Administration Arranged Registering Administration Disclosure and Creation - PowerPoint PPT Presentation

service oriented computing service discovery and composition l.
Skip this Video
Loading SlideShow in 5 Seconds..
Administration Arranged Registering Administration Disclosure and Creation PowerPoint Presentation
Administration Arranged Registering Administration Disclosure and Creation

play fullscreen
1 / 48
Download Presentation
portia-roman
Views
Download Presentation

Administration Arranged Registering Administration Disclosure and Creation

Presentation Transcript

  1. Service-Oriented ComputingService Discovery and Composition

  2. Service Discovery • UDDI • Discovery based on WSDL information • WS-Discovery • Provide an interface for service discovery • Define a multicast discovery protocol • Limitations: No service liveness information, limited service description • Universal Plug and Play (UPnP) • Enables dynamic networking of intelligent appliances, wireless devices, and PCs • Not exactly web services, but is another form of standardization for physical systems

  3. WS-Discovery • Extend UDDI to make it distributed • WS-Discovery multicast message types: • Hello: Sent by a Target Service when it joins a network • Bye: Sent by a Target Service when it leaves a network • Probe: Sent by a Client searching for a Target Service • Search by Type and/or Scope • Resolve: Sent by a Client searching for a Target Service by name • Already know the target service by name, but may not know the communication details • Response uni-cast message types: • Probe Match: a Target Service matches a “Probe” • Resolve Match: a Target Service matches a “Resolve”

  4. UPnP • Devices • When entering a network, register to the control point • When leaving the network, information the control point • Control point • When entering the network, search for devices

  5. UPnP • Device specification • Device name, vender name, URL for the device, etc. • Commands, actions, parameters, etc. • Current state information • Interactions • Control: Control point sends commands/actions (in XML) to activate the device • Event: • Control point can request device to send updates when some variables (specified in the event) are updated • Device can accept the request and respond with a event duration • Presentation • If the device has an URL, control point can request and fetch it • Some devices can be controlled through the URL interface

  6. UPnP • Compare to UDDI • Similar, but has an additional interaction feature • Lack of semantics, hard to compose the devices together to achieve a client goal • Can we wrap devices into high-level services and use the OWL-S technologies to add semantics to devices?

  7. Semantic Web Service Discovery • Semantic search template • User’s service requirements • Semantic web service descriptions • A similarity based matching scheme (just an example) • The Search Template is matched against a set of candidate Web services in a registry and the match Scores are calculated • Overall similarity = Weighted average of syntactic similarity and Functional similarity (normalized sum of operational similarity) • Syntactic similarities = Weighted average of the Name similarity and Description similarity • Operation similarity = Weighted average of syntactic similarity, conceptual similarity, and I/O similarity • … (further decompose the similarity definitions)

  8. Semantic Web Service Discovery • Example search template

  9. Semantic Web Service Discovery • Example candidate service

  10. Semantic Web Service Discovery • Similarity definitions Best match! Compare a desired service name with the names of several candidate services Best match! Compare a desired operation with all the operations of a candidate service

  11. Service Composition • How to put services together to achieve the desired goal • Same old problem • Design patterns has shown to be effective in software design • Many industrial efforts on SOA design patterns • What is a pattern • "A solution to a problem in a context"? • Each pattern describes a problem which occurs over and over again ... and then describes the core of the solution to that problem, in such a way that you can use this solution over and over again • Research • Pattern based research considers how to specify patterns, i.e., how to specify the problem, solution, effects • Semantic Web service composition community consider AI planning techniques for composition reasoning

  12. AI Planning for Service Composition • Planning • move(x,y) • Pre-condition: clear(x) and clear(y) • Effect: on(x,y) and clear(x) • Delete effects: on(x,?), clear(y) • Always • clear(table) • Does not conflict with on(x, table) Initial state on(c,table) on(b,table) on(a,b) clear(a), clear(c) Goal state on(a,table) on(c,a) on(b,c) clear(b) Composition move(a,table) move(c,a) move(b,c) B A C B C A

  13. on(c,table) on(b,table) on(a,b) clear(a) clear(c) clear(table) on(a,c) move(a,c) move(a,table) on(c,table) on(b,table) on(a,b) clear(a) clear(c) clear(table) on(a,table) . . . move(c,a) AI Planning for Service Composition on(c,table) on(b,table) on(a,b) clear(a) clear(c) clear(table) . . . Planning is different from other search algorithms - which generally based on a quantitative measure Planner involves state computation and maintenance

  14. AI Planning for Service Composition • Map semantic web to the planning domain • Definitions for web services • Syntactical definition: I/O parameters • Semantic definition: pre-condition and effects • Supported in OWL-S and WSMO • Map services to actions in planning domain • Pre-condition/effects of the services become the pre-condition/effects of the actions • I/O definitions are translated t o the pre-condition/effects • Map the problem to the planning domain • Define the goal for the problem • Define the initial facts

  15. AI Planning for Service Composition • Limitations of traditional planners for service composition • Atomic actions with deterministic effects, only able to generate sequential plans • Conditional, iterative planning: Construct a plan with branches, taking all possible nondeterministic effects and contingencies into account • Complete knowledge of the world, full observability • Conformant planning: Find a plan which works in any initial situation or incomplete knowledge • Contingency planning: Consider all possible nondeterministic effects or replan when unexpected situation occurs

  16. AI Planning for Service Composition • General limitations for service composition • Pre-conditions and effects, initial states, and goals are mostly simple conjunctions of propositions • Can real-world web services be easily specified based on these? • Has been a core problem in SE for 20+ years!!! Will it work now? • Scaling issues • There may be thousands of services each with multiple ports • Even worse in cyber-physical systems • A lot devices with similar functionalities • Solutions • Hierarchical search: First use keyword based search to filter out unlikely actions, then use planner to explore the possible actions • Service ontology: Categorize services and specify service relations using an ontology

  17. AI Planning for Service Composition • General limitations for service composition • Assume a static and finite set of actions • In SE, a problem can be decomposed and then find the corresponding components • Research work trying to handle partial planning with missing actions • Interaction with users • Planer better be more mixed-initiative • knowledge engineering issues • Efficiently and effectively interacting with XML based information • This should be the simplest problem among the many issues

  18. QoS in Service Composition • QoS (quality of service) • Nonfunctional properties to be satisfied • E.g., availability, performance, price, reliability • Service composition with QoS considerations • Find the best service to meet user QoS requirements • Need to specify client QoS requirements • First, need to define what QoS is (the properties of concern) • Need to know the QoS properties of the services • For a single service, QoS properties can be measured • For a composite service, how to derive the properties of the composed service? • For some properties, property aggregation can be very difficult • Also, need to understand the interaction behaviors among services • Decision making: which services to select?

  19. QoS in Service Composition • Need to have • A formal process to do this systematically, from QoS specification to negotiation, to finalize the selection • An agreement between the involved entities to ensure that the negotiated QoS terms are exercised  Service Level Agreement (SLA) • WS-Agreement • Provides the specification standards for SLA between the client and the service providers • Dynamically established and dynamically managed QoS • Use QoS ontology for QoS specifications, negotiation, and management

  20. Factory Negotiation create() Ops: terminate(limits) negotiate(...) ... SDEs: negotiate() Terms Status Related Agrmts. Negotiator WS-Agreement • Negotiation Layer • Agreement Layer • Service Layer • Factory: creates the instance Factory Agreement create() Ops: terminate(limits) inspect(query) ... SDEs: inspect() Terms Status Related Agrmts. Manager Factory create() Policy Application Instance foo() Consumer Provider

  21. WS-Agreement • Negotiation layer • Provides a Web service-based generic negotiation capability • Newly added (original only has two layers) • Negotiation state transition:

  22. Name Context WS-Agreement • Context • Agreement initiator, responder, expire time, etc. • Service Terms • Identify the specific services to be provided • Guarantee Terms • The service levels that the parties are agreeing upon • Can be used for monitoring Agreement Terms Service Terms Guarantee Terms

  23. WS-Agreement hasGuaranteeTerm GuaranteeTerm hasBusinessValue A guarantee term has a scope – e.g. operation of service hasScope hasObjective hasCondition Scope BusinessValue ServiceLevelObjectives Qualifying Condition hasReward A guarantee term may have collection of service level objectives e.g. responseTime < 2 seconds There might be business values associated with each guarantee terms. Business values include importance, confidence, penalty, and reward. e.g. Penalty 5 USD Reward A guarantee term may have a qualifying condition for SLO’s to hold. e.g. numRequests < 100 Predicate Predicate hasPenalty hasImportance Penalty Parameter Parameter Unit Importance Value ValueExpression Unit Value ValueUnit ValueExpression OWL ontology Assessment Interval Assessment Interval ValueUnit TimeInterval Count Count TimeInterval

  24. WS-Agreement – Agreement Schema WS-Agreement lacks formalism for requirement specification This can make the negotiation and selection process difficult. SWAPS provides these specifics.

  25. Agreement Reasoning • Semantic WS-Agreement Partner Selection (SWAPS) • WS-Agreement • Temporal Concepts: time.owl • OWL version of time (http://www.isi.edu/~pan/damltime/time.owl) • Example concepts: seconds, dayOfWeek, ends • QoS ontology • E.g., Ont-Qos (IBM) • E.g., QoS ontology in Metero-S • Example concepts: responseTime, failurePerDay • Domain Ontology • Represent the domain knowledge • Semantics of predicates rules, such as <, =, etc. • User defined rules • Allow users to customize matchmaking declaratively

  26. Example QoS Ontology

  27. Example Domain Rules • Consumer: • Requirement: Availability is greater than 95% • Provider: • Mean time to recover (MTTR) = 5 minutes • Mean time between failures (MTBF) = 15 hours • Domain rule: • Availability = MTBF / (MTBF + MTTR) • Reasoning • Availability of the provider = 99.4%.

  28. Agreement Reasoning • An alternative alt1 is a suitable match for alt2 if: • (("Gi) such that Gi alt1  requirement(alt1, Gi)  ($Gj) such that Gj alt2  capability(alt2, Gj)  scope(Gi) = scope(Gj)  obligation(Gi) = obligation(Gj)  satisfies(Gj, Gi) • G: a term in the agreement • requirement(alt, G): True if G is a requirement of alt • capability(alt, G): True if G is a capability of alt • scope(G): The scope (the service operation) of G • obligation(G): The obligated party of G • satisfies(Gj, Gi): True if the SLO of Gj is equivalent to or stronger than the SLO of Gi

  29. Agreement Reasoning isEquivalent Provider responseTime < 14 s QC: day of week = weekday Penalty: 15 USD Provider 99% of responseTimes < 14 s Consumer Provider1 Provider FailurePerWeek < 7 Penalty 10USD Provider failurePerWeek < 10 isStronger Provider transmitTime < 4s QC: maxNumUsers < 1000 Penalty: 1 USD Provider processTime < 5 s QC: numRequests < 500 Penalty: 1 USD Provider2 Provider failurePerWeek < 7 Penalty: 2USD

  30. Agreement Reasoning Provider responseTime < 14 s QC: day of week = weekday Penalty: 15 USD isStronger Provider 99% of responseTimes < 14 s Consumer Provider1 Provider FailurePerWeek < 7 Penalty 10USD Provider failurePerWeek < 10 Provider responseTime < 9s QC: maxNumUsers < 1000 QC: numRequests < 500 Penalty: 1 USD Domain Specific Rules: responseTime = processTime + transmitTime  Provider2 Provider failurePerWeek < 7 Penalty: 2USD isStronger

  31. Agreement Reasoning • Problem • There may be no matching providers • Specification can be optimization based • E.g., minimize responseTime  Choose the provider that yields best fit • How to combine different terms • Some non-exact-matching terms, each with a satisfaction level • User define a weighted function to combine them • Multi-objective optimization • When there are a large number of choices • Use some filtering mechanisms first • E.g., consider one term first

  32. Agreement Reasoning • Problem • Only consider selection of a single service • Requirements are frequently specified as end-to-end requirements • E.g., the response time of the composite service is < 1 sec • But the response times are specified only for the atomic services • How to aggregate the properties • For some QoS aspects, this can be very difficult • When there are a large number of choices • Each customer required service may have many providers • Combination can be extensive  Use efficient search algorithms to find the best match • Linear programming, genetic algorithms, etc. have been used

  33. QoS Property Aggregation • Time • Generally is the simplest property to handle • Sum for sequential composition, Max for parallel composition, Min for choices • Is this accurate? • Some works consider multi-tier queuing network for time estimation • How about communication cost • Keeping a table of pair-wise communication cost? O(N2) space • Where to maintain the cost? • Infeasible in case each service may be able to compose with many other services • Cost • Sum for sequential/parallel composition, Min for choices

  34. QoS Property Aggregation • Power consumption • Sum for sequential composition, Min for choices • Is this accurate? How about cumulative heat, etc. • Complicated aggregation for parallel composition • Availability • MTTF / (MTTF+MTTR) • Mean time to failure (MTTF) • Mean time to repair (MTTR) • How to aggregate the availability of the software, hardware, and communication? • Large versus small MTTF & MTTR • Pacemaker: MTTF = 1years, MTTR = 5min, A = 0.99999 • Pacemaker: MTTF = 3 hours, MTTR = .01 sec, A = 0.99999 • Need to consider request frequency and failure consequences

  35. QoS Property Aggregation • Reliability • Software reliability • A: reliability = 1, B: reliability = 0.99 • A+B: reliability = ? • A always invokes B’s incorrect path • A never invokes B’s incorrect path • Need to consider operational profile for software reliability estimation • Determine the probability of each type of input • Mostly partitioned based on a group of program paths • Determine the reliability according to the probability of failure for each input type (testing results) and probability of the input • So, aggregation need to consider mapping of the operational profiles • From A’s operational profile to estimate B’s operational profile • System reliability • Combining software and hardware reliability, how?

  36. QoS Property Aggregation • Quality • Difficult to measure quality even for a single service • Use fuzzy logic to express quality • How to compose them? • Security • Focus on policy based composition, not quantitative measures • No established method to quantitatively measure security yet • Too many factors • Policy issues, attacks, etc. • Generally only consider the policies, but not the attacks • Even for policies, mostly focus on the policies for individual services, not the information flow in the composite service • Attacks may compromise the system, making policies useless

  37. QoS Property Aggregation • Simulation approach • Obtain the aggregate QoS properties based on simulation • Time: Rapidly compose the system and obtain the cumulative time • Reliability: Run test cases with the actual composite service • If there are many different combinations, simulation of each case is infeasible

  38. Search for a Composition Solution • Genetic algorithm • Selection • E.g., Roulette-wheel selector • Probability of selecting Si = E(Si) / k E(Sk) • Mutation, Crossover

  39. Search for a Composition Solution • Multi-object problem and Pareto-optimal solutions • Separate considerations of the objectives • Comparison problem (which is better) • Dominate solution (one is definitely better than the other) • Pareto-optimal solutions: non-dominated solutions

  40. Search for a Composition Solution • Problem • Services may be QoS reconfigurable • E.g., a security service may offer many different security configurations, providing security and time tradeoffs • E.g., an AI search algorithm may offer different thresholds, providing quality and time tradeoffs • Need to consider not only which provider, but also the specific configurations of the provider services • Greatly increase the search space • Consider the paper • QoS-Reconfigurable Web Services and Compositions for High-Assurance Systems

  41. Search for a Composition Solution • Example: Media security gateway for VoIP

  42. Search for a Composition Solution • Example: • Reconfigurable SRTP service: • Implement secure and real-time audio packets transmission • Configurable parameters: Different encryption key sizes, encryption algorithms, authentication key sizes, and authentication algorithms • Reconfigurable AMR codec service • Implement the AMR codec standards • Configurable parameters: Different encoding (compression) rates • Tradeoffs between voice quality, commu/exec times • Reconfigurable Packetization service • Put audio frames into packets and add error correction and retransmission control codes • Configurable parameters: Different error correction codes, retransmission windows, etc. • Tradeoffs between the packet loss rate (which impacts voice quality) and the commu/exec times

  43. Search for a Composition Solution • Compositional approach • First find the Pareto-optimal solutions within the services • Compose the Pareto-optimal solutions • Compose Pareto-optimal solutions as the initial population • If PO-solutions of the components do not guarantee the PO-solutions of the composed system • Only compose Pareto-optimal solutions as the final solutions • Much more efficient due to the greatly reduced search space

  44. Grounding Selections Service Domain - Life detection service - Terrain search service -  Swarm of robots Search for a Composition Solution • Problem: independent admission control in each domain QoS-driven Service Composition Service Composition Service Domain Service Domain Services Services • Execution environment in each domain: • Each service runs on a VM • How to allocate cores to VMS • (or other resources to services) • Manage resources and power such that: • Best satisfy QoS goals of all clients • Minimize power consumption • (e.g., dynamic voltage scaling) QoS properties of the composed system …... Platforms Platforms Platform Services VMs Service Composition Cores …... Need to perform admission control: To ensure that the admitted workloads will have agreed upon levels of QoS and will not violate power constraints • Correspondingly, service composition should consider • Execution environment (e.g., admission control and energy-awareness) • Dynamic adaptation (e.g., selecting configurable services rather than • those that provide the most satisfactory QoS for the time being • System QoS prediction from QoS of individual constituent services • All the above: May have many potential candidates to consider •  efficiency concerns

  45. Search for a Composition Solution • Major issues: • Each domain requires admission control in order to achieve the guaranteed service agreement • But during the service selection time, the execution environment (whether the service can be admitted) is not specifically considered • If we do consider admission at the selection time: • There may be many matching services • Interacting with each matching service to find out the feasibility can be overkill • When considering composition, this can be worse

  46. Search for a Composition Solution • Reference: Service Composition for Real-Time Assurance • Multi-level agreement approach • First: find a set of top-choice compositions based on published provider information, without concerning the admission • When deciding top choices, better ensure that there are no common services in all choices; otherwise, if admission for one of these services fails, no composition is valid • Then proceed to check the admission and proceed with agreement • When to proceed with the agreement? In a composite service, if one of the atomic service selected cannot admit the customer, then all the agreements should be voided • If a provider agrees to provide a service and reserved the resources for customer x, and because of this, it denied another customer y. When customer x revokes all the agreements, what can be done? • Need a pre-agreement phase

  47. Search for a Composition Solution • Multi-level agreement approach • In the first level, use simple QoS estimation methods • May have a large number of compositions to be estimated • Time: simply add together • Reliability: simply multiply together • Security: only evaluate important policies • In the higher levels: Predict QoS with more sophisticated methods • Time: consider queuing analysis • Reliability, consider sophisticated mapping of operational profiles • In the final stage • Simulation to validate the correctness of the composition (and estimation)

  48. QoS in Service Composition • Decision algorithms • Has been the focus of many papers • But is not the major difficulties in QoS-driven service composition • QoS property aggregation • Highly challenging • Admission issues • No perfect solution • Multi-level approach would help • Do not always go for best-choice would help diversify the selections of different customers