed systems offer higherperformance and better software functionseparation
and maintenance.
 CORBAandJava were popularinearlydistributed systems; now, SOAP, XML, and REST
are more common.
1.4.1.2WebServicesandTools (Fig: 1.20)
 Loose coupling and support forheterogeneous implementations make services more
attractive thandistributed objects.
 Two service architectures: web services and REST, eachwithdistinct approachesto
reliabilityand interoperability.
 Webservices: Aim to fully specifyall service aspects using SOAP for communication,
acting as a universal distributed OS. It has had mixed success due to protocol challenges
and implementationdifficulties.
 REST: Focuses on simplicityand delegates complexity to applicationsoftware, with
minimal headerinformationand opaque message bodies. It's suitable for rapid
technologyenvironments anduses "XML over HTTP"insteadof SOAP's XML schemas.
 Web services' ideas maystill be needed in mature systems but at adifferent stack level.
 InCORBAand Java, distributed entities are linked using RPCs. Composite applications
can be created byintegrating entities as objects (e.g., using RMI inJava or similar in
CORBA).
 "Grid" canreferto asingle service or acollectionof services, with sensors as
data-outputting entities and grids/clouds ascollections of services.
BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D

1.4.1.3TheEvolution of SOA (Fig: 1.21)

 Service-OrientedArchitecture (SOA) has evolved to build grids, clouds, interclouds, and
systems of systems.
 Sensor services (SS)collect rawdata, fromdevices like ZigBee, Bluetooth, WiFi, and
smartphones.
 Data interacts withvarious systems: grids, databases, compute clouds, storage clouds,
and filter clouds.
 Filter services (fs)remove unwanted data, forming a filter cloud.
 SOA processes rawdatainto useful information and knowledge, eventuallysupporting
intelligent decision-making.
 Most systems require a webinterface totransform raw dataintoknowledge via
compute, storage, filter, and discoveryclouds.

 Portals like OGFCE and HUBzero provide access to users, integratingwebservice and
Web 2.0 technologies.
1.4.1.4GridsversusClouds
 The boundarybetween grids andcloudsis blurring.
 Workflowtechnologies inweb services coordinate services forbusiness processeslike
two-phase transactions.
BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D
 Workflowapproaches include Pegasus, Taverna, Kepler, Trident, and Swift.
 Grids use static resources, while clouds emphasize elastic resources.
 Some researchers see the maindifference betweengridsand clouds as dynamic
resource allocationvia virtualizationandautonomic computing.
 A grid canbe built from multiple clouds, supportingnegotiatedresource allocationbetter
thanpure clouds.
 Thisleads toa systemof systems like cloud of clouds, grid of clouds, orinter-clouds in
SOA architecture.
1.4.2TrendstowardDistributedOperatingSystems
 Distributed systems have looselycoupled computers, eachrunningindependent
operating systems.
 To enable resource sharing and fast communication, adistributedOS is ideal for
managingresources efficiently.
 Thissystemis typicallyclosed and uses message passing and RPCsforcommunication
betweennodes.
 A distributed OS is crucial for improving performance, efficiency, and flexibilityin
distributed applications.
1.4.2.1DistributedOperatingSystems
 Tanenbaumidentifiesthree approaches for distributing resource management ina
system:
1. NetworkOS: Built over heterogeneous OS platforms, offeringlowtransparency,
mainlyrelyingonfile sharing for communication.
2. Middleware: Provides limited resource sharing, like MOSIX/OS for clustered
systems.
3. Distributed OS: Achieveshighersystemtransparency forbetterresource sharing.
 Table 1.6compares the functionalitiesof these approaches.

BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D

1.4.2.2AmoebaversusDCE
 DCE: A middleware-based systemfor distributed computing, developed bythe Open
Software Foundation(OSF).
 Amoeba: Developed at Free University, the Amoeba, DCE, and MOSIX2are research
prototypesprimarily used in academia, withno commercial success.
 Need forWeb-based OS: New operating systems are needed to support resource
virtualization indistributed environments, still an area of active research.
 Distributed OS Design: Should use a lightweight microkernel approach (like Amoeba)or
extend existing OS (like DCE withUNIX)tobalance resource management across
servers, freeingusers from most management tasks.
1.4.2.3MOSIX2forLinuxClusters
 MOSIX2 is a distributed OS witha virtualizationlayer inthe Linux environment.
 Providesa partial single-systemimage touserapplications.
 Supports bothsequential and parallel applications.
 Discovers resources andmigrates software processes amongLinux nodes.
 Canmanage Linux clusters orgrids of multiple clusters.
 Allows resource sharingamong multiple cluster owners.
 Exploredfor managing resourcesinLinux clusters, GPU clusters, grids, andcloudsusing
VMs.
1.4.2.4Transparencyin ProgrammingEnvironments(Fig1.22)
BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D
 Transparent computing infrastructure separates user data, applications, OS, and
hardware into fourlevels.
 User data is independent of applications.
 OS provides clearinterfaces and systemcalls forapplicationprogrammers.
 Infuture cloud infrastructure, hardware will be separatedfromthe OS by standard
interfaces, allowingusers to choose preferredOSes.
 Cloud applications (SaaS)allowusersto switchservices without binding datato specific
applications.

1.4.3Parallel andDistributedProgrammingModels
 Thissectioncovers fourdistributedcomputing programming models with scalable
performance and flexibility.
 Table 1.7summarizesthree models and associatedsoftware tools.
 MPI is the most popularmodel for message-passingsystems.
 Google’ s MapReduce and BigTable optimize resource use inclouds and datacenters.
 Service cloudsextend Hadoop, EC2, and S3for distributed computing overstorage
systems.

BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D

1.4.3.1Message-PassingInterface(MPI)
 Message-Passing Interface (MPI) is a primary standard for developing parallel and
concurrent programs on distributed systems.
 MPI is a libraryfor C orFORTRAN usedto write parallel programs.
 It supports clusters, grid systems, and P2P with upgraded web services and utility
computing applications.
 Parallel Virtual Machine (PVM) is another low-level primitive for distributed
programming, alongside MPI.
1.4.3.2MapReduce
 MapReduce is a web programming model forscalable data processing onlarge clusters.
 It is mainlyusedinweb-scale search and cloud computing applications.
 Users specify a Map function to generate key/value pairs and a Reduce function to
merge intermediate values withthe same key.
 MapReduce scales highly, handling terabytes of data across tens of thousands of
machines.
 Google executesthousands of MapReduce jobsdailyacross itsclusters.
1.4.3.3HadoopLibrary
 Hadoop is a software platform developed by Yahoo! for running applications on
distributed data.
 It scales easilytostore and processpetabytes of data.
 Hadoopis cost-effective, with anopen-source MapReduce versionminimizingoverhead.
 It processes data with high parallelism across many nodes and ensures reliability by
keeping multiple data copies forfailover.
BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D

1.4.3.4Open GridServicesArchitecture(OGSA)
 Gridinfrastructure development isdriven bylarge-scale distributed computing
applications requiring highresource anddata sharing.
 OGSA is astandard forpublic grid services, withGenesis II being akeyimplementation.
 Features include adistributedexecutionenvironment, PKI services, trust management,
and securitypolicies.

1.4.3.5GlobusToolkitsandExtensions
 Globus is amiddleware librarydevelopedbyU.S. Argonne National Laboratoryand USC
InformationScience Institute.
 Implements OGSA standards forresource discovery, allocation, and securityingrid
environments.
 Supports PKI certificates formultisite mutual authentication.
 GT4, the current version, hasbeeninuse since 2008, withIBMextendingit for business
applications.
1.5.1PerformanceMetricsandScalabilityAnalysis
[Link] Performance Metrics
 Performance Factors:
o MIPS (Million Instructions PerSecond)
o Tflops (TeraFloating-Point Operations per Second)
o TPS (Transactions PerSecond)
o Job response time
o Networklatency
 Preferred Network:
o Low-latency, high-bandwidthinterconnectionnetwork
 System Overhead:
o OS boot time
o Compile time
BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D
o I/O data rate
o Runtime support system
 OtherPerformance Metrics:
o QoS forInternet and web services
o System availabilityanddependability
o Security resilience against network attacks
[Link] Dimensions ofScalability
 ScalabilityinDistributed Systems:
o Size Scalability: Achieving higher performance by increasingmachine size
(processors, memory, I/O). Not all systems are equally scalable (e.g., IBM
BlueGene/L scaledto 65,000 processors).
o Software Scalability: Upgrading OS, compilers, libraries, and environments. Some
software may not work with larger systems and require testing and fine-tuning.
o ApplicationScalability: Matchingproblem size withmachine size. Increasing
problemsize (data or workload)can improve systemefficiency without
expanding machine size.
o Technology Scalability: Adapting to changes in technology(hardware and
software). Includes:
Time: Generation scalability, consideringthe impactof upgrading
components.
Space: Packaging, energyconcerns, and portability.
Heterogeneity: Use of components fromdifferent vendors, which maylimit
scalability.
[Link] Scalability versus OSImage Count
 Scalable Performance: Refers toachieving higherspeeds byadding more processors,
memory, disk capacity, or I/O channels.

 OS Image: The number of independent OS images in acluster, grid, P2P network, or

cloud.
BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D
 SMP (SymmetricMultiprocessor): A single systemimage, limitedto afew hundred
processors. Scalabilityis constrained by packaging and interconnect.
 NUMA(Non-UniformMemory Access): Composed of SMP nodes with distributed
shared memory. Can scale to afewthousand processors withmultiple OS images(e.g.,
2,048processors with32 OS images).
 Cluster: Canscale better thanNUMA, withthe numberof OS images basedonthe
numberof nodesused. Cloudscan be virtualized clusters, scaling upto thousands of
VMs.
 Grid: Mayhave fewerOS images thanprocessors (e.g., hundreds or thousands fewer).
 P2P Networks: Canscale to millions of peer nodes, but performance depends onQoS in
public networks.
[Link] Amdahl’ s Law
 Parallel Execution: Ona uniprocessor, the program takesT minutes. Whenparallelized
fornprocessors, α of the code is sequential (sequential bottleneck), and (1− α)canbe
parallelized.
 ExecutionTime Formula: The total executiontime is
o αT+(1− α)T/n where the first term is the sequential executiontime, and the
secondis parallel.
 Speedup: Amdahl’ s Lawdefines speedup as
 S=T/[αT+(1− α)T/n]= 1/[α+(1- α)/n] ------------------------------------------(1.1)
 Maximum Speedup: The maximumspeedup of nisachievedwhen α= 0, Asthe cluster
becomes sufficientlylarge, that is, n→ ∞, S approaches 1/α, anupper bound onthe
speedupS.
 Surprisingly, this upper boundis independent of the clustersize n.
 KeyInsight: Amdahl's Law shows that reducingthe sequential bottleneckis crucial for
maximizingspeedup, andsimplyincreasing the cluster size maynot yieldbetter
performance.
[Link] Problem with FixedWorkload
we have assumedthe same amount of workload forbothsequential and parallel executionof
the program witha fixed problem size or data set.
To execute a fixed workload onnprocessors, parallel processingmaylead to asystem
efficiencydefined as follows:
E=S/n=1/½αn+1− α----------------------------------------------------(1.2)
Very often the system efficiency isratherlow, especiallywhenthe clustersize is very large. To
execute the aforementionedprogramon aclusterwithn= 256 nodes, extremelylow efficiency
E=1/[0.25× 256+0.75]=1.5% isobserved. This isbecause onlya few processors (say, 4)are
kept busy, while the majorityof the nodes are left idling
BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D

1.5.1.6Gustafson’ sLaw
To achieve higherefficiency whenusing a large cluster, we must considerscaling the problem
size to matchthe cluster capability. This leads to the following speedup lawproposed byJohn
Gustafson, referred as scaled-workload speedup
Let W be the workloadina givenprogram. Whenusing ann-processorsystem, the user scales
the workload toW′ = αW+(1− α)nW. Note that onlythe parallelizable portionof the workload
is scaled ntimes in the second term. This scaledworkloadW′ is essentially the sequential
executiontime ona single processor.
The parallel executiontime of ascaledworkload W′ onn processors is defined bya
scaled-workload speedup as follows:
---------------------------------- (1.3)

1.5.2 System Availability

 System Availability: Dependent onhardware, software, and networkcomponents; any

failure that disruptsthe whole systemis asingle point of failure.
 Designingfor Availability: Avoid single pointsof failure by adding redundancy, improving
reliability, andensuringtestability.
 Effect ofSystemSize: Asthe systemsize increases, availability generallydecreases due
to higherchancesof failure and difficulty isolating issues.

BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D

 VulnerabilityinSystems: SMP and MPP systemsare vulnerable withcentralized

resources. NUMAmachines improve availabilitybyusing multiple OSes.
 Clusters, Clouds, and Grids: Clusters oftenhave high availabilitywithfailover capabilities.
Clouds, built fromvirtualizeddata centers, have similar availability. Grids offerhigher
availability due to fault isolationbut still experience decreasing availabilityas size grows.
 P2P Networks: P2P networks, withmanyclient machines, have highaggregationbut
operate with lowavailability, especiallyduring simultaneous node failures.
1.5.3 Network Threats and Data Integrity
[Link] Threats toSystems andNetworks
 Network Viruses: Widespread attackshave caused worm epidemics, affecting routers,
servers, andleading to billions inlosses across business, government, and services.
 Typesof Attacks:
o InformationLeaks: Loss of confidentiality.
o DataIntegrity Loss: Caused byuser alteration, Trojan horses, andspoofing
attacks.
o Denialof Service(DoS): Loss ofsystem operations and connections.
o UnauthorizedAccess: Lack of authenticationallowsillegitimate use of
resources.
 TargetedSystems: Data centers, P2P networks, and grid/cloudinfrastructures maybe
targeted.
 ProtectionNeeded: Clusters, grids, clouds, and P2P systems needprotection against
malicious intrusions, or they shouldn’ t be trusted for outsourced work.

BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D

 Impact: Attacks candestroyhosts, networks, and storage resources, hindering the

acceptance ofpublic-resource computing services.
[Link] Security Responsibilities
 Three Key SecurityRequirements: Confidentiality, integrity, andavailability.
 Responsibility inCloudModels:
o SaaS: Provider handles all securityfunctions.
o PaaS: Provider manages dataintegrityandavailability; users handle confidentiality
and privacy.
o IaaS: Users manage almost all security, withthe provider ensuring availability.
1.5.3.3Copyright Protection
 Collusive piracyisa significant issue inP2P networks.
 Paidclients (colluders) illegallyshare copyrightedcontent withunpaid clients (pirates).
 Thispiracyaffects commercial content deliveryover open P2P networks.
 A proactive content poisoning scheme canprevent collusive piracy.
 The scheme uses identity-based signatures and timestamped tokens todetect pirates.
 Thisapproach prevents piracy without harming legitimate users.
[Link] System DefenseTechnologies
 First generation: Focused onpreventingintrusions withaccess controls, cryptographic
systems, etc.
 Second generation: Detects intrusions andtakes remedial actionsusingfirewalls, IDS,
PKI services, reputation systems, etc.
 Thirdgeneration: Provides more intelligent responses to intrusions.
[Link] Data Protection Infrastructure
 User level: Trust negotiation and reputation aggregationare needed.
BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D
 Applicationlevel: Security measures against worms, viruses, andDDoS attacks, plus
protection against piracyand copyright violations.
 Cloud security: Providers handle platform availability; IaaS users manage confidentiality;
IaaS providers manage data integrity. In PaaS/SaaS, bothparties share responsibility for
data integrity and confidentiality.
1.5.4 EnergyEfficiency in DistributedComputing
 Primarygoals: Highperformance and throughput withfault tolerance andsecurity.
 Emerging issues: Energy efficiency, workload/resource outsourcing, and sustainabilityof
large-scale systems.
 Energyconcerns: Energyconsumptioninservers and HPC systems (distributedpower
management).
 Costs: Systemslike EarthSimulatorand Petaflop have peak powercosts of $1,200 and
$10,000 per hour, respectively.
 Coolingissues: Hightemperatures negatively affect electronic components, increasing
operational costs andreducing component lifespan.
[Link] Energy Consumption ofUnusedServers
 Serverfarmcosts: Companies spend heavilyon hardware, software, operational
support, andenergy.
 Underutilization: 15% of servers (about 4.7millionglobally)are idling daily.
 Savings potential: Turning off unused servers could save $3.8billioninenergycosts and
$24.7billionintotal costs.
 Environmental impact: Wastedenergyresults in11.8milliontons of CO₂ annually,
equivalent to the pollutionof2.1 million cars.
 Action: IT departmentsshould identifyandmanage underutilized servers.
[Link] Reducing Energyin Active Servers
Inadditionto identifying unused/underutilized servers forenergysavings, it is also necessaryto
applyappropriate techniques to decrease energyconsumptioninactive distributedsystems
withnegligible influence on theirperformance. Powermanagement issues indistributed
computing platforms canbe categorized intofour layers (see Figure 1.26): the applicationlayer,
middleware layer, resource layer, and network layer.

BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D

1.5.4.3Application Layer
 Traditional apps: Focus onincreasing speedorquality.
 Energy-aware apps: Challenge isto manage energywithout compromisingperformance.
 Keyfactors: Energyconsumptiondependsonthe number of instructionsand memory
transactions.
 Correlation: Compute and storage impact completion time and energy use.
[Link] Middleware Layer
 Middleware layer: Bridges the applicationand resource layers.
 Functions: Provides resource broker, communication, task analysis, scheduling, security,
reliability, andinformationservices.
 Energy-efficiency: Focusesonenergy-efficient taskscheduling.
 New approach: Combines makespan (executiontime) and energyconsumptionin
scheduling.
[Link] Resource Layer
 Resource layer: Manages computingnodes, storage, and hardware resources in
distributed systems.
 Powermanagement: Includes Dynamic Power Management (DPM)and Dynamic
Voltage-FrequencyScaling (DVFS)techniques.
 DPM: Allows hardware, like CPUs, to switch to lowerpowermodes.

BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D
 DVFS: Controls powerconsumptionbyadjusting frequencyandvoltage, optimizing
executiontime andenergyuse.
1.5.4.6Network Layer
 Networklayer: Responsible forrouting packets and enabling network services to the
resource layerindistributed systems.
 Challenges:
1. Models shouldfullyunderstand the interactions of time, space, and energy.
2. Need fornew energy-efficient routingalgorithms andprotocols tocombat
network attacks.
 Data centers: Crucial forinformationstorage and processing but suffer fromhighcosts,
complexmanagement, lowsecurity, andexcessive energyconsumption. Next-gen
technologies are needed for improvement.
[Link] DVFS Methodfor Energy Efficiency

BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25
 CloudComputingandSecurity-BIS613D

BGSCET/AI&ML/2022-SCHEME/MODULE-01/2024-25