Export (0) Print
Expand All
Separating DSL Semantics from Implementation
Expand Minimize

Domain-Specific Modeling Languages: Moving from Writing Code to Generating It

Cc168592.skyscrapr_banner(en-us,MSDN.10).gif

Steven Kelly

December 2007

Summary: The purpose of this article is to help you generate full production code from precise, high-level models in your own modeling language. (6 printed pages)

Contents

Introduction
If You Have the Code, Why Generate It?
When and Where Is DSM Applicable?
Tool Support for DSM
Why Is DSM Different?
Conclusion
Critical-Thinking Questions
Sources
Glossary

 

Introduction

I always remember the day when my bluff was called. I had just been handed 50 pages of Java code and was told, "Generate that!" I had always said that if you gave me a sufficiently precise model of a system, and the code that implements that system, I could build a code generator that would produce that system from the model.

While I normally prefer to work on code in a text editor, this task clearly called for more scalable weapons from the architect's tool chest: hard copy, scissors, sticky tape, and the full palette of highlighter pens. Shutting the door, I set to work at analyzing the code to find repeated sections, common blocks in which just a few elements changed, and larger scale patterns. These then had to be crosschecked with the graphical model of the system, to find which parts of the model determined them. Several hours—and more than a few shouts of, "But that's just the same as that!"—later, I figured that I had it nailed. In a couple of more hours, I had implemented a generator that crawled the structures in the model to spit out the corresponding blocks of code—replacing the changing elements with values from the models.

Running the generator the first time produced... exactly nothing. After fixing the offending minor bug, I got to work on its bigger brothers: comparing the generated output with the code that I had been given. After an hour or two, I had it down to about five or six spots in which I just could not figure out why the code was what it was. Clearly, a given situation sometimes would require one kind of code; but, in one or two spots, that code was either different or not present. I sent the results off to the customer—asking for clarification of what improvements were needed to the modeling language, to capture the differences between the situations.

The next day was scheduled for implementing those changes to the modeling language, and updating the modeling language and generator. When I got the customer's reply the next morning, I realized that, for once, I would go home well-rested. The generator was already finished: Every single one of the differences was a mistake in the hand-written Java code.

If You Have the Code, Why Generate It?

The background to this story was a German company that wanted to build an e-commerce site on which their customers could compare insurance policies. The server would understand the policies sufficiently enough to be able to calculate things such as costs and benefits. Their problem was that this effectively required "implementing" the policies in Java, so that the Web site could "run" them for comparison. The Java to do this was nontrivial, and the policies themselves were written in dense legalese that only the insurance experts could understand. And the insurance experts did not happen to be Java programmers.

Because the final system would have hundreds of such policies, each with a mass of code similar to the preceding story, something had to be done. In effect, there was too large a disconnect between the level of abstraction of the problem and the level of abstraction that Java offered. If this were something that had to be faced only once, in a small section of the application, it might have been acceptable. But here, there clearly had to be a better way: something at a higher-level of abstraction than Java, closer to the language that was used by the domain experts, and yet with the precision to enable us to generate code.

This is, in fact, a problem that is faced by any company that is building a range or family of applications in the same domain. How can we raise the level of abstraction beyond today's third-generation programming languages? One way is to move away from trying to specify all kinds of applications by using only one generic set of program-language concepts. Instead, each company could use the concepts of the problem domain in which it works—giving each concept its own visual representation.

Symbols in such a domain-specific modeling language (DSM language), as well as the rules about how they can be connected and used, would thus come directly from the world in which the application is to run. This is a whole level of abstraction higher than UML models. It results in a very expressive, yet bounded, design language that can only specify applications in the problem domain that it was designed to cover. While it might be useless elsewhere, it is both concise and precise in its own domain.

Together with the insurance experts, we thus built a DSM language for describing insurance policies. It had symbols for concepts such as "risk," "subpolicy," "insured," "accident," and "payout calculation," as well as relationships that connected these to show how the policy was made up, and who paid what under which conditions. Figure 1 shows an example from a top-level diagram.

Click here for larger image

Figure 1. An example diagram in the domain-specific modeling language for insurance (Click on the picture for a larger image)

Based on one policy's handwritten code from their Java programmers, we built the code generator that turned this modeled information about the static structure and dynamic behavior of the product into J2EE Java code. Now, the insurance experts could easily describe a policy in the new modeling language, and the appropriate code could be generated instantly. New policies could be added to the system up to five times faster, and with fewer errors than with manual coding.

When and Where Is DSM Applicable?

Domain-specific modeling requires a DSM language and a matching code generator. Although defining these is much easier than defining a generic modeling language, it is still not a task for every developer. It requires a good knowledge of the problem domain and the code that is written for it. This domain expertise is usually found in situations in which we deal with product-family development, continuous development of the same system, or configuration of a system to customer-specific requirements (phone-routing systems, CRM, Workflow, payment systems, and so on).

Experienced developers who work in these types of domains know what are the concepts of their domain, are familiar with the rules that constrain them, and know how code or configuration files should be written best. Sometimes, it is one expert who possesses this knowledge; sometimes, the knowledge is shared among a small group of people who have different skills, as in the preceding example. Whether it is one person or several, the expertise that they possess makes them well-qualified to define the automation mechanism that will make the rest of the developers much more productive.

Benefits of DSM

DSM requires both an investment of development resources and time to set up the DSM environment. Although this might clash with the urgency of current development projects, the improved productivity is usually worth it. Industrial experiences of DSM consistently report productivity being between 5 and 10 times higher than with current development approaches. Reported examples include Nokia's cellular phones, Lucent's 5ESS telecommunications switch systems, Microsoft's Whitehorse tools for SOA, and EADS Tetra terminals.

The last time that our industry saw such productivity gains was with the move from assembly language to third-generation languages (GLs) such as FORTRAN or BASIC: That produced a leap in productivity of 450 percent. The 40 years since then have had little effect: From BASIC to Java is an improvement of just 20 percent [Software Productivity Research, 2005]. This is not so surprising, if you consider that all three GLs are at roughly the same level of abstraction.

Traditional modeling languages such as UML have not increased productivity, because the core models from which code can be generated are on the same level of abstraction as the programming languages that are supported. Current model-driven architecture (MDA) tools are based on UML and, thus, suffer from its problems. Even a tool-vendor–sponsored study found only an increase in productivity of 35 percent, compared with hand coding [TheServerSide, 2003].

In addition to productivity gains, it is surely better that the expert formalize development practices once, and that other developers' generated code automatically follow them, instead of having all developers try to follow best practices manually all the time. All architects will be familiar with how well their wonderful guidelines and standards documents are actually read and followed by developers! If the rules and guidelines are embedded in a tool, it can enforce the rules and guide or warn developers appropriately.

DSM models capture the information about the actual system and its behavior, instead of its implementation, with a certain programming language and set of libraries. This means that the exact, same models can be used to generate code for a different programming language or libraries: All that is needed is for one developer to write the generator. Flexibility such as this significantly reduces the risk of obsolescence and the costs of migration to different technology. On a smaller scale, code changes that are required by newer versions of the same technology can also be made by one person in the generator, instead of by every developer in every piece of code.

Tool Support for DSM

A software-development method is of little value if it has no tool support. Designs in a DSM language on a whiteboard do not generate any code; thus, while they help in defining the functionality of the application or system, they do not improve productivity. DSM has been around for quite some time, but companies that chose to adopt it often had to resort to building their own environments to support their DSM language and generators. The huge investment in person-years that is required to do this often led to the decision to look for other ways in which to improve productivity.

While a DSM language and corresponding code represent the core competency of the company that makes them, in most cases, generic modeling-tool functionality is not something in which the company has experience. Also, the basic behavior of a modeling tool generally remains the same, irrespective of the specific modeling language. It is, therefore, possible for a company that has modeling-tool experience to make a generic modeling tool, which can be customized appropriately by each organization that is building a DSM language.

In early examples of this kind, the customization was more a case of writing your own code that ran on top of the vendor's framework. Today's more advanced tools make it possible to build your own modeling language without writing any code: The modeling-language definition is put in as data through graphical models and forms, and the generic modeling tool configures itself automatically from that data. This has the significant benefit that any changes to the modeling language can be instantly reflected in the modeling tool—even updating existing models.

Most existing DSM takes place with DSM environments—either commercial, such as MetaCase's MetaEdit+, or academic, such as Vanderbilt University's GME. The increasing popularity of DSM has led to DSM frameworks being added to existing IDEs—for instance, Eclipse's EMF and GMF, or Microsoft's DSL Tools for software factories [Greenfield and Short, 2004]. A list of tools is available in the industry forum [DSM Forum, 2007] at http://www.dsmforum.org/tools.html.

Why Is DSM Different?

As few domains are static, the ability of DSM environments to let you evolve the modeling language is vital. This ability is something that separates DSM from earlier attempts to generate full code directly from models. In the computer-aided software-engineering (CASE) tools of the 1980s, both modeling language and code generator were provided by the vendor. If the modeling language did not fit your way of describing applications, you had to change; if the code was not good for your needs, you were stuck. In UML, the modeling language is fixed by committee and the generator by vendors. In DSM, the modeling language is designed by your best domain expert, the generator is crafted by your best developer; and, if either of them needs changing or fixing, a good DSM environment will let you do it straightaway—updating existing models.

Conclusion

DSM allows an organization to make major improvements in productivity, where developers would normally be forced to resort to copy-paste reuse of code snippets. A top developer in the company creates a new graphical modeling language whose concepts map to the company's problem domain, and a generator that maps these concepts into the code that the developer would teach others to write. The other developers build models with the new language, and full code is automatically generated. Today's DSM environments make creating and using such languages and generators accessible to all sizes of project.

Critical-Thinking Questions

· Where are your developers resorting to copy-paste reuse?

· How much better is the code from the top 20 percent of your developers than from the bottom 40 percent?

· How much time is spent on problems stemming from poor communication between domain experts and developers?

Sources

· [DSM Forum, 2007] "DSM Tools." DSM Forum Web site. 2007. http://www.dsmforum.org/tools.html (Accessed January 10, 2007.)

· [Greenfield and Short, 2004] Greenfield, Jack, and Keith Short. Software Factories: Assembling Applications with Patterns, Models, Frameworks, and Tools. Indianapolis , IN: Wiley Publishing, 2004.

· [Software Productivity Research, 2005] "Programming Languages Table(TM) (PLT2005)." Software Productivity Research Web site. 2005. http://www.spr.com

· [TheServerSide, 2003] "Enterprise Java Community: TheServerSide Symposium Coverage June 2003." TheServerSide Web site. (See "Productivity Analysis: Model-Driven, Pattern-Based Development with OptimalJ.") 2003. http://www.theserverside.com/tt/articles/article.tss?l=SymposiumCoverage (Accessed January 10, 2007.)

Glossary

DSM—Domain-specific modeling. Creating and using modeling languages that are targeted for a narrow problem domain; often, to generate full code from the models.

MDA—Model-driven architecture. The Object Management Group's approach to model-driven development, relying largely on UML.

Software factories—Microsoft's approach to automating development, using DSM in a central role.

UML—Unified Modeling Language. The Object Management Group's set of general-purpose modeling languages for any problem domain, but focused on implementation with an object-oriented programming language.

About the author

Dr. Steven Kelly is the CTO of MetaCase and cofounder of the DSM Forum. He has over 10 years of experience with building metamodeling and modeling environments, as well as acting as a consultant on their use in domain-specific modeling. Dr. Kelly writes and speaks frequently in major industry venues and journals, such as SD Best Practices, OOPSLA, and Dr. Dobbs.

 

This article was published in Skyscrapr, an online resource provided by Microsoft. To learn more about architecture and the architectural perspective, please visit skyscrapr.net.

 

Show:
© 2014 Microsoft