Model-Driven Engineering of Component-based Distributed Real-time & Embedded Systems

1. Institute for Software Integrated Systems Vanderbilt University Nashville, Tennessee Model-Driven Engineering of Component-based Distributed Real-time & Embedded Systems STI Project 2007 Status Report Krishnakumar Balasubramanian, Amogh Kavimandan {kitty, amoghk}@dre.vanderbilt.edu

2. STI Project 2007 Status Report • Vanderbilt’s contributions • Model-Driven Engineering tool-chain • Current year’s focus on two areas • Model-driven Application Specific Optimizations (Kitty) • Model-driven QoS Mapping (Amogh)

3. Distributed Real-time & Embedded (DRE) Systems • Stringent Quality-of-Service (QoS) demands, e.g., real-time constraints • Operate under limited resources • e.g., avionics mission computing • Enterprise DRE Systems • Simultaneous execution of multiple applications with varying importance • Highly heterogeneous platform, language & tool environments • e.g., Total Shipboard Computing Environment (TSCE) • Use COTS middleware technologies • CORBA, RT-Java • Use COTS Component/Service technologies • CORBA Component Model (CCM), EJB, Web Services

4. Motivation for Application Specific Optimizations • Middleware tries to optimize execution for every application • Collocated method invocations • Optimize the (de-)marshaling costs by exploiting locality • Specialization of request path by exploiting protocol properties • Caching, Compression, various encoding schemes • Reducing communication costs • Moving data closer to the consumers by replication • Reflection-based approaches • Choosing appropriate alternate implementations at run-time

5. Related Research

6. Related Research: What’s missing? • Lack of a high-level notation to guide optimization frameworks • Missing AST of application

7. Related Research: What’s missing? • Lack of a high-level notation to guide optimization frameworks • Missing AST of application • Detection at run-time (reflection) • Additional overhead in the fast path • Not suitable for DRE systems • Intrusive optimizations, i.e., not completely application transparent • Requires providing multiple implementations, e.g., EJB • Optimization performed either • Too early, or too late

8. Application Specific Optimizations: Unresolved Challenges • Overhead of platform mappings • Blind adherence to platform semantics • Inefficient middleware glue code generation per component • Example: • Creation of a factory object per component • Servant glue-code generation for every component Need optimization techniques to build large-scale component systems!

9. Solution Approach: Physical Assembly Mapper (PAM) • Physical Assembly Mapper • Uses the application model as the input • Exploits knowledge about platform semantics to rewrite the input model to a functionally equivalent output model • Generates middleware glue-code • Generates deployment configuration files • Operates just before deployment • Can be viewed as a “deployment compiler”

10. PAM Inputs • Set of 3 input graphs • G1: (V1, E1), V1 = { Application Components} E1 = { Connections between application components} • G2: (V2, E2), V2 = V1 U {QoS configuration options on components} E2 = { Connections between the components and the QoS options} • G3: (V3, E3), V3 = V1 U {nodes in the target deployment domain} E3 = {Connections between the components and the nodes}

11. PAM Output • Produces an output graph • Go: (Vo, Eo), • Vo = {Physical assemblies created on a single node} • Eo = {Connections between the composites} • Creation of physical assemblies subject to a number of constraints • Average case • |Vo| < |V1| • Worst case • |Vo| = |V1|, i.e., the optimizer couldn’t create any physical assemblies • Equivalent to deployment of original application

12. Physical Assembly • Physical Assembly • Created from the set of components that are deployed within a single process of a target node • Subject to various constraints • Example constraints include: • No two ports of the set of components should have the same name • No two components should have incompatible RT-CORBA policies • No changes required to individual component implementations

13. Physical Assembly Generation • Given set of components deployed on a single process of a target node • Compute pair-wise intersections of component port name sets • If intersection is null, merge the two components into a physical assembly • Additional RT-CCM constraint • If intersection is null & RT-CORBA policies of the components are compatible • e.g., PriorityPropagation Models, PriorityBands • Physical Assembly indistinguishable to external clients • All valid operations on individual components are still valid

14. Middleware Concepts Exploited by Physical Assemblies • Opacity of object references • Components don’t rely on specific details of object references, e.g., location of type information • Allows replacing references transparent to component implementations

15. Main Main Component Component Component Component Specific Specific Executor Executor Context Context CCMContext CCMContext Executors Executors Executors Executors Executors Executors CCMContext CCMContext EnterpriseComponent EnterpriseComponent Servant Servant POA POA Container Container Middleware Concepts Exploited by Physical Assemblies • Opacity of object references • Components don’t rely on specific details of object references, e.g., location of type information • Allows replacing references transparent to component implementations • Presence of a component context • Components access ports of other components using a context object • Allows replacing context transparent to component implementations s Container l e a c n a r e f r t CORBA e x t E n I Component user implemented code Internal Interfaces POA

16. Middleware Concepts Exploited by Physical Assemblies • Opacity of object references • Components don’t rely on specific details of object references, e.g., location of type information • Allows replacing references transparent to component implementations • Presence of a component context • Components access ports of other components using a context object • Allows replacing context transparent to component implementations • Clean separation between glue-code & component implementation • Allows modifications transparent to component implementations Technique can be applied to other middleware with these properties

17. Physical Assembly Evaluation Criteria • Footprint of physical assembly components • Compared to vanilla Component-Integrated ACE ORB (CIAO) • Different scenarios • Simple scenarios (<= 10 components) • ARMS GateTest scenarios (150+ components) • Reduce static & dynamic footprint • Reduce no. of homes by (n – 1) / n • Reduce no. of objects registered with POA by (n – 1) / n • Reduce no. of context objects created by (n – 1) / n n = no. of components deployed on a single process of a target node

18. Evaluating Physical Assemblies • Boeing’s BoldStroke Basic Single Processor Scenario • Consists of 4 components • Timer – Periodically sends refresh signal to the GPS • GPS – Calculates new co-ordinates of the aircraft in response to Timer signal • Airframe – Processes new location inputs from GPS • Display – Updates the new location of the aircraft in the navigation display

19. Applying Physical Assemblies to BasicSP Scenario • Assumption: GPS, Airframe and Display components are deployed on a single processor board • Applying physical assemblies • Combines GPS, Airframe and Display components into a single physical assembly (BasicSPAsm) • Maintains the same number of connections • Timer component not combined (due to clash in port names)

20. Experimental Results: BasicSP Scenario • Testbed • Linux 2.6.20 FC6 • Dual 2.4Mhz processor • 1.5GB RAM • Evaluation for larger applications (~150 components) is in progress Dynamic footprint reduction of ~10% Static footprint reduction of ~45% Physical assembly mapping reduces the footprint significantly

21. Application Specific Optimizations: Unresolved Challenges • Lack of application context • Optimization decisions relegated to run-time • e.g., every invocation performs check for locality • Settle for near-optimal solutions • Missed middleware optimization opportunities

22. Application Specific Optimizations: Unresolved Challenges • Lack of application context • Optimization decisions relegated to run-time • e.g., every invocation performs check for locality • Settle for near-optimal solutions • Missed middleware optimization opportunities • e.g., allocation of RT-CORBA thread pools and lanes • Large overhead compared to collocated calls Invocation to a different pool/lane (150µs) Invocation within the same pool/lane (12µs) Cannot be solved efficiently at middleware level alone!

23. Solution Approach: Use Application Context from Models • Use application structure & context available in models • Create fast path within middleware for physical assemblies • Cross-pool/lane proxy

24. Solution Approach: Use Application Context from Models • Use application structure & context available in models • Create fast path within middleware for physical assemblies • Cross-pool/lane proxy • Utilize available fast path in PAM • e.g., Use matching real-time policies as additional constraint when creating physical assemblies

25. Solution Approach: Use Application Context from Models • Use application structure & context available in models • Create fast path within middleware for physical assemblies • Cross-pool/lane proxy • Utilize available fast path in PAM • e.g., Use matching real-time policies as additional constraint when creating physical assemblies • Configure middleware resources efficiently • e.g., allocate physical assemblies with matching policies in the same thread pool or thread pool with lanes

26. Context-driven Optimization Evaluation Criteria • Improve performance • t = no. of cross-pool/lane interactions between components within an assembly • Transform t remote calls to t cross-pool/lane calls • Eliminate mis-optimizations • Check incompatible POA policies • Incompatible invocation semantics (oneway or twoway) • No changes to individual component implementations • Eliminate need for a local vs. remote version • Customizable & application transparent

27. Experimental Results: Cross pool/lane proxy Average latency improved by 60 – 66% Standard deviation <= 3µs 99% latency improved by 60-66% Max latency improved by ~50% Significant performance benefits with cross pool/lane proxy!

28. Summary of Research Contributions

29. QoS Mapping Context Benefits of QoS-enabled middleware technologies Raise the level of abstraction Support many quality of service (QoS) configuration knobs Component Component Reference Reference Container Priority Band Container End-to-End Priority Component Component Propagation Home Home s COMPONENT s COMPONENT t t e e x l x l F F c c e e EXECUTORS a t EXECUTORS a a t a t c n t n c p p e o o e t e e s t C C s s l s c e c l a t e t e c e a n n c n a R R n r e a e f r e r f n n e t r e s s t o n e t o t e t n t I e n p n p I n c n S E c I S E e r m I r e i v m u i v v n u v o e n o e o E k Callback o E n k C Callback s C S n t S s t Interfaces Interfaces POA POA ORB Protocol Properties Portable Priorities COMPONENT SERVER 2 COMPONENT SERVER 1 Publisher(Subscriber) multi-threaded access Event Filtering Event Dispatching 29

30. QoS Mapping Context Benefits of QoS-enabled middleware technologies Raise the level of abstraction Support many quality of service (QoS) configuration knobs Component Reference Container Component Home s COMPONENT t e x l F c e EXECUTORS a t a t c n p e o t e s C s l c e a t e c n n a R r e f e r n t e s o n t t e I n p n c E S I e r m i v u v n o e o E k Callback n C s S t Interfaces POA ORB COMPONENT SERVER • Drawbacks of QoS-enabled middleware technologies • Achieving desired QoS increasingly a system QoSconfiguration problem, not just an initial system functionaldesign problem Lack of effective QoS configuration tools result in QoS policy mis-configurartions that are hard to analyze & debug 30

31. Example Application: Satellite Mission • Satellite mission consists of four identically instrumented spacecraft & a ground control system • Collect mission data • Send it to ground control at appropriate time instances

32. Example Application: Satellite Mission Satellite mission QoS requirements span two dimensions Multiple modes of operation Varying importance of data collection activity of satellite sensors based on mission phase Need to translate QoS policies into QoS configuration options & resolve QoS dependencies • Slow Survey • Fast Survey • Burst 32

33. Challenge 1: Translating QoS Policies to QoS Options Policy for handling of dangling(ill-behaved) publishers(subscribers) (QoS Policy) must be mapped to control policy and control period (QoS configuration) Prioritized service invocations (QoS Policy) must be mapped to Real-time CORBA Banded Connection (QoS configuration) • Large gap between application QoS policies & middleware QoS configuration options • Bridging this gap is necessary to realize the desired QoS policies • The mapping between application-specific QoS policies & middleware-specific QoS configuration options is non-trivial, particular for large systems

34. Challenge 1: Translating QoS Policies to QoS Options Conventional mapping approach requires deep understanding of the middleware configuration space e.g., multiple types/levels of QoS policies require configuring appropriate number of thread pools, threadpool lanes (server) & banded connections (client) Client Propagation & Server Declared Priority Models Static Scheduling Service Standard Synchonizers Explicit Binding Thread Pools Portable Priorities Protocol Properties Request Buffering 34

35. Challenge 2: Choosing Appropriate QoS Option Values • Individually configuring component QoS options is tedious & error-prone • e.g., ~10 distinct QoS options per component & ~140 total QoS options for entire satellite mission • Manually choosing valid values for QoS options does not scale as size & complexity of applications increase

36. Challenge 3: Validating QoS Options • Each QoS option value chosen should be validated • e.g., Filter priority model is CLIENT_PROPAGATED, whereas Comm priority model is SERVER_DECLARED • Each system reconfiguration (at design time) should be validated • e.g., reconfiguration of bands of Analysis should be validated such that the modified value corresponds to (some) lane priority of the Comm 36

37. Challenge 4: Resolving QoS Option Dependencies “QoS option dependency” is defined as: Dependency between QoS options of different components Manually tracking dependencies is hard – or in some cases infeasible Dependent components may belong to more than one assembly Dependency may span beyond immediate neighbors e.g., dependency between Gizmo & Comm components Empirically validating configuration changes by hand is tedious, error-prone, & slows down development & QA process considerably Several iterations before desired QoS is achieved (if at all) ThreadPool priorities of Comm should match priority bands defined at Gizmo 37

38. Solution Approach: Model-Driven QoS Mapping • QUality of service pICKER (QUICKER) • Model-driven engineering (MDE) tools model application QoS policies • Provides automatic mapping of QoS policies to QoS configuration options • Validates the generated QoS options • Automated QoS mapping & validation tools can be used iteratively throughout the development process

39. QUICKER Enabling MDE Technologies Enhanced Platform Independent Component Modeling Language (PICML), a DSML for modeling component-based applications QoS mapping uses Graph Rewriting & Transformation (GReAT) model transformation tool Customized Bogor model-checker used to define new types & primitives to validate QoS options 39

40. QUICKER Enabling MDE Technologies Enhanced Platform Independent Component Modeling Language (PICML), a DSML for modeling component-based applications QoS mapping uses Graph Rewriting & Transformation (GReAT) model transformation tool Customized Bogor model-checker used to define new types & primitives to validate QoS options CQML Model Interpreter Bogor Input Representation CQML Model interpreter generates Bogor Input Representation (BIR) of DRE system from its CQML model 40

41. Resolving Challenge 1: Translating Policies to Options (1/2) • Expressing QoS policies • PICML modes application-level QoS policies at high-level of abstraction • e.g., multiple service levels support for Comm component, service execution at varying priority for Analysis component, prioritizing event dispatching at Analysis component • Reduces modeling effort • e.g., ~25 QoS policy elements for satellite mission vs. ~140 QoS options

42. Resolving Challenge 1: Translating Policies to Options (2/2) Mapping QoS policies to QoS options GReAT model transformations automate the tedious & error-prone translation process Transformations generate QoS configuration options as CQML models Allow further transformation by other tools e.g., code optimizers & generators Simplifies application development & enhances traceability 47

43. Resolving Challenges 2 & 3: Ensuring QoS Option Validity • CQML model interpreter generates BIR specification from CQML models • BIR primitives used to check whether a given set of QoS options satisfies a system property • e.g., fixed priority service execution, a property of Comm component • Supports iterative validation of QoS options during QoS configuration process QUICKER

44. Resolving Challenge 4: Resolving QoS Option Dependencies • Dependency structure maintained in Bogor used to track dependencies between QoS options of components, e.g.: • Analysis & Comm are connected • Gizmo & Comm are dependent • Change(s) in QoS options of dependent component(s) triggers detection of potential mismatches • e.g., dependency between Gizmo invocation priority & Comm lane priority Detect mismatch if either values change Dependency Structure of Satellite Mission Components

45. Summary of Research Contributions QUICKER provides Model-Driven Engineering (MDE) for QoS-enabled component middleware Maps application-level QoS policies to middleware-specific QoS configuration options Model transformations automatically generate QoS options Model-checking extensions ensure validity of QoS options at component- & application-level 50