QoS-based Scheduling of e-Research Application Workflows on Global Grids

1
QoS-based Scheduling of e-Research Application
Workflows on Global Grids

• Dr. Rajkumar Buyya

Grid Computing and Distributed Systems (GRIDS)
LaboratoryDept. of Computer Science and Software
EngineeringThe University of Melbourne,
Australiawww.gridbus.org
Gridbus Sponsors
2
GRIDS Lab _at_ Melbourne
Education
R D

• Youngest and one of the rapidly growing research
  labs in our School/University
• Founded in 2002
• Houses
• Research Fellows/PostDocs
• Research Programmers
• PhD candidates
• Honours/Masters students
• Funding
• National and International organizations
• Australian Research Council DEST
• Many industries (Sun, StorageTek, Microsoft, IBM,
  Microsoft)
• University-wide collaboration
• Faculties of Science, Engineering, and Medicine
• Many national and international collaborations.
• Academics
• Industries
• Software
• Widely in academic and industrial users.
• Publication

Community Services e.g., IEEE TC for Scalable
Computing
3
Agenda

• Introduction
• Utility Networks and Grid Computing
• Application Drivers and Various Types of Grid
  Services
• Global Grids and Challenges
• Security, resource management, pricing models,
• Service-Oriented Grid Architecture and Gridbus
• Market-based Management and Gridbus Software
  Stack
• Grid Workflows and QoS Scheduling
• Architecture, Design and Implementation
• Performance Evaluation Simulation based
  workflows
• SLA-based Resource Allocation
• Utility based allocation, pricing, performance
  results
• Summary and Conclusion

4
Power Grid Inspiration Seamlessly delivering
electricity as a utility to users
5
(5) Computing Grid Delivering IT services as the
5th utility after water, gas, electricity, and
telephone
eScience eBusiness eGovernment eHealth Multilingua
l eEducation
6
Grid-like Vision

• In 1969, Leonard Kleinrock, one of the chief
  scientists of the original ARPA project which
  seeded the Internet, wrote
• "As of now, computer networks are still in their
  infancy, but as they grow up and become
  sophisticated, we will probably see the spread of
  "computer utilities", which, like present
  electric and telephone utilities, will service
  individual homes and offices across the country
• Despite major advances in hardware and software
  systems over the past 35 years, we are yet to
  realize this vision. How far are we still from
  delivering computing as a utility?

7
Computing and Communication Technologies
Evolution 1960-2010!

HTC

P2P

PDAs
Minicomputers


PCs

Workstations

Mainframes

Grids
COMPUTING

PC Clusters
Computing as Utility

Crays

MPPs

WS Clusters

XEROX PARC worm
e-Science
e-Business
IETF
W3C
TCP/IP
Ethernet
Communication
Mosaic
HTML
Web Services
Email
Sputnik
SocialNet
Internet Era
WWW Era
XML
ARPANET
1960
1970
1975
1980
1985
1990
1995
2000
2010
Control
Centralised
Decentralised
8
What is Grid? (It means different things to
different people)

• IBM
• On Demand Computing
• Microsoft
• .NET
• Oracle
• 10g
• Sun
• N1 Sun Grid Engine
• HP
• Adaptive Enterprise
• Amazon
• Electric Cloud Services
• United Devices and related companies
• Harvesting Unused Desktop resources

9
What is Grid?Buyya et. al.

• A type of parallel and distributed system that
  enables the sharing, exchange, selection,
  aggregation of geographically distributed
  autonomous resources
• Computers PCs, workstations, clusters,
  supercomputers, laptops, notebooks, mobile
  devices, PDA, etc
• Software e.g., ASPs renting expensive special
  purpose applications on demand
• Catalogued data and databases e.g. transparent
  access to human genome database
• Special devices/instruments e.g., radio
  telescope SETI_at_Home searching for life in
  galaxy.
• People/collaborators.
• depending on their availability, capability,
  cost, and user QoS requirements.

Widearea
10
How does Grids look like?A Bird Eye View of a
Global Grid
Grid Information Service
Grid Resource Broker
Application
R2
R3
R4
R5
RN
Grid Resource Broker
R6
R1
Resource Broker
Grid Information Service
11
Classes of Grid Services / Types of Grids

• Computational Services CPU cycles
• Pooling computing power SETI_at_Home, TeraGrid,
  AusGrid, ChinaGrid, IndiaGrid, UK Grid,
• Data Services
• Collaborative data sharing generated by
  instruments, sensors, persons LHC Grid, Napster
• Application Services
• Access to remote software/libraries and license
  managementNetSolve
• Interaction Services
• eLearning, Virtual Tables, Group Communication
  (Access Grid), Gaming
• Knowledge Services
• The way knowledge is acquired, processed and
  manageddata mining.
• Utility Computing Services
• Towards a market-based Grid computing Leasing
  and delivering Grid services as ICT utilities.

Utility Grid
Users
Knowledge Grid
Interaction Grid
ASP Grid
Data Grid
infrastructure
Computational Grid
12
How Are Grids Used?
Utility computing
High-performance computing
Collaborative design
Financial modeling
High-energy physics
E-Business
Drug discovery
Life sciences
Data center automation
E-Science
Natural language processing Data Mining
Collaborative data-sharing
13
e-Science Environment Supporting Collaborative
Science

E-Scientist
Peers sharing ideas and collaborative

interpretation of data/results



Cyberinfrastructure
Distributed data
Remote
Visualization

2100

2100

2100

2100

Distributed computation

2100

21
00

2100

2100

Distributed instruments

Data Compute Service

14
Agenda

• Introduction
• Utility Networks and Grid Computing
• Application Drivers and Various Types of Grid
  Services
• Global Grids and Challenges
• Security, resource management, pricing models,
• Service-Oriented Grid Architecture and Gridbus
• Market-based Management and Gridbus Software
  Stack
• Grid Workflows and QoS Scheduling
• Architecture, Design and Implementation
• Performance Evaluation Simulation based
  workflows
• SLA-based Resource Allocation
• Utility based allocation, pricing, performance
  results
• Summary and Conclusion

15
Grid Challenges
16
Some Grid Initiatives Worldwide

• Australia
• Nimrod-G
• Gridbus
• DISCWorld
• GrangeNet.
• APACGrid
• ARC eResearch
• Brazil
• OurGrid, EasyGrid
• LNCC-Grid many others
• China
• ChinaGrid Education
• CNGrid - application
• Europe
• UK eScience
• EU Grids..
• and many more...
• India
• Garuda

• USA
• Globus
• GridSec
• AccessGrid
• TeraGrid
• Cyberinfrasture
• and many more...
• Industry Initiatives
• IBM On Demand Computing
• HP Adaptive Computing
• Sun N1
• Microsoft - .NET
• Oracle 10g
• Infosys Enterprise Grid
• Satyam Business Grid
• StorageTek Grid..
• and many more
• Public Forums
• Global Grid Forum

27 million
1.3 billion 3 yrs
2? billion
120million 5 yrs
450million 5 yrs
486million 5 yrs
1.3 billion (Rs)
1 billion 5 yrs
http//www.gridcomputing.com
17
Open-Source Grid Middleware Projects
18
Driving ThemeCommunity Grids vs. Utility Grids
Type Feature Community Grids Utility Grids
User QoS Best effort Contract/SLA
Service Pricing Not considered / free access Usage, QoS level, Market supply and demand
Example Middleware Globus, Condor, OMII, Unicore Nimrod-G, Gridbus, many inspired efforts
19
The Gridbus Project _at_ MelbourneEnable Leasing
of ICT Services on Demand
WWG
Gridbus
Pushes Grid computing into mainstream computing
20
(No Transcript)
21
Agenda

• Introduction
• Utility Networks and Grid Computing
• Application Drivers and Various Types of Grid
  Services
• Global Grids and Challenges
• Security, resource management, pricing models,
• Service-Oriented Grid Architecture and Gridbus
• Market-based Management and Gridbus Software
  Stack
• Grid Workflows and QoS Scheduling
• Architecture, Design and Implementation
• Performance Evaluation Simulation based
  workflows
• SLA-based Resource Allocation
• Utility based allocation, pricing, performance
  results
• Summary and Conclusion

22
What do Grid players want?

• Grid Consumers
• Execute jobs for solving varying problem size and
  complexity
• Benefit by utilizing distributed resources wisely
• Tradeoff timeframe and cost
• Strategy minimise expenses
• Grid Providers
• Contribute resources for executing consumer jobs
• Benefit by maximizing resource utilisation
• Tradeoff local requirements market opportunity
• Strategy maximise return on investment

23
What do Grid players require?

• They need tools and technologies that help them
  in value expression, value translation, and value
  enforcement.
• Grid Service Consumers (GSCs)
• How do I express QoS requirements ?
• How do I trade between timeframe cost ?
• How do I map jobs to resources to meet my QoS
  needs?
• How do I manage Grid dynamics and get my work
  done?
• Grid Service Providers (GSPs)
• How do I decide service pricing models ?
• How do I specify them ?
• How do I translate them into resource allocations
  ?
• How do I enforce them ?
• How do I advertise attract consumers ?
• How do I do accounting and handle payments?

24
Principle 1 Service Oriented Architecture (SOA)

• A SOA is a contractual architecture for offering
  and consuming software as services.
• There are four entities that make up an SOA
• service provider,
• service registry, and
• service consumer (also known as service
  requestor).
• The functions or tasks that the service provider
  offers, along with other functional and technical
  information required for consumption, are defined
  in
• the service definition or contract.

registry
contract
provider
consumer
25
Principle 2 Market-Oriented (Grid) Computing-
(a) Sustained Resourced Sharing and (b)
Effective Management of Shared Resources
Grid Economy
26
Market-based Systems Self-managed and
Self-regulated systems.

• Complexity present in Grid systems is similar to
  one present in human economies.

27
Service-Oriented Grid Architecture
Data Catalogue
Grid Bank
Information Service
Grid Market Services
Sign-on
HealthMonitor
Info ?
Grid Node N

Grid Explorer

Secure
ProgrammingEnvironments
Job Control Agent
Grid Node1
Applications
Schedule Advisor
QoS
Pricing Algorithms
Trade Server
Trading
Trade Manager
Accounting
Resource Reservation
Misc. services

Deployment Agent
JobExec
Resource Allocation
Storage
Grid Resource Broker

R1
R2
Rm
Core Middleware Services
Grid Service Consumer
Grid Service Providers
28
Gridbus and Complementary Technologies
realizing Utility Grid

Grid Applications
Portals
Science
Commerce
Engineering
Collaboratories

X-Parameter Sweep Lang.
Workflow
ExcellGrid
Gridscape
MPI
User-LevelMiddleware

Grid Brokers
Gridbus Data Broker
Workflow Engine
Nimrod-G
Grid Exchange Federation
Grid MarketDirectory
Globus
Unicore
Grid Storage Economy
GridBank

Core Grid Middleware
Alchemi
NorduGrid
XGrid
Grid Economy
.NET
JVM
Condor
SGE
Tomcat
PBS
Libra
Grid Fabric Software
Mac
AIX
Solaris
Windows
Linux
IRIX
OSF1
Grid Fabric Hardware
Worldwide Grid
29
Agenda

• Introduction
• Utility Networks and Grid Computing
• Application Drivers and Various Types of Grid
  Services
• Global Grids and Challenges
• Security, resource management, pricing models,
• Service-Oriented Grid Architecture and Gridbus
• Market-based Management and Gridbus Software
  Stack
• Grid Workflows and QoS Scheduling
• Architecture, Design and Implementation
• Performance Evaluation Simulation based
  workflows
• SLA-based Resource Allocation
• Utility based allocation, pricing, performance
  results
• Summary and Conclusion

30
Workflow-based Applications

• Workflow applications
• Scientific and engineering domains (e.g.,
  biology, astronomy, chemistry)
• Task execution is based on their control and data
  dependencies.

(Protein annotation workflow London e-Science
Centre)
31
Workflow for VR-Based Respiratory Treatment
Planning System
32
Driving ThemeCommunity Grids vs. Utility Grids
Type Feature Community Grids Utility Grids
User QoS Best effort Contract/SLA
Service Pricing Not considered / free access Usage, QoS level, Market supply and demand
Example Workflow Systems Triana, MyGrid, Askalon, DAGMan, Pegasus, GrADS Kepler Gridbus Grid Workflow Engine
33
Workflow Scheduling

• Scheduling on Community Grids
• Minimize the execution time based on best effort
  (ignores factors such as monetary cost of
  resource access and various users QoS
  satisfaction levels.)
• Scheduling on Utility Grids
• Focuses on mapping workflow tasks on services to
  satisfy users QoS constraints (e.g. deadline,
  the quality of produced data).
• Supports negotiation and establishment of SLA as
  a contract between users and providers
• Optimize performance under most important QoS
  constraints imposed by users.
• Minimize execution cost while meeting a specified
  deadline.
• Minimize execution time while meeting a specified
  budget.
• Support SLA-based allocation of resources so that
  multiple competing demands from users can be
  managed with the aim of enhancing providers
  profit.

34
Cost-based Workflow Scheduling

• Objective Function
• Minimize the execution cost and yet meet the
  time constraints imposed by users.

task

35
Workflow Management Systems

• Support composition, deployment, and execution
  management of workflow applications
• Workflow language
• Graphical environment for workflow composition
  and monitor
• Grid middleware integration
• Data management
• Fault-tolerance
• QoS-based SLA negotiation
• Scheduling
• ...

36
(No Transcript)
37
Architecture
GSP Grid Service Provider
Workflow Management System
SLA Service Level Agreement
Workflow Specification
Grid MarketDirectory
QoSRequest
Service Discovery
Grid Service
Performance Estimator
Grid Service
Workflow Planning
GSP
Grid Service
ReservationRequest(SLA)
Advance Reservation
contract violation
Grid Service
ServiceRequest(SLA)
Workflow Execution
QoS Monitor
Executor
Feedback
marketplace
Workflow Scheduling
38
Methodology

• Discover available services and estimate
  execution time for every task.
• Group workflow tasks into task partitions.
• Distribute users overall deadline into every
  task partition.
• Query available time slots, generate optimized
  schedule for each task partition and make advance
  reservations.
• Start workflow execution and reschedule when the
  initial schedule is violated at run-time.

39
Predicting Execution Time

• Reservation-enabled Utility Services
• Resource services
• Provide proportions of hardware resources (e.g.
  computing processors, network bandwidth, storage,
  memory) as a service for remote client access.
• Simulation, analytical modeling, empirical and
  historical data.
• Application services
• Allow remote clients to use their specialized
  applications.
• Provide estimated service times based on the
  metadata of users service requests.

40
Workflow Task Partitioning
Simple task
Branch
Synchronization task
T3
T2
T2
T3
T4
T4
T5
T6
T6
T5
T1
T1
T14
T14
T11

T11
T8
T7
T8
T7
T13
T13
T10
T10
T12
T12
T9
T9
Before partitioning.
After partitioning.
41
Deadline Assignment/Distribution
(43)

• P1. Any assigned sub-deadline must be greater
  than or equal to the minimum processing time of
  the corresponding task partition.
• P2. The overall deadline is divided over task
  partitions in proportion to their minimum
  processing time.
• P3. The cumulative sub-deadline of any
  independent path between two synchronization
  tasks must be same.
• P4. The cumulative sub-deadline of any path from
  entry task to exit task is equal to the overall
  deadline.

(152)
350
(217)
(284)
(350)
(53)
(120)
(187)
(187)
(187)
350
(253)
(269)
(269)
(350)
42
Planning

• Generates an optimized schedule for advanced
  reservation and run-time execution.
• Solve the problem based on divide-and-conquer.
• Generate a optimized schedule for each partition
  based on its assigned sub-deadline.
• A local optimized schedule minimizes execution
  cost while meeting its assigned sub-deadline.
• A optimized schedule constructed by local
  schedules.
• Task partition optimization
• Synchronization Task Scheduling
• Branch Task Scheduling

43
Task Partition Scheduling

• Synchronization task scheduling
• Only one task.
• Solution select the cheapest service that can
  process the task and transfer data within the
  assigned sub-deadline.
• Branch task scheduling
• One simple task in a branch.
• Multiple tasks in a branch.
• Model a branch as a Markov Decision Process (MDP)
  

44
Experiments

• Different Workflow Structures

(fMRIs neuroscience workflow)
(Protein annotation workflow London e-Science
Centre)
Pipeline
Parallel
Hybrid structure
45
(Simulation) Experiments

• MI (million instructions) represents length of
  tasks
• MIPS (Million Instructions per Second) represents
  the processing capability of services.
• Service type represents different types of
  services.
• 15 types of services, each supported by 10
  different service providers with different
  processing capability.

Table I. Service speed and corresponding price
for executing a task.
Table II. Transmission bandwidth and
corresponding price.
Service ID Processing Time (sec) Cost (G)
1 1200 300
2 600 600
3 400 900
4 300 1200
Bandwidth (Mbps) Cost/sec (G/sec)
100 1
200 2
512 5.12
1024 10.24
46
Experiments

• Compared heuristics
• Greedy cost
• sorts services by their prices.
• assigns as many tasks as possible to cheapest
  services without exceeding the deadline.
• Deadline-level
• divides workflow tasks into levels based on their
  depth in the workflow graph.
• assigns sub-deadlines to each task level equally.
  

47
Results
48
Results
49
Agenda

• Introduction
• Utility Networks and Grid Computing
• Application Drivers and Various Types of Grid
  Services
• Global Grids and Challenges
• Security, resource management, pricing models,
• Service-Oriented Grid Architecture and Gridbus
• Market-based Management and Gridbus Software
  Stack
• Grid Workflows and QoS Scheduling
• Architecture, Design and Implementation
• Performance Evaluation Simulation based
  workflows
• SLA-based Resource Allocation
• Utility based allocation, pricing, performance
  results
• Summary and Conclusion

50
Utility-driven Cluster RMS Architecture for GSPs
51
Economy-based Admission Control Resource
Allocation

• Uses the pricing function to compute cost for
  satisfying the QoS of a job as a means for
  admission control
• Regulate submission of workload into the cluster
  to prevent overloading
• Provide incentives
• Deadline --
• Execution Time --
• Cluster Workload --
• Cost acts as a mean of feedback for user to
  respond to

52
Impact of Penalty Function on Utility
53
Normalised Comparison of FCFS, Libra Libra
54
Impact of Increasing Dynamic Pricing Factor on
GSP Profitability
55
Agenda

• Introduction
• Utility Networks and Grid Computing
• Application Drivers and Various Types of Grid
  Services
• Global Grids and Challenges
• Security, resource management, pricing models,
• Service-Oriented Grid Architecture and Gridbus
• Market-based Management and Gridbus Software
  Stack
• Grid Workflows and QoS Scheduling
• Architecture, Design and Implementation
• Performance Evaluation Simulation based
  workflows
• SLA-based Resource Allocation
• Utility based allocation, pricing, performance
  results
• Summary and Conclusion

56
Summary and Conclusion

• Grids exploit synergies that result from
  cooperation of autonomous entities
• Resource sharing, dynamic provisioning, and
  aggregation at global level ?Great Science and
  Great Business!
• Grids have emerged as enabler for
  Cyberinfrastructure that powers e-Science and
  e-Business applications.
• SOA Market-based Grid Management Utility
  Grids
• Grids allow users to dynamically lease Grid
  services at runtime based on their quality, cost,
  availability, and users QoS requirements.
• Delivering ICT services as computing utilities.
• QoS Scheduling of Workflows and SLA-based
  resource allocation enables ability of Grids to
  serve as IT backbone for delivering utility
  computing services.

57
Thanks for your attention!
We Welcome Cooperation in Research and
Development! http/www.gridbus.org
eScience2007.org
58
Backup

• MDP etc.

59
Markov Decision Process (MDP)

• Effective for solving sequential decision
  problems.
• A MDP model contains
• A set of possible system states
• A set of possible actions
• A real valued reward (penalty) function
• A transition of each actions effects in each
  state

60
MDP Model

• States
• A state consists of current execution task, ready
  time and current location.
• Actions
• An action in the MDP is to allocate a time slot
  on a service to a task.
• t input data transmission time plus the
  processing time of the service.
• c transmission cost plus the service cost.

61
MDP Model

• Immediate penalty obtained from taking action a
  in state s and transitioning to state s.
• Expected penalty
• The sum of immediate penalties from current state
  to a terminal state.
• The optimal action for state s is

, sub-deadline

, otherwise
Expected penalty
62
Implementation

• Value iteration
• is a standard dynamic programming algorithm
• compute a new value function for each state based
  on the current value of its next state.
• value iteration proceeds in an iterative fashion
  and can converge to the optimal solution quickly.
• record a number of candidate solutions while
  finding the optimal time slot.

63
Rescheduling

• Re-adjust sub-deadline and re-compute optimal
  schedules for unexecuted task partitions.
• Reschedule minimum number of tasks.