Ben Butchart Wolfgang Emmerich University College London [email protected]
description
Transcript of Ben Butchart Wolfgang Emmerich University College London [email protected]
Orchestration of an OGSI-enabled scientific application using the Business Process Execution Language
Ben Butchart
Wolfgang Emmerich
University College London
Grids and Applied Language Theory (GAL2003)
Overview
• How we became interested in workflow• Why we chose BPEL• A BPEL appetizer• Shortcomings of existing BPEL implementations• Shortcomings of BPEL language itself• Overcoming some problems• Comparison with GSFL• Conclusion and future work• Discussion
The Beginning…
A messy but beautiful data flow diagram
Control logic of the Polymorph prediction application (not distributed)
configure system
loop 1..29 space groups
loop 1..X high denisty structurescollate results
molpak
dmarel
yes
is last structure ?
no
yes
no
is last spacegroup ?
Control logic of the distributed polymorph prediction application
molpak molpak molpak
dmarel dmarel dmarel dmarel dmarel dmarel
collate results
configure system
So why BPEL ?
• WSFL and XLANG deprecated• Microsoft, IBM, BEA triumvirate powerful
market force• Submitted to OASIS for standardization• First implementations already available for
evaluation (IBM BPWS4J, Collaxa)• Do not want to diverge from standard Web
Services technology set
Instantiating SOAP message
<soapenv:Body> <runPoly xmlns="http://poly"> <ns1:atomArray xmlns:ns1="http://www…/cml2/core"> <ns1:atom elementType="H" x="3.17" y="2.43“ z="0.02"/> <ns1:atom elementType="H" x="-1.74" y="2.61" z="-0.38"/> </ns1:atomArray> <spacegroups> <spacegroupLabel>AA</spacegroupLabel> <spacegroupLabel>AB</spacegroupLabel> </spacegroups> <ns2:crystal z="1" xmlns:ns2="http://www…/cml2/core"> <ns2:scalar title="alpha" units="deg">0</ns2:scalar> <ns2:symmetry id="" spaceGroup=""/> </ns2:crystal> </runPoly> </soapenv:Body>
BPEL Example
<sequence name="sequence"> <receive name="receive" partner=“user" portType="p:PolyPort" operation="runPoly" container="request“ createInstance="yes"/> <assign> <copy> <from container="request" part="atomArray"/> <to container="molpakrequest" part="atomArray"/> </copy> </assign>
<assign> <copy> <from container="request" part="crystal"/> <to container="molpakrequest" part="crystal"/> </copy> </assign>
BPEL Example
<assign> <copy> <from container="request" part="spacegroups“
query="./spacegroupLabel[position()=1]" /> <to container="molpakrequest" part="crystal"
query="./ns2:symmetry/@id" /> </copy></assign>
<invoke name="invoke1" partner="provider" portType="mpak:MolpakPort"
operation="runMolpak" inputContainer="molpakrequest"
outputContainer="molpakresponse"/>…
</sequence>
BPEL Features
• Loop and control logic• XPath query assignment• Recursive Composition• Synchronous and asynchronous
communication• Concurrent execution• Correlated messages• Event handling• Fault handling
Shortcomings of BPEL implementations
• Binding to a URL expected at deployment time
• Early implementations could not cope with OGSI extensions to WSDL
• Or properly handle imports• Dynamic creation of new process instances to
improve scalability• Steering – change process during execution
Shortcomings of BPEL language
• Need placeholder objects to support some assignments
• Data aggregation - No way to temporarily collate results
• Cannot specify mechanism for binding (WSDL, UDDI, OGSAFactory)
• Limited lifecycle interface (just receive, pick terminate)
• Verbose• Does not support peer communication
The scalability problem
After Krishnam, Wagstrom, LaszewskiGSFL: A Workflow Framework for Grid Services
Workflow Engine Workflow Engine
Compute Service
Compute Service
Compute Service
Compute Service
Compute Service
Compute Service
<invoke name="invoke1" partner="provider" portType="mpak:MolpakPort" operation="runMolpak" inputContainer="molpakrequest" outputContainer="molpakresponse"/>
getBindingURL(PortType)
BPEL Parser
AbstractBinding
WSDLBinding
UDDIBinding
OGSA Registry
OGSA FactoryBinding
OGSA Registry
UDDI Registry
Molpak serviceCreate instance
<service name="MolpakService"><port><soap:address location= "http://tapas.cs.ucl.ac.uk: 8080/ogsa/services/"/></port></service>
Late binding
Scalability
Workflow Engine
Compute Service
Compute Service
Compute Service
Workflow Engine
Compute Service
Compute Service &Workflow Engine
Compute Service
compute1
compute2
compute3
How we solve the problems
• late binding: extra URL binding mechanisms• scalability: recursive composition• scalability: group binaries into single WS port type • scalability: dynamic deployment and delegation• steering: use BPEL event handling functionality• aggregation: use a dedicated web service• verbose: use GUI/XSLT
Other Web Service Workflow Languages
• WSFL and XLANG – now superseded by BPEL
• GSFL – extension to WSFL to support Grid requirements
• SWFL – another extension to WSFL • Scufl (taverna) / Freefluo
Comparison with GSFL
• recursive composition
• correlated messages
• separates data flow
• peer data flow modelling
• binding mechanism configurable
• WS standard
• cross vendor support
• fault handling and compensation
• concurrent control flow
• message part assignment
• Xpath assignment
• implementations available
• recursive composition
• correlated messages
• separates data flow
• peer data flow modelling
• binding mechanism configurable
• WS standard
• cross vendor support
• fault handling and compensation
• concurrent control flow
• message part assignment
• Xpath assignment
• implementations available
GSFL BPEL
Conclusion
• Can use BPEL for modelling grid applications – no real need for extensions
• But we need an implementation that is Grid enabled
• Might not need full implementation • We should stick with WS technology set if at
all possible
Integration with Scheduler (Sun Grid Engine)
2.getBindingURL(PortType)Trader
Binding
SGE
3. Run a proxy job (jobID, traderURL)
molpak serviceisntance
OGSARegistry
Trader
SGE Wrapper
4. Run a proxy job (jobID, traderURL)
Dummy Process(blocks)
5. runDummy.sh(jobID,traderURL)
BPEL Engine
1.getBindingURL(PortType)
9.invocation completed
6. SGEchooseMe(jobID,myURL)
10.kill
7. search for instance
8.invoke service
compute node
Late binding: deployment descriptor
<service name="poly/PolyProviderFactoryService/poly1"> <requestFlow> <handler type="BPELRequestHandler"/> </requestFlow> <responseFlow> <handler type="BPELResponseHandler"/> </responseFlow> <parameter name="name" value="Polymorph process workflow"/> <parameter name="className" value="poly.PolyProvider“ <parameter name=“bindingMechanism" value=“Registry"/></service>