Modeling of Event-based Communication in Component-based Architectures

Transcription

1 Modeling of Event-based Communication in Component-based Architectures Christoph Rathfelder, Benjamin Klatt and Samuel Kounev Keynote talk presented by Samuel Kounev at ETAPS 2012 Tallinn, Estonia, March 31, 2012 DESCARTES RESEARCH GROUP, CHAIR FOR SOFTWARE DESIGN AND QUALITY INSTITUTE FOR PROGRAM STRUCTURES AND DATA ORGANIZATION, FACULTY OF INFORMATICS KIT University of the State of Baden-Wuerttemberg and National Research Center of the Helmholtz Association

2 Agenda Introduction to Event-based Communication Design-time Modeling and Analysis Palladio Component Model (PCM) Meta-Model Extensions for Event-based Communication Case Studies Run-Time Quality-of-Service Management Descartes Meta-Model (DMM) Outlook

3 References (1) Christoph Rathfelder. Modelling Event-Based Interactions in Component-Based Architectures for Quantitative System Evaluation. PhD Thesis, KIT, In preparation, March

4 References (2) [1] B. Klatt, C. Rathfelder and S. Kounev, Integration of event-based communication in the palladio software quality prediction framework, in Proceedings of the joint ACM SIGSOFT conference - QoSA and ACM SIGSOFT symposium - ISARCS on Quality of software architectures - QoSA and architecting critical systems - ISARCS, QoSA-ISARCS '11, Seiten 43-52, New York, NY, USA; Best Paper Nomination, [2] C. Rathfelder and B. Klatt, A quality-prediction tool for component-based architectures, in Proceedings of the 2011 Ninth Working IEEE/IFIP Conference on Software Architecture, WICSA '11, Seiten , Washington, DC, USA, [3] C. Rathfelder and S. Kounev, "Model-based performance prediction for event-driven systems," in Proceedings of the Third ACM International Conference on Distributed Event-Based Systems, DEBS '09, Seiten 33:1-33:2, [4] C. Rathfelder and S. Kounev, Modeling Event-Driven Service-Oriented Systems using the Palladio Component Model, in Proceedings of the International Workshop on the Quality of Service-Oriented Software Systems (QUASOSS), Seiten , [5] C. Rathfelder, D. Evans and S. Kounev, Predictive Modelling of Peer-to-Peer Event-driven Communication in Component-based Systems, in Proceedings of the 7th European Performance Engineering Workshop (EPEW'10), volume 6342 of Lecture Notes in Computer Science, Seiten , University Residential Center of Bertinoro, Italy, [6] C. Rathfelder, B. Klatt, S. Kounev and D. Evans, Towards middleware-aware integration of event-based communication into the palladio component model, in Proceedings of the Fourth ACM International Conference on Distributed Event-Based Systems, DEBS '10, Seiten 97-98, [7] C. Rathfelder, S. Kounev and D. Evans, Capacity Planning for Event-based Systems using Automated Performance Predictions, in 26th IEEE/ACM International Conference On Automated Software Engineering (ASE 2011), Seiten , Oread, Lawrence, Kansas; Annahmequote: 14.7% (37/252), [8] C. Rathfelder, B. Klatt, F. Brosch and S. Kounev, Performance Modeling for Quality of Service Prediction in Service-Oriented Systems, in Handbook of Research on Service-Oriented Systems and Non-Functional Properties: Future Directions, Seiten , IGI Global, [9] C. Rathfelder, B. Klatt, K. Sachs und S. Kounev, Modeling Event-based Communication in Component-based Software, Journal on Software and Systems Modeling; Theme Issue on Models for Quality of Software Architecture, 2012; Invited paper under review

5 References (3) Kai Sachs. Performance Modeling and Benchmarking of Event-Based Systems. PhD Thesis, TU Darmstadt, ISBN-13: , Sierke Verlag, August 30,

6 Agenda Introduction to Event-based Communication Design-time Modeling and Analysis Palladio Component Model (PCM) Meta-Model Extensions for Event-based Communication Case Studies Run-Time Quality-of-Service Management Descartes Meta-Model (DMM) Outlook

7 Event-based Communication Decoupling in the 3 dimensions [Eugster] Communicating parties: 1. Do not need to be active at the same time (Time) 2. Do not need to know each other (Space) 3. Are not blocked when exchanging messages (Synchronization)

8 Event-based Communication (2) Event A significant change in state Source Producer, publisher, sender, generator, or monitoring component. Transmission System Notification service, event service, event-based middleware, channel or event bus. Sink Reactive components, consumers, subscribers, or receivers

9 Event Delivery Model

10 Motivating Example Store Scenario Source UpdateStockData Sink UpdateStockData RFID Scanner Cashdesk Service Source UpdateStockData Event Bus Sink UpdateStockData Order Managment Service Prov. Interface CreateOrder Inventory Management Service Req Interface CreateOrder Logging Service

11 Motivating Example Store Scenario Changing Usage Source UpdateStockData Prediction Max. Throughput Processing Time Sink UpdateStockData Prov. Interface CreateOrder RFID Scanner Order Managment Service vs. Inventory Management Service Sizing Req Interface CreateOrder vs. vs. Cashdesk Service Logging Service Source UpdateStockData Analyses Event Bus Bottleneck Sink Resource UpdateStockData Utilization

12 Agenda Introduction to Event-based Communication Design-time Modeling and Analysis Palladio Component Model (PCM) Meta-Model Extensions for Event-based Communication Case Studies Run-Time Quality-of-Service Management Descartes Meta-Model (DMM) Outlook

13 Palladio Software Quality Prediction Framework Palladio Component Model Domain specific modelling language Aligned with UML 2 design models Component-based architectures Eclipse-based modelling and prediction tool Design-time quality prediction Transformations into prediction models Simulation code Layered queueing networks Queueing Petri nets

14 Palladio Modelling Approach Performance Analyses: Component Repository System Model Deployment Model Usage Model

15 Model Solution Techniques Component Model Simulation Code Service-Level Prediction Architecture Model Layered Queueing Networks Resource Utilization Deployment Model Usage Model Palladio Approach Queueing Petri Nets Stochastic Regular Expr. Response Time

16 Agenda Introduction to Event-based Communication Design-time Modeling and Analysis Palladio Component Model (PCM) Meta-Model Extensions for Event-based Communication Case Studies Run-Time Quality-of-Service Management Descartes Meta-Model (DMM) Outlook

17 Integration of Event-Based Communication in the Palladio Software Quality Prediction Framework Event Benjamin Klatt, FZI Christoph Rathfelder, FZI Samuel Kounev, KIT B. Klatt, C. Rathfelder and S. Kounev. Integration of Event-Based Communication in the Palladio Software Quality Prediction Framework. In 7th ACM SIGSOFT International Conference on the Quality of Software Architectures (QoSA 2011), pages 43-52, Boulder, Colorado, USA, June 20-24,

18 Events and Components

19 Direct Point-to-Point Connections

20 Graphical Notation Event Source Event Sink Interface Component Component Component Component Component Component

21 Publish/Subscribe Communication

22 PCM Meta-Model Extensions Explicit modelling of Events Source and sink ports Many-to-many connections Event producer (Source) and handler (Sink) Sources, Sinks & Connectors Components & Roles Emit Event Action

23 Approach Quality-Prediction of component-based architectures with event-based communication integrated into Palladio Explicit event modeling with reduced effort Considering event communication middleware influences Reuse existing prediction techniques Extended PCM Model Event Original PCM Model Platform Specific Model Prediction Result Middleware Middleware Transformation Weaving Prediction

24 1. Meta-Model Extension Extended PCM platform-independent Extension of the PCM meta-model Semantically correct modelling of Events Source and sink ports 1-many connections Event handlers

25 2. Model-to-Model Transformation Extended PCM platform-independent Classical PCM platform-independent Transformation Extensions are substituted with existing elements Performance equivalent model Allows reuse of existing prediction techniques Does not consider platformspecific resource demands

26 3. Weaving of Platform Specific Components Extended PCM platform-independent Transformation Middleware repository Platform-specific components Behaviour Resource demands Classical PCM platform-independent Weaving Middleware Repository platform-specific Weaving with platformindependent model Input for quality predictions Final Model platform-specific Quality Prediction

27 Transformation Example Ext.PCM Class. PCM Middleware Repository Final Model Prediction A Source A Sink B B Sink C C A IEvent Source SourcePort Distribution Preparation Source A Event Distribution Event Sender Event Sender IEvent Receiver IEvent Receiver Event Receiver Sink C SinkPort C IEvent Sink Platform independent IMiddleware SourcePort IMiddleware DistributionPreparation IMiddleware EventDistribution IMiddleware Sender IMiddleware Receiver IMiddleware SinkPort Platform specific Middleware SourcePort Middleware Hub Middleware SinkPort

28 Agenda Introduction to Event-based Communication Design-time Modeling and Analysis Palladio Component Model (PCM) Meta-Model Extensions for Event-based Communication Case Studies Run-Time Quality-of-Service Management Descartes Meta-Model (DMM) Outlook

29 Predictive Modelling of Peer-to-Peer Event-driven Communication in Component-based Systems Event Christoph Rathfelder FZI, Germany David Evans University of Cambridge Samuel Kounev KIT, Germany C. Rathfelder, D. Evans, and S. Kounev. Predictive Modelling of Peer-to-Peer Event-driven Communication in Component-based Systems. In Proceedings of the 7th European Performance Engineering Workshop (EPEW'10), University Residential Center of Bertinoro, Italy, volume 6342 of Lecture Notes in Computer Science, pages Springer-Verlag Berlin Heidelberg,

30 Motivating Example Traffic Monitoring in Cambridge TrafficLight Sensors Acis Location Acis Sensors RedlightBus Proximity DB Location Storage vs

31 Motivating Example Design Alternatives Traffic Monitoring in Cambridge One component per Traffic light Intersection District New components Speeding detection Prediction TrafficLight Sensors vs Acis Location Acis Sensors Changing Usage vs Analyses Processing Time Resource Utilisation Max. Throughput Bottleneck Event delivery latency RedlightBus Proximity DB Location Storage Deployment and Sizing vs

32 Preliminary Case Study - Application Traffic Monitoring in Cambridge Output of TIME-EACM research project Real world scenario SBUS (Stream Bus) middleware Supports RPC as well as event streams Developed in Cambridge GPS Location Data ACIS Location Storage Traffic Light Status SCOOT Proximity Detection Project website:

33 SBUS Middleware

34 Preliminary Case Study System Model

35 Preliminary Case Study Evaluation Prediction of Event processing time CPU utilisation Different scenarios Several instances Different deployment Up to 4 quad-core machines Variation of event rates Prediction error Mostly < 10% Never exceeded 20% Processing time prediction CPU utilisation prediction

36 Preliminary Case Study Evaluation (2) Initial Modelling Effort Model Adaptation Effort

37 Capacity Planning for Event-based Systems using Automated Performance Predictions Event Christoph Rathfelder FZI, Germany Samuel Kounev KIT, Germany David Evans University of Cambridge C. Rathfelder, S. Kounev, and D. Evans. Capacity Planning for Event-based Systems using Automated Performance Predictions. In 26th IEEE/ACM International Conference On Automated Software Engineering (ASE 2011), pages , Oread, Lawrence, Kansas, November

38 Motivating Example Cam Cam License Plate Recognition Toll Event Bus Bus Sensors Traffic Control Location Bus Proximity Speeding

39 Motivating Example Changing Load System Evolution Cam vs. Cam License Plate Recognition New components New algorithms Vs. Toll Event Bus Sizing and Capacity Planning Bus Sensors Traffic Control Max. throughput Resource utilization Location Latency Bottlenecks Bus Proximity Speeding Over-provisioning of hardware resources

40 Capacity Planning using Performance Predictions System Modeling/ Resource Demand Estimation Capacity Planning System Deployment/ Reconfiguration System Evolution/ Workload Changes Model Adaptation Performance Predictions Result Evaluation Capacity Planning QoS requirements fullfilled? Yes Variation of Architecture/Deployment No Resources used efficiently? Yes Yes End of life?

41 Automated Performance Prediction Process 1 Parameter Variation 2 M2M-Event- Transformation Middleware- Weaving No Yes End of 3 Parameter Range? Solving/ Simulation Transformation Prediction Model 1. Systematic variation of simulated workload intensity 2. Automated model transformations 3. Simulation-based performance prediction using PCM

42 Transformation Example No Yes Parameter Variation End of Parameter Range? Solving/ Simulation M2M-Event- Transformation Transformation Prediction Model Middleware- Weaving B A C A IEvent Source SourcePort Distribution Preparation Source A Event Distribution Event Sender Event Sender IEvent Receiver IEvent Receiver Event Receiver Sink C SinkPort C IEvent Sink Platform independent IMiddleware SourcePort IMiddleware DistributionPreparation IMiddleware EventDistribution IMiddleware Sender IMiddleware Receiver IMiddleware SinkPort Platform specific SBUS SourcePort SBUS Middleware SBUS SinkPort

43 Case Study Traffic Monitoring System Based on output of TIME-EACM research project Real world scenario with data from the City of Cambridge SBUS (Stream Bus) middleware Supports RPC as well as event streams Developed in Cambridge Scenarios System variations Evaluation of deployment options Maximal throughput Hardware utilization Latency Bottlenecks ACIS SCOOT Cam Cam Cam Speeding Location License Plate Recognition Toll Bus Proximity

44 Specification of Architecture-level Prediction Model Component Repository Controlled experiments for each component Resource demand estimation based on time measurements Probabilistic and parameter dependent 1 Variant System Model Instantiation and connection of components Variations depending on scenario (e.g., load balancing) 3 Variants Deployment and Hardware Specification of hardware resources Deployment of components depending on scenario 7 Variants Usage Model Automated variation of workload > 100 Variants

45 Scenario 2 - Growing Workload Adding additional cameras causes additional workload License Plate Recognition (LPR) is bottleneck Result: Small improvement by deploying LPR on dedicated server CPU utilization [%] AllOnOne Distrib. LPD LPR Distrib. Other , Cam Cam Cam Cam Cam Cam License Plate Recognition Cam Cam Cam Cam Cam Cam License Plate Recognition Timespan Image/event between rate two [1/s] images [s] Speeding Speeding ACIS Location ACIS Location SCOOT Bus Proximity SCOOT Bus Proximity AllOnOne Distributed

46 Scenario 3 - Additional Component New Toll component LPR is the bottleneck 2 additional servers Load balancing on 3 LPR instances Result: Centralized deployment of Speeding and Toll CPU utilization [%] LPR: LPR 3 LPR cent.: LPR 3 LPR: decent.: LPR Proc. central decent Timespan Image/event between two rate images [1/s] [s] Cam Cam Cam Cam Cam Cam License Plate Recognition Cam Cam Cam Cam Cam Cam License Plate Recognition Cam Cam Cam Cam Cam Cam License Plate Recognition Speeding Toll Speeding Toll Speeding Toll ACIS Location ACIS Location ACIS Location SCOOT Bus Proximity SCOOT Bus Proximity SCOOT Bus Proximity LPR 3 LPR cent. 3 LPR decent.

47 Scenario 4 Improved Cameras New cameras with higher resolution Improved LPR success rate Higher overall CPU demand for processing Result: Decentralized deployment of Speeding and Toll CPU utilization [%] Cent. LPR Decent. LPR Cent. Proc. CPU utilization [%] Cent. LPR Decent. LPR Cent. Proc Image/event rate [1/s] Timespan between two images [s] Old cameras Image/event rate [1/s] Timespan between two images [s] New cameras

48 Evaluation of Prediction Accuracy Deployment of each scenario in our testbed Load drivers with configurable event rate using real world data Compare measured and predicted values in different load situations S1 S2 S3 Gigabit Switch S4 S5 S6 S7 S8 S9 S10 S11 S12 Each machine equipped with: Intel Core 2 Quad Q6600 2,4GHz, 8GB RAM, Ubuntu 8.04 Experiment Controller

49 [%] [%] Evaluation Results CPU utilization Scenario3 CPU utilization Scenario 3 LPR LPD Meas. (decent.) LPR LPD Pred. (decent.) LPR LPD Meas. (cent.) LPR LPD Pred. (cent.) Proc. Meas. (cent.) Proc. Pred. (cent.) Frequency Image/event of images rate per [1/s] Cam [1/s] CPU utilization Scenario 4 CPU utilization Scenario 4 Meas. (cent., old) Pred (cent., old) Meas. (decent., old) Pred. (decent., old) Meas. (cent., new) Pred. (cent., new) Meas. (decent., new) Pred. (decent., new) Frequency Image/event of images rate per [1/s] Cam [1/s] Prediction error: Utilization always underestimated Mean error < 20%, Max error < 25%

50 Effort Reduction Architecture-level prediction models Eased variation of architecture and deployment Automated model transformation for events 80% less manual element creations compared to manual modeling Automation of performance predictions Time saving Prediction time: 3 min Experiment run time: 2.7 hours Automated load-variation

51 Conclusion Capacity Planning Based on automated performance predictions Prediction error < 25% Always underestimated Improves the system s efficiency Often over-provisioning by factor 2 and more Effort Reduction Modeling effort for event-based systems reduced by 80% Significant time saving by using prediction techniques

52 Agenda Introduction to Event-based Communication Design-time Modeling and Analysis Palladio Component Model (PCM) Meta-Model Extensions for Event-based Communication Case Studies Run-Time Quality-of-Service Management Descartes Meta-Model (DMM) Outlook

53 Descartes Meta-Model (DMM) Architecture-level modeling language for self-aware run-time systems management of modern IT systems, infrastructures and services Main Goal: Provide Quality-of-Service (QoS) guarantees Performance (current focus) Response time, throughput, scalability and efficiency Or more generally, dependability Including also availability, reliability and security aspects

54 1) Self-Reflective Aware of their software architecture, execution environment and hardware infrastructure, as well as of their operational goals 2) Self-Predictive Able to anticipate and predict the effect of dynamic changes in the environment, as well as the effect of possible adaptation actions 3) Self-Adaptive Proactively adapting as the environment evolves to ensure that their operational goals are continuously met

55 Descartes Meta-Model (DMM) Collection of several meta-models each focusing on different system aspects

56 PCM and DMM Palladio Component Model (PCM) Descartes Meta-Model (DMM) Design-time aspects Run-time aspects

57 Design-time vs. Run-time Models Two orthogonal dimensions Modeling of design-time vs. run-time aspects Use of models at design-time vs. run-time Fine granular differentiating factors 1. Model purpose 2. Model target users / consumers 3. Degrees of freedom in model use case scenarios 4. Model structure & parameterization 5. Possibilities for model calibration 6. Required model flexibility

58 Descartes vs. Palladio PCM DMM

59 Example Usage Scenarios Example: Design-time QoS analysis Example: Run-time QoS management Example: Elasticity evaluation at design-time Example: Design-time QoS analysis in a DMM resource landscape

60 DMM Technical Report

61 Discussion