A Distributed Data Flow Model for Composing Software...

Stanford University Ph.D. Oral ExaminationDepartment of Electrical Engineering

A Distributed Data Flow Model for Composing Software Services

David W. LiuMay 9, 2003

2

Presentation OutlinePresentation Outline

• Motivation and Objectives

• Theoretical Analysis

• FICAS Service Composition Infrastructure

• Summary

3

Motivation and Objectives

4

Paradigm Shift in Software EngineeringParadigm Shift in Software Engineering

Coding

Integration

1970 1990 2010

Coding

Integration

1970 1990 2010

Courtesy of Professor Gio Wiederhold

5

Distributed Service ModelDistributed Service Model

AutonomousService

AccessProtocol

OperatingSystem

Host

A

A

A

AutonomousService

AccessProtocol

OperatingSystem

Host

N

N

N

Communication Network

Megaservice:Conceptual Composition of

Autonomous Services

AutonomousService

AccessProtocol

OperatingSystem

Host

B

B

B

......

6

Research ObjectivesResearch Objectives

• Demonstrate the efficiency of the distributed data-flow model• Define a framework for constructing software services• Provide tools for composing software services• Investigate techniques for performance optimization

Service1

Service3

MegaService

(a) Centralized Data-flow Model (b) Distributed Data-flow Model

Control-flow

Data-flow

Service2

Service1

Service3

MegaService

Service2

7

Theoretical Analysis

8

Service Integration ModelsService Integration Models

M

S1

S2 S3

S4

M

S1

S2 S3

S4

M

S1

S2 S3

S4

M

S1

S2 S3

S4

(a) Centralized Control-flow andCentralized Data-flow Model (1C1D)

Control-flows

Data-flows

MegaserviceM

AutonomousServices

S

(b) Centralized Control-flow andDistributed Data-flow Model (1CnD)

(c) Distributed Control-flow andCentralized Data-flow Model (nC1D)

(d) Distributed Control-flow andDistributed Data-flow Model (nCnD)

9

System ModelingSystem Modeling

CP2 CPn

CPn

CP1

CP0CM0n/CMn0

S1: f1 (SI1, SP1, SO1)

CM01/C

M10

S2: f2 (SI2, SP2, SO2) Sn: fn (SIn, SPn, SOn)

Sn: fn (SIn, SPn, SOn)M: (_, MP, _)

CP: processing powerCM: communication bandwidthλ: message header sizeδ: data distribution coefficient

f: invocation frequencySI: input data sizeSP: processing loadSO: output data size

iijij SOdd=δ

10

Aggregated CostAggregated Cost

Aggregated cost = Amount of system resource consumedby a megaservice

• Centralized data-flow model incurs more data traffic• Distributed data-flow model incurs more message overheads

( )∑=

+×=−n

imessagedatadc iDiDMCOSTMCOST

1)()()()( β

where

×−−××=

−××=

∑=

n

jjimessage

iiidata

jimfimfiD

SOfiD

1

0

),())0,(1()(

)1()(

λ

δ

≠=

=0100

),(ij

ij

ifif

jimδδ

Output data from service

Portion returned to megaservice

Weight of communication cost

Data messages among services

Data messages returnedto megaservice Data message overhead

11

Messaging Cost for a Service InvocationMessaging Cost for a Service Invocation

0

2

4

6

8

10

12

14

16

18

20

1 2 3 4 5 6 7 8λ d/λ c

Invo

catio

n M

essa

ging

Cos

t (λ

c)

1C1D 1CnD

1CnD performs better, when data messages are much larger

than control messages

(Data Message Size / Control Message Size)

12

Response TimeResponse Time

Response time = Time consumed to execute a megaservice

• Distributed data-flow model performs better ifCMki ≥ CM0i for all k ≠ 0 and i ≠ 0

MP_c SI3_d SP3_d SO3_d

MP_f

SI2_b SP2_b SO2_b

SI1_a SP1_a SO1_a

SI4_e SP4_e SO4_e

Access links of megaservice

Communication networkamong services

Time Marked Graph Representationof a Megaservice

13

Computing Networks and Integration ModelsComputing Networks and Integration Models

Service1

Service3

MegaService

(a) Internet and Corporate Intranet(Fit for distributed data-flow model)

Service2

Communication Backbone

Service1

Service3

MegaService

(b) Dedicated Service Environment(Fit for centralized data-flow model)

Service2

14

Summary of FindingsSummary of Findings

Distributed data-flow model is suited for coarse grain service integration

Performance optimization for megaservice• Establish direct data exchanges among services• Distribute computations to where data is located

System architecture• Improve the communication network among the services

for distributed data-flow model• Improve the access links of the megaservices for

centralized data-flow model

15

FICAS

16

FICASFICAS

Flow-based Infrastructure for Composing Autonomous Services

• Autonomous Services– Wrap legacy software applications– Provide an access protocol

• Buildtime Environment– Specify composition logic

• Runtime Environment– Coordinate service execution– Conduct performance optimization

17

Autonomous Services

18

Autonomous Service MetamodelAutonomous Service Metamodel

Data-flow

Control-flow

Input Data Container

Output Data Container

Input Event Queue

Output Event Q

ueue

Service Core(ASAP Support)

Megaservice C

ontroller

EncapsulatedSoftware

Application

AutonomousServiceWrapper

Service Core• Provide service functionalities• Wrap software applications

Two Data Containers• Handle I/O data• Enable distributed data-flows

Two Event Queues• Handle inqueries and issue requests• Support asynchronous invocations• Form control-flows

Megaservice Controller• Coordinate megaservice execution

19

Autonomous Service Access ProtocolAutonomous Service Access Protocol

ASAP• Light-weight, asynchronous and event-based• Define how autonomous services respond to events • Use XML as transport medium for both control and data

Events• SETUP: Initialize a service• TERMINATE: Terminate a service• INVOKE: Start execution of a service• MAPDATA: Establish a data-flow between two services• CONTROLFILE: Execute a megaservice

20

Autonomous Service WrapperAutonomous Service Wrapper

public interface ServiceCore{

public boolean setup (Container inc,Container outc,FlowId fid );

public boolean execute (Container inc,Container outc,FlowId fid );

public boolean terminate (Container inc, Container outc,FlowId fid );

}


initialize invoke terminate

EncapsulatedSoftware

Application

21

Buildtime Environment

22

Architecture of Buildtime EnvironmentArchitecture of Buildtime Environment

Executable

CompilationCLAS

Compiler

FICASControl

Sequences

JavaCompiler

CLASPrograms

MobileClass

SourceCodes

MobileClasses

Source

CompositionalSpecification

ComputationalSpecification

Composition• Invocation of services• Dependencies among services• Process flow of services

Computation• Processing of service data

23

CLASCLAS

Compositional Language for Autonomous Services• High-level and declarative• Based on CLAM developed in CHAIMS• Simple (for domain experts, NOT technical experts)• Separation between composition and computation

Features• Decomposition of a CALL statement into 4 primitives

– SETUP, INVOKE, EXTRACT, TERMINATE

• Control primitives– IF … THEN … ELSE– WHILE

24

Sample CLAS ProgramSample CLAS Program

SchedulingDemo http://ficas.stanford.edu/Megaprogram{

/* Setup Services */psl_svc = SETUP("SIPsl")p3_svc = SETUP("SIP3")notification_svc = SETUP("SINotification")

/* Invoke services */psl = psl_svc.INVOKE("to-psl", "CEIL")ceil = psl.EXTRACT()p3 = p3_svc.INVOKE("reschedule", ceil)ceil2 = p3.EXTRACT()oracle = psl_svc.INVOKE("to-oracle", ceil2)status = oracle.EXTRACT()IF (status == “SUCCESS”)THEN {

notif = notification_svc.INVOKE("171.64.55.32", 8250, status)}…

}

25

Mobile ClassMobile Class

Mobile Class• Java-based and reusable• Perform complex computations

Usage of Mobile Class• Arithmetic operation• Relational operation• Data aggregation and

abstraction• Type conversion

/* A mobile class for type conversion */

public class int2float implements MobileClass

{

public DataElement execute(Vector params) {

DataElement arg =

(DataElement) params.firstElement();

int val = arg.getIntValue();

return new DataElement().setValue(

new Double(val).doubleValue());

}

}

/* Using mobile class in a CLAS program */

floatnum = MCLASS("int2float", num)

26

Mobile Class for Type MediationMobile Class for Type Mediation

M

S1

(a) Type Brokers

T1

T2T1_T2 T2_T3

S2

T3

M

S1

(b) Type Mediation Mobile Classes

T3 S2

mobile classT1_T2

mobile classT2_T3

Control-flow

Data-flow

27

Runtime Environment

28

Architecture of Runtime EnvironmentArchitecture of Runtime Environment

Megaservice Controller

ServiceCore


ServiceCore

CommunicationNetwork

AutonomousService

Directory

FICASControl

Sequence

MobileClasses


ServiceCore

Autonomous Service Wrapper

From FICAS

Buildtime

29

Megaservice ControllerMegaservice Controller

ASAP Event Receiver

Outgoing Event Pool

FICASControl Sequence

Control Manager

FlowDependency

Table

VariableCache


Autonomous Service Wrapper

FromAutonomous

Services

Input Event Queue

Output Event Q

ueue



ServiceCore

Active Mediator

ToAutonomous

Services

FICAS Control Sequence

...

<INVOKE><INVOCATIONHANDLE>Invocation1</INVOCATIONHANDLE><SERVICEHANDLE>Service1</SERVICEHANDLE>

</INVOKE><INVOKE><INVOCATIONHANDLE>Invocation2</INVOCATIONHANDLE><SERVICEHANDLE>Service2</SERVICEHANDLE>

</INVOKE><EXTRACT><VARIABLE>A</VARIABLE><INVOCATIONHANDLE>Invocation1</INVOCATIONHANDLE>

</EXTRACT><EXTRACT><VARIABLE>B</VARIABLE><INVOCATIONHANDLE>Invocation2</INVOCATIONHANDLE>

</EXTRACT><INVOKE><INVOCATIONHANDLE>Invocation3</INVOCATIONHANDLE><SERVICEHANDLE>Service3</SERVICEHANDLE><VALUELIST><VARIABLE>A</VARIABLE><VARIABLE>B</VARIABLE>

</VALUELIST></INVOKE>

...

30

Extract Data Dependencies from MegaserviceExtract Data Dependencies from Megaservice

Invocation1 Invocation2

Invocation3

Invocation4

A B

C

D

Invocation1 = Service1.INVOKE();

Invocation2 = Service2.INVOKE();

A = Invocation1.EXTRACT();

B = Invocation2.EXTRACT();

Invocation3 = Service3.INVOKE(A, B);

C = Invocation3.EXTRACT();

Invocation4 = Service4.INVOKE(C)

D = Invocation4.EXTRACT();

31

Event Dependency Graph (1C1D)Event Dependency Graph (1C1D)

INVOKE(Service1)

INVOKE(Service2)

MAPDATA (A, Service1,Megaservice)

INVOKE(Service3)

MAPDATA(C, Service3,Megaservice)

MAPDATA(B, Service2,Megaservice)

INVOKE(Service4)

MAPDATA(D, Service4,Megaservice)

MAPDATA (A, Megaservice,

Service3)

MAPDATA (B, Megaservice,

Service3)

MAPDATA(C, Megaservice,

Service4)

Service 1

Service 2

Service 3

MegaService

Service 4

32

Event Dependency Graph (1CnD)Event Dependency Graph (1CnD)

INVOKE(Service1)

INVOKE(Service2)

MAPDATA (A, Service1,

Service3)

INVOKE(Service3)

MAPDATA(C, Service3,

Service4)

MAPDATA(B, Service2,

Service3)

INVOKE(Service4)

MAPDATA(D, Service4,Megaservice)

Service 1

Service 2

Service 3

MegaService

Service 4

33

Performance Evaluation Performance Evaluation –– SOAP vs. FICASSOAP vs. FICAS

SwitchMegeService

S1producesa string

10mbps

S2consumes

a string

out10mbps

in

(LAN) in = 10 mbps; out = 10 mbps (Wireless) in = 2 mbps; out = 0.5 mbps

MultiService { a = S1(size) S2(a) }

SingleService { a = S1(size) }

34

Performance in LAN SettingPerformance in LAN Setting

0

2000

4000

6000

8000

10000

12000

14000

16000

18000

0 100 200 400 800 1600 3200

Data Volume (KB)

Meg

aser

vice

Exe

cutio

n Ti

me

SOAP (SingleService) SOAP (MultiService) FICAS (MultiService)

FICAS incurs higher control-flow cost

SOAP incurs higher data-flow cost

35

Performance in Wireless SettingPerformance in Wireless Setting

0

10000

20000

30000

40000

50000

60000

70000

80000

0 100 200 400 800 1600 3200

Data Volume (KB)

Meg

aser

vice

Exe

cutio

n Ti

me

SOAP (LAN) SOAP (802.11b) FICAS (LAN) FICAS (802.11b)

SOAP creates bottleneck on the megaservice

communication link

FICAS is little affected since data-flows are distributed

36

Active MediatorActive Mediator


MobileClass

Fetcher

MobileClass

Runtime

ExceptionHandling

Mobile ClassCache

Mobile ClassRepository



MobileClassAPI

Library

Active Mediator

37

Example Megaservice Utilizing a Mobile ClassExample Megaservice Utilizing a Mobile Class

SwitchMegeService

S1producesa string

10mbps

S2consumes

a string

10mbps

Mobile ClassFILTER

10mbps

Invocation1 = S1.INVOKE(size) A = Invocation1.EXTRACT()

B = MCLASS ("FILTER", A)

Invocation2 = S2.INVOKE(B)

38

Placement of Mobile ClassPlacement of Mobile Class

S1

Megaservice

S2

S1

Megaservice

S2

mobile classFILTER S1

Megaservice

S2mobile classFILTER

(c) Using an autonomous service(a) Placing FILTER at S1 (b) Placing FILTER at S2

1

3

1

2

1

2

Data-flowServiceInvocation

A

B

B A FILTER2

39

Performance Comparison for Mobile ClassPerformance Comparison for Mobile Class

0

2000

4000

6000

8000

10000

12000

100 200 400 800 1600 3200

Data Volume (KB)

Meg

aser

vice

Exe

cutio

n Ti

me

Mobile Class on S1 Mobile Class on S2 Utility Autonomous Service

Mobile class placement affects megaservice performance

Mobile class is more efficient than autonomous service

40

Infrastructure forEngineering Services

41

An Integrated Service EnvironmentAn Integrated Service Environment

Designers

Project Managers

On Site Personnel

CommunicationNetwork

Spreadsheets

Autonomous ServiceWrapper

ProjectPlanning

Tools


ModeilingTools


Mobile Classes forData Integration

IntegratedWork Processes

42

Data Mediation Among the ToolsData Mediation Among the Tools

PSLXML

4DViewer

ActiveMediator

Palm DesktopBrowser

PrimaveraP3

MicrosoftProject

RelationalData

MicrosoftExcel

Related work by Jim Cheng at Engineering Informatics Group, Stanford University

43

Review Design in 4D ViewerReview Design in 4D Viewer

44

Review Schedule in PrimaveraReview Schedule in Primavera

45

Review Schedule in Microsoft ProjectReview Schedule in Microsoft Project

46

View Schedule on SiteView Schedule on Site

http://med...!! HistorySCHEDULE

Review the schedule and make appropriate updates by changing thevalue in duration:

SCHEDULEIDSTARTDATEDURATION

18T1-3320101-31-2001 1.….………………….. Update

18T1-3324102-01-2001




18T1-3320101-31-2001 1..…………………….. Update

18T1-3324102-01-2001

47

Modifying Schedule OnModifying Schedule On--sitesite




18T1-3320101-31-2001 40…………………….. Update

18T1-3324102-01-2001




18T1-3320101-31-2001 40…………………….. Update

18T1-3324102-01-2001

Change duration of activity18T1-33201 (“Erect Roof Element 1”)From 1 day to 40 days

48

Invoke Rescheduling Megaservice Invoke Rescheduling Megaservice

Network

ReschedulingMegaService

RescheduleService { model = PSLModel(modelname, 'oracle-to-psl') new_model = P3Scheduling(model, 'reschedule') PSLModel(new_model, 'psl-to-oracle') excel_data = MCLASS("psltoexcel", new_model) ExcelService(excel_data, 'show') ChangeNotify(modelname) }

PSLPSL

ModelService

P3Scheduling

Service

ChangeNotify

Service

ExcelService

Primavera

49

Review Modified Schedule in ExcelReview Modified Schedule in Excel

Actual Change

Affected activity 18T1-33201

50

Review Changed Activities in BrowserReview Changed Activities in Browser

Actual change

Affected Activities

51

Review Modified Schedule in PrimaveraReview Modified Schedule in Primavera

Updated activity

52

Review Modified Design in 4D ViewerReview Modified Design in 4D Viewer

Roof constructionis delayed

53

Review Modified Schedule in ProjectReview Modified Schedule in Project

Actual change

Affected activities

54

Summary

55

ContributionsContributions

• Data-flow distribution improves megaservice performance

• Distributed data-flow model is supported in service composition

– Separate data from controls in services

– Separate computation from composition

– Establish direct data communications among services

• Distribution of computations facilities service composition

– Mobile class allows performance optimization

– Active mediation enhances the flexibility of services

• FICAS provides comprehensive support for service composition

56

PublicationsPublications

D. Liu, J. Cheng, K. H. Law, G. Wiederhold, and R. D. Sriram. "An Engineering Information Service Infrastructure for Ubiquitous Computing", Journal of Computing in Civil Engineering, 2003.

D. Liu, J. Cheng, K. H. Law, and G. Wiederhold. "Ubiquitous Computing Environment for Project Management Services", Proceedings of the Civil Engineering Conference and Exposition, Washington DC, 2002.

D. Liu, K. H. Law, and G. Wiederhold. "Data-flow Distribution in FICAS Service Composition Infrastructure", Proceedings of the 15th International Conference on Parallel andDistributed Computing Systems, Louisville, KY, 2002.

D. Liu, K. H. Law, and G. Wiederhold. "Analysis of Integration Models for Service Composition", Proceedings of the 3rd International Workshop on Software and Performance, Rome, Italy, 2002.

D. Liu, K. H. Law, and G. Wiederhold. "CHAOS: An Active Security Mediation System", Proceedings of the International Conference on Advanced Information Systems Engineering, pp. 232-246, 2000.

D. Liu, J. Peng, K. H. Law, G. Wiederhold, and R. D. Sriram. “Composition of Autonomous Services with Distributed Data Flows and Computations", Submitted to ACM Transactions on Internet Technology, 2003.

J. Peng, D. Liu, and K. H. Law. "An Engineering Data Access System for a Finite Element Program", Journal of Advances in Engineering Software, vol. 34(3), pp. 163-181, 2003.

57

AcknowledgementsAcknowledgements

• Defense committee

• Members of Engineering Informatics Group

• Research support– Dr. Ram D. Sriram, NIST– Dr. Charles S. Han, Autodesk– Mr. Jim Cheng, Stanford University– Prof. Martin Fisher and his research group, Stanford University

• This work is supported in part by– Air Force (Grant F49620-97-1-0339, Grant F30602-00-2-0594)– CIFE, Stanford University– Product Engineering Program at NIST– Technology for Education 2000 Equipment Grant, Intel Corporation

58

End of PresentationEnd of Presentation

A Distributed Data Flow Model for Composing Software...

Documents

Transcript of A Distributed Data Flow Model for Composing Software...