Flexible Scientific Workflows Using Dynamic Embedding

1. Flexible Scientific Workflows Using Dynamic Embedding Anne H.H. Ngu, Nicholas Haasch Terence Critchlow Shawn Bowers, Timothy McPhilips, Bertram Ludaescher

2. Outline • Scientific Workflow • Problems with static scientific workflow • Frame actor • Dynamic Embedding • Implementation of Dynamic Embedding • TSI case study • Conclusion

3. Scientific Workflows A model of the way a scientist works with their data and tools • Mentally coordinate data export, import, analysis and visualization via various software tools Goals: • Design • Automation • Component reuse … make data analysis and management tasks easier for the scientist!

4. SPA/Kepler Scientific Workflow System • Scientific Process Automation (SPA) • Modeling Workflows using actor-oriented framework • Executing and Monitoring Workflows • Built on top of Ptolemy II (Berkeley) • Graphical User Interface • Similar to a charting program (Visio) • Drag-and-Drop Components • Connect components • Execute workflows • Monitor Execution SPA PtolemyII

5. Actor-Oriented Modeling Actors • single component or task • well-defined interface (signature) • given input data, produces output data

6. Actor-Oriented Modeling Parameter • Input that is configured statically • User changes at design • Produces data

7. Actor-Oriented Modeling Sub-Workflows (aka Composite Actors) • composite actors “wrap” sub-workflows • like actors, have signatures (i/o ports of sub-workflow) • hierarchical workflows (arbitrary nesting levels) for abstraction • versus ‘Atomic Actors'

8. Actor-Oriented Modeling Directors • define the execution semanticsof workflow graphs • schedule and execute workflow graphs • sub-workflows may be governed by different directors • Examples: Synchronous Data-Flow (SDF), Process Networks (PN), Discrete Event (DE), Finite State Machine (FSM)

9. Problems in current SPA/Kepler modeling framework • Workflow is static and must be completely specified before orchestration • Specific tools used in the workflow must be picked at design time (It may not be picking the best tool) • All alternative tools must be enumerated exhaustively (resulted in complex workflow)

10. Problems in current SPA/Kepler modeling framework (cont.) • Our work aims to provide an abstraction of a tool, a resource, or an algorithm • Specific tool or resource resource is selected during execution

11. Terascale Supernova Initiative (TSI) Workflow

12. Different TSI Workflows A Each containing a Submit Job Actor B C

13. TSI Submit Job Actors…when we look at the sub-workflows Each workflow does job submission in a different way C A B Each containing a remote execution

14. Goals • There exists actors (atomic and composite) that perform a similar task • Submit Job • Transfer Files • Process Files • RemoteExection • SSH2Exec • SSHWrapper • InvokeSubmitJob • Can we provide an abstraction for encapsulating different implementations of a specific task that can be reused across different workflows? • Can we execute the workflow flexibly? • Choosing a specific implementation depending on runtime condition

15. Requirements • Select and execute actor based on run time conditions • Execute with data process networks • Built in capability for streaming data • Built in concurrency • Selected actors instantiated on demand • Reusable actor can be nested • Actors are usable by scientist

16. a a F a Frames • Actors are concrete • Correspond to particular implementations • Frames are abstract • Placeholder for actor / composite actor • input, output, and parameter ports • An embedding occurs when a actor is placed in a frame • A refinement is an actor that can be embedded in a frame SSH Actor a RemoteExecutionFrame F F Embedding F[a]

17. SSH2 Exec SSH2 Exec Web Service Web Service Static Embedding Frames • A refinement to a Frame is embedded during design • Frames become concrete and cannot be reconfigured during workflow execution Design Time Run Time F SSH2 Exec Web Service

18. F F SSH2 Exec Web Service Static Embedding Frames (cont.) • Execute with data process networks • Refinement instantiated as needed • Can be nested • Actors are usable by scientist • Select refinements based on run time conditions Run Time

19. SSH2 Exec SSH2 Exec Web Service Web Service Dynamic Embedding Frames • Refinement to frames are embedded during execution • Frames are not concrete during workflow execution Run Time F SSH2 Exec Web Service

20.  Why Dynamic Embedding? • Select refinements based on run time conditions • Execute with data process networks • Refinement instantiated as needed • Can be nested • Actors are usable by scientist

21. Implementation Of Dynamic Embedding Construct a new workflow to execute the actor Generates Remote Execution Frame Generated Workflow

22. Dynamic Embedding Process Remote Execution Frame • Wait for inputs to arrive to the Frame. • Select a refinement. • Transfer of input tokens from Frame to the refinement. • Select mappings • Input Port • Output Port • Parameter • Constructs internal workflow. • Run internal workflow. • Transfer of output tokens from the refinement to the Frame. Model Generated By Dynamic Frame

23. ModelReference Actor • A higher order actor that can execute a given model (workflow) through its input port. • It fits most of the requirement for dynamic embedding except: • User must pre-construct the given model • Output tokens from the given model are transferred only after completion of the internal workflow. • Our implementation of dynamic embedding leverage the capability of ModelReference actor with two major improvements: • The given model is constructed automatically • Output tokens are transferred synchronously • Frame is thus implemented as a subclass of ModelReference actor with four additional components: SelectActor(), FrameSourceActor(); FrameSinkActor() and PortWiring()

24. Implementing a Type of Frame • Subclass Frame and Implement • Selection Process • selectActor() • Configure Ports and Parameters • getIntputMappings() • getOutputMappings() • getParameterMappings()

25. Implements selection policy Returns a refinement Refinement that is returned gets automatically embedded Remote Execution String selectedActor selectActor(){ If(testWebService()) selectedActor = “webservice” return getWebServiceActor() else selectedActor = “ssh2exec” return getSSH2ExecActor() } Selection Process: selectActor()

26. Transfer Token Frame Input PortActor Input Port Actor Output PortFrame Output Port Expressed as list containing pairs of strings {“hostname”,”hostname”} {“command”,”cmd”} F hostname errors out hostname SSH2 Exec errors cmd command stdout Port Wiring Remote Execution String selectedActor getInputPortMapping(){ if(selectedActor==“SSH2Exec”) return {{“hostname”,”hostname”} {“command”,”cmd”}} else if(selectedActor==“webservice”) return {{“hostname”,”url”} {“command”,”method”}} }

27. TSI Case Study TSI-A Workflow SubmitJobFrame TSI-B Workflow

28. Remote Execution Frame TSI-A subworkflow TSI-B subworkflow SubmitJobFrame

29.  Benefits • Select refinements based on run time conditions • Execute with data process networks • Refinement instantiated as needed • Can be nested • Actors are usable by scientist

30. Limitations • Limitations • Unable to type check internal workflow before execution • There is overhead of creating an additional workflow to execute a refinement • Change in selection process requires recoding/recompiling • Can not monitor internal workflow (Useful for debugging)

31. Future Work • Semantic binding of Ports and Parameters • Configurable selection criteria • Intelligent brokering • Ptolemy Expression Language • Python • Perl • … • Simplified refinement creation • Caching of generated workflows • Design time type checking of internal workflow

32. UCRL-ABS-226047 Work performed under the auspices of the U. S. Department of Energy by Lawrence Livermore National Laboratory under Contract W-7405-Eng-48