Upload
clarence-banks
View
214
Download
0
Tags:
Embed Size (px)
Citation preview
A Novel Approach to Workflow Management in Grid EnvironmentsFrank Berretz*, Sascha Skorupa*,Volker Sander*, Adam Belloum**
15/04/2010
* FH Aachen - University of Applied Science, GER ** University of Amsterdam, NL
© FH AACHEN UNIVERSITY OF APPLIED SCIENCES 15. April 2010 | 2
Taxonomy of Grid Workflow Systems
> Common Architecture
Push-based Job Distribution
> Requirements
> UNICORE Workflow System
> Consequences
Pull-based Approach
> Benefits & Challenges
> General Architecture
Prototype Implementation
> UNICORE Grid Middleware
> jBPM Workflow Engine
Outline
© FH AACHEN UNIVERSITY OF APPLIED SCIENCES 15. April 2010 | 3
Taxonomy of Grid Workflow SystemsCommon Architecture
Workflow Design & Definition
Grid Users
Build Time
Run Time
Workflow Execution & Control
Interaction with Grid resources
Grid Workflow Application Modeling & Definiton Tools
Workflow Scheduling
Data Movement Fault Movement
Grid Workflow Enactment Service
Grid Middleware
Grid Resources
Grid Information Services
Resource Info Service
Application Info Service
interaction withinforamtion services
workflow change
Grid Workflow Specification
Source: Jia Yu and Rajkumar Buyya
© FH AACHEN UNIVERSITY OF APPLIED SCIENCES 15. April 2010 | 4
Push-based job distribution requires
> Efficient resource discovery and selection processes
> Detailed knowledge of available resources
> Well-defined interfaces of resources
> Up-to-date and confidential information systems
> Adapt VO schedulers and local schedulers
> Proper Access Control Lists from resource providers
Push-based Job DistributionRequirements
© FH AACHEN UNIVERSITY OF APPLIED SCIENCES 15. April 2010 | 5
Push-based Job DistributionUNICORE Workflow System
1. Splitting workflow into sequence of WAs and send them to a orchestrator.
2. Filter appropriate resources by requesting an information service.
3. Information Service requests all available Grid sites.
4. Evaluate requests from multiple VOs.
5. Response concrete resource endpoint.
6. Service Orch. forwards JSDL to known interface.
7. Site performs authorization by mapping CA to a local account.
8. XNJS sends job through the TSI to a physical computing resource.
9. – 11. Callback chain
Workflow Engine
Service Orchestrator
Target System Interface
UNICORE Atomic
Services
OGSA Interface
XNJS
Workflow Management System
Computing Resource
WA
JSDL
Information Service(CIS / GLUE 2.0)
XUUDB
12
3
4
5
6
7
8 9
10
11
© FH AACHEN UNIVERSITY OF APPLIED SCIENCES 15. April 2010 | 6
Scalability: Information systems and schedulers may become bottlenecks with respect to the amount of…
> Users and resources
> Parallel branches inside a workflow (parameter studies)
Cross Grid Scheduling:
> Side effects caused by resources in multiple Grids
> Limitation of resource candidates to resources of a particular Grid (Open community approach?)
Heterogeneity: Grid Workflow Systems typically deal with computational resources
> Cumbersome integration of special resources like human interaction (lack of integrating emerging standards)
> Complex decision processes by humans might influence further workflow steps (e.g. qualitative assessment criteria)
Push-based Job DistributionConsequences
© FH AACHEN UNIVERSITY OF APPLIED SCIENCES 15. April 2010 | 7
Alternative IdeaPull-based Job Distribution Strategy
Workflow system sends task to an intermediary repository
Resources act autonomously and adapt to the repository
Any kind of resource can actively request the repository
Resources apply for defined roles to receive tasks according to their capabilities
Resources have to authenticate against the task repository
Grid Workflow Application Modeling & Definition Tools
Grid WorkflowSpecification
Grid Middleware
Grid Resources
Grid WorkflowEnactment Service
Workflow Scheduling
DataMovement
FaultManagement
Grid WorkflowEnactment Service
DataMovement
FaultManagement
Task Repository
Community Task Management
workflowchange
workflowchange
WFMS push jobs to Grid resources Grid resources pull
Jobs from WFMS1
2
12
Task Client
GridResources
HumanResources
© FH AACHEN UNIVERSITY OF APPLIED SCIENCES 15. April 2010 | 8
Architectural Concept for the Pull Model
Workflow Repository
Abstract Workflow
Abstract Workflow
Abstract Workflow
finished workflow
instantiate
Workflow Engine
Concrete Workflow
finished task
deploy task
Task Repository
Stored Tasks
Gat
eway
Rich Client(WF-Deployer)
Rich Client(WF-Executor)
Actor
Actor
Actor
Actor
deploy workflow
lookup workflows
execute workflow
notify
lookup tasks
claim tasks
execute tasks
finished tasks
© FH AACHEN UNIVERSITY OF APPLIED SCIENCES 15. April 2010 | 9
Benefits
> Scheduler and brokering components are now optional
> Simplified integration of special resources like humans, telescopes or medical devices
> Reduced administrative VO management overhead at resource sites
> Support actors across organizational boundaries (community approaches)
Challenges
> Bottleneck problem should not be shifted to the task repository
> Submitted jobs run the risk of starvation (SLAs!)
> Appropriate security and provenance frameworks needed
Benefits & Challenges of Pull-based Approach
© FH AACHEN UNIVERSITY OF APPLIED SCIENCES 15. April 2010 | 10
Integrate Pull Approach to an existing Grid Middleware
Extending UNICORE Grid middleware
Use existing XML Tuple Space as repository
Integrate jBPM Workflow Engine as client for the space
1.User starts workflow2.Engine writes job to space3.Resource takes job from space4.Resource executes job locally5.Resource finishes job6.Engine receives notification by the space and resumes workflow
Grid User(s)UNICORE Rich Client
jBPM Engine
Workflow Management System
WA
1
11
Resource Provider(s)Job Taker
XML Space(s)
2
3 5
6
4
UN
ICO
RE
Host
ing
Env.
© FH AACHEN UNIVERSITY OF APPLIED SCIENCES 15. April 2010 | 11
Conclusions:
> Pull-based job distribution strategies are currently missing in Grid systems
> But it could be used as an alternative model for certain application scenarios
(heterogeneous resources, high-throughput computing, …)
Summary and Outlook
© FH AACHEN UNIVERSITY OF APPLIED SCIENCES 15. April 2010 | 12
Further steps:
> XML space should be replaced by a scalable task repository
> Among the simple Job-Takers more complex client systems should be implemented to integrate special resources (humans) into Grids
> Hybrid push/pull distribution strategies as an option
> Performance and scalability analysis
This work should result in a refined architecture to address the challenges of pull-based approaches
Summary and Outlook
FH Aachen University of Applied ScienceCampus Jülich
Ginsterweg 1 52428 JülichT +49. 241. 6009 [email protected], [email protected] www.fh-aachen.de
© FH AACHEN UNIVERSITY OF APPLIED SCIENCES | FRANK BERRETZ, SASCHA SKORUPA | GINSTERWEG 1 | 52428 JÜLICH | WWW.FH-AACHEN.DE