Keynote at Code Generation 2014: The business cases of modeling and generators

Post on 27-Aug-2014

1.983 views 2 download

Tags:

description

There is no business case for modelling and generators – only for a specific language and generator in a specific situation. The right language in the right situation can improve productivity by an order of magnitude; the wrong language will reduce it. So what makes a language “right”? In this talk we look what kinds of modelling languages and generators tend to be more beneficial than others –from the early days of modelling right up to the latest research. Getting more out of the models than their original creation required, raising the level of abstraction and addressing a specific need seem to be common characteristics of successful modelling and code generation approaches. To evaluate the full business case we also need to look at the costs and benefits of creating languages, generators and tools. By applying the above principles to language creation itself, we can improve the quality of the resulting modelling language, whilst lowering the time and effort needed to create and maintain it. Other features important to the language developer include enabling tools to support language evolution, and improving tool scalability to tackle larger systems and teams.

Transcript of Keynote at Code Generation 2014: The business cases of modeling and generators

12 April 2014

Juha-Pekka Tolvanen

The business cases for modeling and generators

Various styles of modeling

MDD done correctly

Work at a higher level of abstraction and automatically produce code, library calls, configuration, simulation...

Specify less, generate more!

Gains

“The work that earlier took a whole week, is now done during an afternoon”, an engineer at Elektrobit

The measurement (n=6) revealed that Domain-Specific Modeling solution was on average 10 times as productive as the current development approach, study1 at Polar

Productivity improvement is 289%, study2 at USAF

– Over 130 tasks, both initial development and maintenance

– The differences in the average performance of the subjects are statistically significant at confidence levels exceeding 99%

1) Kärnä, J., et al., Evaluating the Use of Domain-Specific Modeling in Practice, Workshop on DSM, 2007

2) Kieburtz, R., et al. A software engineering experiment in software component generation, ICSE, 1996

MDD research

Recent empirical research in industry practices found that:

success with modeling is most likely when companies develop their own bespoke modeling approaches, languages and tools rather than simply applying off-the-shelf solutions

1) Hutchinson, J., Whittle, J., Rouncefield, M. and Kristoffersen, S., “Empirical Assessment of MDE in Industry,” 2011 International Conference on Software Engineering (ICSE), pp.471-480.

2) Hutchinson, J., Rouncefield, M. and Whittle, J., “Model Driven Engineering Practices in Industry,” 2011 International Conference on Software Engineering (ICSE), pp. 633-642.

Number of railway languages?

1) Mark Blackburn, Peter Denno, Virtual Design and Verification of Cyber-physical Systems: Industrial Process Plant Design, Procedia Computer Science, Volume 28, 2014

Detected productivity increase

* Productivity proportional to earlier practice

500 %

300 %

750 %

600 %

200 %

750 %

500 %

600 %

0 % 100 % 200 % 300 % 400 % 500 % 600 % 700 % 800 %

Embedded UI applications

Message translation &

validation

Phone switch features

Call processing services

Adaptive cruise control

Heart rate monitor

J2EE web application

Home automation

Domain

Percent Increase

PROCESS abc;

START;

TASK b_subscriber := subscriber_t_find (b_subnum);

DECISION b_subscriber.bill_paid AND reserve_vc (b_vc);

(T):

OUTPUT initiate_answering_s (b_subscriber);

NEXTSTATE works_done;

(F):

OUTPUT setup_nak_s TO cci_party;

STOP;

ENDDECISION;

STATE works_done;

INPUT ringing_s;

TASK sso_party := SENDER;

OUTPUT setup_ack_s TO cci_party;

NEXTSTATE waiting_off_hook_s;

INPUT input_busy_s;

OUTPUT disconnect_fwd_req_s TO cci_party;

NEXTSTATE waiting_disconnect_fwd_ack_s;

ENDSTATE wait_ringing;

STATE waiting_disconnect_fwd_ack_s;

INPUT disconnect_fwd_ack_s;

STOP;

ENDSTATE waiting_disconnect_fwd_ack_s;

STATE waiting_off_hook_s;

INPUT off_hook_s;

...

Who creates and maintains them?

1. Standardization bodies:

– ISO, OMG, etc.

2. Industry associations and groups:

– ITU, SAE, AUTOSAR, EAST-ADL association, etc.

3. Tool vendors:

– Labview, Simulink, Mendix, etc.

4. In-house:

– Language engineers, tool-smiths, generator developers

6 ways to get the tools needed

1. Find an existing and suitable tool

2. Write own tool from scratch

3. Write own tool based on frameworks

4. Metamodel, generate tool skeleton, add code for the rest

5. Metamodel, generate full tool over a framework

6. Metamodel, configure an existing tool with own language & generator

Use tools to get your tool! – Meta-tools, CASE shells, metasystems, metaCASE tools, DSL

tools, language workbenches

PSL/PSA

Defining languages with:

– Objects

– Properties

– Relationships

1 PROCESS hourly-enployee-procesing;

2 /* DATE OF IAST CHIANGE - JUN 26, 1976, 13:56:44 */

3 DESCRIPTIONS;

4 this process performs those actions needed to interpret

5 time cards to produce a pay statement for each hourly

6 employee.;

7 KEYWORDS: independent;

8 ATTRIBUTES ARE:

9 complexity-level

10 high;

11 GENERATES: pay-statement, error-listing,

12 hourly-employee-report;

13 RECEIVES: time-card;

14 SUBPARTS ARE: hourly-paycheck-validation, hourly-emp-update,

15 h-report-entry-generation,

16 hourly-paycheck-production;

17 PART OF: payroll-processing;

18 DERIVES: pay-statement

19 USING: time-card, hourly-employee-record;

20 DERIVES: hourly-employee-report

21 USING: time-card, hourly-employee-record;

22 DERIVES: error-listing

23 USING: time-card, hourly-employee-record;

24 PROCEDURE;

25 1. compute gross pay from time card data.

26 2. compute tax, from gross pay.

27 3. subtract tax from gross pay to obtain net pay.

28 4. update hourly employee record accordingly.

29 5. update department record accordingly.

30 6. generate paycheck.

31 note: if status code specifies that the employee did not work

32 this week, no processing will be done for this employee.;

33 HAPPENS:

34 number-of-payments TIMES-PER pay-period;

35 TRIGGERED BY: hourly-emp-processing-event;

36 TERMINATION-CAUSES:

37 new-employee-processing-event;

38 SECURITY IS: company-only;

Teichroew, D., Hershey, E., PSL/PSA, IEEE Trans. on Software Engineering, Vol SE-3 , 1, 1977

Eclipse TBK

Three parts in modeling tool definition:

– Types

• Node and link

– Representation

• Symbols containing shapes

– UI

• Menus, selectors…

MacRoibeaird, S., Developing & Integrating Tools In Eclipse/PCTE, 1990

MetaEdit: abstract & concrete syntax

MetaEdit Personal, MetaCase, 1991

Quality of the language depends on its definition!

Definition of UML contains rules for elements that have been removed several versions ago - 5 years ago

Fragmented definition

• Concepts in metamodel

• Rules in constraint language

• Notation in symbol definitions

• Transformations in code, …

• Tool functions in code, .ico …

Integrated definition

• Change in one place updates it everywhere:

• in rules and constraints,

• in symbols,

• in generators,

• toolbars, icons, etc.

361 errors in UML 2.0: Bauerdick et al, in Procs of UML 2004, LNCS 3273, Springer, 2004

320 errors in UML 2.3: Wilke & Demuth, 2010, journal.ub.tu-berlin.de/eceasst/article/download/669/682‎

Quality of resulting language (& tool)

Incremental language development

Nobody gets the language right at the first try!

The best way to build languages is in incremental steps together with language users

• Language defined disconnected from the users

• Language defined as a spec on paper

• Partial, like plain schema

• Language changed without considering work done

• Language users are directly involved

• Each “unit” of the language tried out in the real world

• Influence of language change checked against other parts of the language

Effort to define language (& tool)

Two insurance-specific languages: “Insurance experts specify (with a tool) insurance products while at the backend the models are exported to a runtime execution environment”

Case from CodeGen’11 talk Case in www.dsmbook.com

PropertySpec:

/ Renewal DateRoleSpec:

Named Driver

RuleSpec:

Minimum

Insured Age

Calculation:

Calculate

Renewal

Date

Product:

FleetPlan

Product :

FleetN-N

Product :

FleetN-N

Big variety in the effort to define DSL:

– Insurance DSL in CodeGen’11: 6 calendar months, 7 persons

– Insurance DSL in dsmbook.com: 11 days, 1 person

Fully-loaded annual pay rate: 80.000 USD

– Cost varies a lot based on organizations and countries

Case from CodeGen’11 talk

Calendar months: 6

Persons: 7

Investment in time: 3½ years

Investment in $: 280.000

Case in www.dsmbook.com

Calendar months: 0,5

Persons: 1

Investment in time: 11 days

Investment in $: 3.520

Effort to define language (& tool)

Empirical tool comparison*:

– Same language, different tools

* El Kouhen et al. Evaluation of Modeling Tools Adaptation, 2012, http://hal.archives-ouvertes.fr/docs/00/70/68/41/PDF/Evaluation_of_Modeling_Tools_Adaptation.pdf

Effort to get tool support

Days to implement

BPMN language

From products to projects

Product development

Apply many times

– Several apps/features

Maintenance

Predictability

Initial investment shared over time/ products/variants

Projects

Apply mainly for the current case

Next customer?

A single project does not usually have a budget for language and generator development

To apply languages and generators in project organizations…

Build them fast

Accept uncertainty on the language constructs

Be ready to change frequently

Be ready to throw away, and start another

Allow language users to participate

Language Workbenches must support project organizations too!

Time needed to create languages and generators must be short

63 language concepts XML

generator

60 language concepts

C, HTML, build script generators

36 language concepts

Assembler generator

77 language concepts

Python generator

Java generator for simulation

143 language concepts J2EE

generator

Man days

MetaEdit Personal, MetaCase, 1991

Evolution (of language & models)

Maintenance is the biggest part in development!

Insurance DSL in CodeGen 2011: “took 5 months to update models when language evolved”

Language update scenario 1

1. Change language

2. Change tooling

3. Share/package updates

4. Install updates for all...

5. Update existing models

Language update scenario 2

1. Change language

– shared automatically

– automatic tool update

– models update automatically

Business case in terms of money

5 months of work:

– By one person: 33.000 USD (or whole team: 233.000 USD)

– What the other language users are doing meanwhile?

Missed opportunity: 5 months later is often bigger cost than 5 months of work!

Calendar months: 5

Investment in $:

33.000 - 233.000

Calendar months: 0

Investment in $: 0

Generator development process

Generator developer needs to master many things: metamodel (&models), generator language, target language/libraries

Fragmented:

• Metamodel in X(SD)

• Model in X(ML)

• Generator in X(SLT)

• Output in .x files

Integrated: • Access metamodel from

generator definition • Get from generated

code to source models • Access models while

debugging generators

Generator speed

A company visiting CodeGen few years ago was using 4-core Pentium machine to run their generator for a nightly build... and nights became too short!

Big differences between tools and generators

Comparison with an example*:

• Read model

• Read metamodel

• Use temporary model storage/M2M

• + execution

Cuadrado & Molina, Building Domain-Specific Languages for Model- Driven Development, IEEE Software, 2007, http://doi.ieeecomputersociety.org/10.1109/MS.2007.135 & Kelly, S., blog: http://www.metacase.com/blogs/stevek/blogView?entry=3385914921

Scalability, Collaboration

Development is about collaboration!

MDD is more than a few diagrams, it can be gigabytes of models

Single-user tool

• 1 (XML) file

• 1 person edits or…

• diff & merge done later

• Opening a larger model may take minutes...

Collaboration tool

• Repository with projects

• Collaborative editing

• No need to diff & merge

• Handle billions of elements, use lazy loading, etc.

MDD solution created correctly

Work at higher level of abstraction during language creation too!

Define only the needed parts, get rest automatically

Special attention to:

– Improving quality of language definitions

– Incremental development approach, user participation

– Expect (continuous) change in the domain

– Effort and costs of language development

– Generator development process and generator speed

– Scalability of the tools

Thank you!

For references, contact jpt@metacase.com * Some of the examples shown are available to download at www.metacase.com

Europe: MetaCase

Ylistönmäentie 31 FI-40500 Jyväskylä, Finland

Phone +358 14 641 000 Fax +358 420 648 606

USA: MetaCase

5605 North MacArthur Blvd. 11th Floor, Irving, Texas 75038

Phone (972) 819-2039 Fax (480) 247-5501