Cloud  Compu)ng  and  Virtualisa)on  

John  Brooke    ITS  and  IMG  group  in  CS  

March  31st  2014    

Thanks  to  Ralf  Ratering  for  material  on  slides  9-­‐18  

Structure  and  aims  

•  The  lecture  seeks  to  place  the  phenomenon  of  Cloud  Compu)ng  in  its  context.    

•  Cloud  Compu)ng  is  a  par)cular  form  of  Distributed  Compu)ng  which  is  crucially  dependent  on  the  concept  of  Virtualisa)on.  

•  We  first  discuss  previous  forms  of  virtualised  distributed  compu)ng,  in  par)cular  the  Grid  Programming  Environment  from  Intel.  

•  Then  we  consider  what    dis)nguishes  Cloud  Compu)ng  and  why  it  arouses  so  much  current  interest.  

CS6022 3

What is Computational Power?

•  Power in physics is defined as energy/time.

•  In computing we consider data-processing/time.

•  The processing needs of an organisation may vary, with spikes in consumption.

•  This leads to the concept of elastic computing, extra resource available as needed.

•  Ultimately this is scheduling problem, with a sufficiently large pool of resources, the illusion of elasticity can be provided.

The  power  of  scale  

•  Organisa)ons  such  as  Amazon,  Google,  MicrosoV  own  a  scale  of  resources  that  can  be  used  to  support  elas)c  compu)ng.    

•  However  we  must  introduce  uniformity  in  the  compu)ng  interfaces,  we  cannot  go  looking  for  libraries,  storage  spaces,  etc..  as  we  move  the  compu)ng  around  the  datacenter.    

•  It  is  necessary  to  run  virtualised  images  or  services  which  package  all  the  necessary  soVware  dependencies.    

CS6022 5

Costing computational resource

Cost vector

Job vector Cost

User Job

Scalar cost

in tokens

1

2

CS6022 6

Computational Resource

•  Computational jobs ask questions about the internal structure of the provider of computational power in a manner that an electrically powered device does not.

•  For example, do we require specific compilers, libraries, disk resource, visualization servers?

•  What if it goes wrong, do we get support? If we transfer data and methods of analysis over the Internet is it secure?

CS6022 7

RR space

RP space

RP space

RP space RR space

request

request

Request referral

sync

Figure 1: Request from RR space at A mapped into resource providers at B and C, with C forwarding a request formulated in RR space to RP space at D. B and C synchronize at end of workflow before results returned to the initiator A.

A B

C D

RR and RP Spaces

CS6022 8

WSRF-­‐  enabled  Servers  

WSRF-­‐  enabled  Storage  

WSRF-­‐  enabled  Network  

Web  Service  Resource  Framework  (WSRF)  

Open  Grid  Services  Architecture  (OGSA)  

Grid  Programming  Library  GPE

High-level Grid API

Operations n  Job management n  File transfers n  Brokering n  Steering, etc.

Descriptions n  Resources (CIM) n  Jobs (JSDL) n  Workflows (BPEL)

Grid  SDK   Grid  Beans   Client  Framework  

Applica)ons  

Grid  Compu)ng  

CS6022 9

• JSDL (Job Submission Description Language) –  High level job description that can be submitted to all target systems offering a JSDL interface

• CIM (Common Information Model) –  Used to describe resources –  Usage of CIM management interfaces for Grid administration

• BPEL (Business Process Execution Language) –  Integration of Grid Bean services into larger business process workflows

• WS* (WS-Addressing, WSRF, WSN, etc.) –  Interoperation with other Grid Middleware

• OGSA (Open Grid Services Architecture) –  Share components with other architectures

Standards

CS6022 10

Grid Services GPE Client Framework

Expert Client

Application Portal

Web Client

Portlets

Application Client

GridBean

GridBean Service

Grid SDK

Plugins Plugins GridBeans

Registry

Workflow Engine

WSR

F Se

rvic

es

UN

ICO

RE,

Glo

bus,

OG

SA-D

AI,

etc.

Load  GridBeans  

Submit  workflows  

Query  target  system  info  

Invoke  

Plugins Plugins GridBeans

Deploy  

GridBe

ans  

Register  

Invoke  

GPE Concept

CS6022 11

Atomic Services

GTK4  UNICORE  

GPE  Client  Framework  

Atomic  Service  Interfaces  

GTK4  Implementa)on  

UNICORE/GS  Implementa)on  

Common  set  of  WSRF  service  interfaces  

–  Atomic service interfaces define basic set of operations and properties that have to be available on a Grid

–  Different impementations of interfaces for different infrastructures

Java  TSI  

GPE  Implementa)on  

CS6022 12

Web Server

Client

Target System

File Export Service (FES)

File Import Service

(FIS)

File Management

Service (FMS)

Job Management

Service (JMS)

Target System Service (TSS)

Target System Interface

Invoke  

GPE Atomic Services

Invoke  

Manage  target  system  

Manage  jobs  on  target  system  

Manage  files  on  storage  

Manage  imports  to  storage  

Manage  exports  from  storage  

GPE  Java  TSI  

CS6022 13

Job3  

Operations Service Resource Type Properties

Target System Service (TSS)

Target    System  

Memory=1Gb Running Jobs=23 etc.

Extends  WS-­‐RP/LT  Job-EPR submit(JSDL, InitialTT)

Job Management Service (JMS)

Job2  

Status=Running OriginalJSDL=<JSDL..> ExecutionJSDL=<JSDL..> etc.

Extends  WS-­‐RP/LT  start() abort() hold() resume()

File Management Service (FMS)

Storage1  

Storage2  

AvailableFileSpace=127kb SupportedProtocols= {GridFTP, UPL, scp} etc.

Extends  WS-­‐RP/LT  listDirectory(path) createDirectory(path) changePermissions(path, perm) copy(path, path) rename(path, path) delete(path) import(path, protocol, isPipe) export(path, protocol, isPipe)

File Import Service (FIS)

File  Import  1  

File  Import  2  

Sourcefile=c:/tmp/test.txt DestFile=/tmp/test.txt Transferred=300kb etc.

Extends  WS-­‐RP/LT  putChunk(offset, size)

File Export Service (FES)

File  Export  1  

File  Export  2  

Sourcefile=/tmp/output.txt DestFile=c:/tmp/output.txt Transferred=12kb etc.

Extends  WS-­‐RP/LT  getChunk(offset, size)

GPE Atomic Service Interfaces

Job1  

create

create

TS-Property

CS6022 14

Site A

Discovering services: the role of a registry

–  Registry collects resource properties from target systems

–  Target systems notify Registry about property changes

•  WS-Notification –  Registry may be

installed •  per site •  across different

organizations

Client

Target System Service (TSS)

Broker

Registry

Target System Service (TSS)

Site B

Registry

Registry

Register/Update

Query

CS6022 15

GPE Registry

–  Registry may be installed on one site/system with the actual TSS, but also can be used across different organizations

–  TSS contacts Registry on startup to register and update the information using notifications

Registry

TSS

TSS

TSS DataBase  

TSS Static Properties

TSS installed Applications and application interfaces

TSS Dynamic Properties IDB  

IDB: Static Properties

IDB: Incarnations

Register/ Update

Query

CS6022 16

Web Server

GPE Workflows

Web Server

File Export Service

File Import Service

File Management

Service

Job Management

Service

Target System Service

Workflow Instance

Target System Service

Plain Web

Service

orchestrate

Client

Workflow Engine

EPR submit(BPEL)

create

getStatus()

–  Use BPEL to orchestrate WSRF services in complex workflows –  Allows integration into larger business processes

CS6022 17

Virtualization in Grid Computing

Virtual Target System

Virtual Target System

Physical Hardware Resources

VMM (Virtual Machine Manager)

Virtual  Machines  are  dynamically  created  and  configured  according  to  user  request  

• Protection of sensitive user data • Other partitions on the same machine will remain unaffected if one partition crashes • Virtual machines may even migrate during run-time

–  for instance when the hosting hardware is needed from a different user with high priority, when the system needs to shut down, etc.

CS6022 18

C L I E N T

OS Image Repository

Real Machine

Management Interface

1. send resource request

7. reference to deployed VM

query, delete OS images

8. query, submit, store, shutdown, pause, migrate, ...

4. create VM with OS image

2. request matching OS image

5. request missing software for OS image

request missing software from job request

6. Install software

3. deploy OS image

Virtual Target System Factory

Application Deployment

Service

Virtual Target System Service

manage VM

store OS image

Virtual Machine

9. submit job

OS

TSI

Virtual Machine

OS

TSI

TSI = Target System Interface

Virtualization Architecture in GPE

CS6022 19

Characteristics of Cloud Computing

As defined by the US NIST •  On-demand self service. •  Broad network access. •  Resource Pooling. •  Rapid elasticity. •  Measured service.

CS6022 20

Service oriented computing

NIST defines the following service models •  Software as a Service (SaaS) – example

Office 365 •  Platform as a Service (PaaS) – example

Google Apps. •  Infrastructure as a Service (IaaS) –

example Amazon Web Services.

CS6022 21

Private and public clouds

Clouds can be classified by accessibility •  Private Cloud, entirely owned by one

organisation – very like a private Grid •  Community Cloud – pooled between

organisations, again like a Grid. •  Public cloud widely available over the

Internet, payment models. •  Hybrid cloud, private but allows

“Cloudbursting”

A    model  for  portal  access  to  Clouds  

Virtual  Node  

Virtual  Node  

Virtual  Node  

Virtual  Node  

Virtual  Node  

Virtual  Node  

Virtual  Node  

Virtual  Node  

AutoPilot  Gateway  

AutoPilot  Gateway  

Accoun)ng  Database  

Cloud  Client  Library  

Virtual  Disks  

Pladorm  Image  

Pladorm  Image  

Cloud  

Portal  

control,  mount,  exec  

describe  

1:  request  VMs  by  pafern  

Service  Interface  

2:  access  VMs  

Auto  Pilot  Engine  

Pladorm  Descriptors  

Taverna  in  the  cloud  

•  Taverna  is  a  widely  used  system  for  running  workflows  that  process  scien)fic  data  mainly  from  the  life  sciences.    

•  Taverna  currently  runs  mainly  on  desktop  machines,  however  they  do  not  have  the  power  to  process  the  quan))es  of  data  coming  from  modern  genomics  sequencing  (called  Next  Genera)on  Sequencing)  

•  A  possible  solu)on  is  to  run  Taverna  in  the  Cloud  via  a  server  side  component,  Taverna  Server  

Server  Host  

Standard  View  of  Taverna  Server  

Portal  

Taverna  Server  Front  End  

TavServ  Back  End  

TavServ  Back  End  

TavServ  Back  End  

Service  

Service  

Service  

Server  Host  

Cloud  Back-­‐End  Services  

Portal  

Taverna  Server  Front  End  

TavServ  Back  End  

TavServ  Back  End  

TavServ  Back  End  

Service  

Service  

Service  

Simple  Taverna  Cloud  

Portal  

Taverna  Server  Front  End  

TavServ  Back  End  

TavServ  Back  End  

TavServ  Back  End  

Service  

Service  

Service  

Cloud  VM  Instance  

Discre)zed  Taverna  Cloud  

Portal  

Taverna  Server  Front  End  

TavServ  Back  End  

TavServ  Back  End  

TavServ  Back  End  

Service  

Service  

Service  

Cloud  VMs  

Cloud  Controller/Messaging  e.g.  Amazon  

Cloud  Compu)ng  and  Mobile  Compu)ng  

•  Cloud  compu)ng  and  mobile  compu)ng  can  be  very  complementary,  Cloud  is  the  server  side  and  mobile  is  the  the  agile  interface.    

•  Cloud  allows  data  and  applica)ons  to  be  accessed  from  mul)ple  devices  and  loca)ons  without  )me  consuming  and  error  prone  data  copying  e.g  Apple’s  IoS  links  to  the  Apple  Cloud.  

•  Cloud  provides  extra  power  to  mobile  devices,  e.g.  Siri.  

   

29  

Case study: supporting water engineers

Cloud Computing Infrastructure

Computa(onal  Model  

Real-World System Predic(ons  of  Future  

Events/Condi(ons  

Real-time Sensor Observations

On-demand Resource Allocation & Usage

Simulations Results

Wireless Sensor Network

•   Sensor-­‐Grid  Compu)ng  =  WSN  +  Cloud  Compu)ng  

•   Allows  construc)ng  real-­‐)me  models  

•   Applica)ons  includes  the  monitoring  and  control  of  natural  hazards,  built  environment  and  target/human  surveillance  

30  

Cloud Computing Infrastructure

Wireless Sensor Network

Computa(onal  Model  

Real-World System

Real-time Sensor Observations

On-demand Resource Allocation & Usage

Predictions of Future Events and Conditions

Proactive and Reactive Operational Decisions

A Distributed System of

Human Workforce

Current State

Integration of a distributed workforce with mobile devices

31

Managing physical infrastructure •  Focus of research work is on existing distributed Infrastructures:

§  cover longer distances

§  comprise of hundreds of thousands of elements

§  having complex topologies e.g. all-main models

§  operate for long period up to a century

•  Examples of Distributed Infrastructure:

§  Water Distribution System

§  Electricity Grids

§  Oil and Gas Network

•  UK water distribution and sewer pipes

§  Water pipes - 397,401 km (Earth - Moon: 384,403 km)

§  Sewer pipes - 354,066 km

Network of about 4000 nodes

Junc)ons  and  pipes  overlay  

Satellite  view  

34  

Design of the functionality of the system

We build on current Grid

middleware for computation

(heavyweight Grid/Cloud) and link it

to mobile and roaming devices with a role-based

lightweight architecture based

on messaging protocols. In this

way we include the engineering

process.

35  Wireless  Sensor  Network  

Data  Acquisi(on  &  Processing  Agent  

SPRING  Framework    

Real-­‐)me    raw  data  

Data    warehouse  

Filtered/Processed    Data  

Extended    Simula(on  Toolkit  

Conven(onal    Computa(onal    

Model  

Grid/Cloud    Compu(ng    Resources  

Asynchronous  Messaging    Queues  

Concurrency    Control    

Mechanism  

REST  based  API  

Sensors  Update  

Scheduling  of  Future  Decisions  

Alerts  

Queries/Results  

Dynamic    and    

Predic(ve  DST  

Dynamic    and    

Predic(ve  DST  

Alerts  

Queries,  Planned  Future  Ac)ons,  Alert  Subscrip)ons  

Results/Alerts  

GIS  Data  

Web  based  Clients  

36  

Interac)on  of  DST  with  RESTfull  Service  Client- RESTful Service Simulation Service

XML-GIS_Data

GET-GIS_DATA()

createExperiment()

PlanWhatIfDecisions()POST-runExperiment()

XML-Simulation_Result

Data Warehouse

GetGISData()

runSimulation()

extractWhat-IfDecisions()

runSimulation()

GetCurrentStatus()

addSimTS()simulationResult()

PlanFutureDecisions()POST-Register_Decision()

GetScheduledDecisions()

AcquireResourceLock()If SimTS(Si)>WTS(Ri) then write PFAix

if SimTS(Si)>WTS(Ri) then SUCCESS else ERROR

37  

Implementa)on  on  PDAs  and  SmartPhones  

•     

Reasons  for  current  success  of  Cloud  Model  

•  Easy  to  understand  business  model.  Payment  by  credit  card  or  by  subscrip)on.    

•  Organisa)ons  with  data  centres  that  are  on  a  scale  where  “elas)c  compu)ng”  and  economies  of  scale  become  possible.    

•  A  user  is  now  represented  by  mul)ple  devices,  several  of  them  mobile.  A  central  virtual  repository  is  needed  to  keep  the  user’s  data  consistent.  Dropbox  is  the  classic  example  of  this.    

Poten)al  problems  for  Cloud  Model  

•  Compared  to  Grid  Compu)ng  standards  in  the  Cloud  world  are  very  undeveloped.  Strong  risk  of  vendor  lock-­‐in.    

•  With  Cloud  not  only  your  soVware  but  also  your  data  and  even  your  iden)ty  become  locked  to  vendors.  Controversies  over  Facebook  show  the  the  very  high  stakes  here.    

•  For  very  high  performance  applica)ons,  virtualisa)on  is  a  performance  hit.  

•  Network  access  is  expensive  and  slow  rela)ve  to  all  other  features  of  Cloud  Compu)ng.  

More  informa)on  

•  Amazon  AWS  documenta)on  gives  a  good  descrip)on  of  how    IaaS  works.    

•  www.gridcafe.org  a  good  source  for  Grid  compu)ng  informa)on.    

•  hfp://en.wikipedia.org/wiki/Cloud_compu)ng  .  

•  There  are  lots  of  books,  I  have  used  Cloud  Compu)ng  Bible  by  Barry  Sosinsky  for  general  informa)on  and  Programming  Amazon  EC2  by  van  Vliet  and  Paganelli,  O’Reilly  books  for  hands-­‐on.