View on GitHub


Structural Model Subtyping with OCL Constraints

This document describes how to install and use the tool supporting the results of the paper Structural Model Subtyping with OCL Constraints by Artur Boronat (University of Leicester, UK), accepted for publication at SLE@SPLASH’17.


Our tool facilitates the reuse of model management operations that are defined for metamodel specifications, which consist of an EMF metamodel together with its well-formedness constraints (in OCL). Specifically, our tool infers whether two metamodel specifications, whose metamodels and constraints need not be related a priori, are compatible. This captures the notion of subtype polymorphism in model management operations using structural model subtyping with optional OCL constraints. Our tool implements the structural subtyping relationship in a type-theoretic framework with multiple inheritance semantics and can be used for implementing various use cases in flexible model-driven engineering, for example: structural (semantic) refinement, evolution of domain-specific modelling languages, and reuse of model management operations.

To deal with structural subtyping, the tool synthesizes an extension metamodel that is employed both to check the compatibility of OCL constraints and to reuse model management operations without requiring manual intervention. The extension metamodel can be obtained both for single and multiple inheritance contexts. Moreover, the tool produces complements of the supertype and subtype metamodel, w.r.t. the chosen subtyping relation, which can be used for guiding the reuse model management operations. In addition, our structural subtyping mechanism is expressive enough so as to support variants of model subtyping, including multiple, partial, and dynamic model subtyping.

Our tool is available as a Java library together with a self-contained Gradle project that includes all the dependencies required for using the tool together with additional resources implementing the examples of the paper. The tool has been tested on macOS 10.12.4 (Sierra) and Linux Ubuntu 15.04 (Vivid Vervet). In addition, a VirtualBox image with the Linux configuration is available for download.

This document contains three main sections: an overview of the contents in the zipped file, an installation guide, and step-by-step instructions arranged in the form of specific scenarios for each of the use cases mentioned above. These step-by-step instructions illustrate the core contributions of the paper and show the full extent of the examples used in it. Moreover, these scenarios have been implemented in executable test cases in order to ensure the experimental reproducibility of the results of the paper.

Overview: content of the archive

The contents of the archive are as follows:

Setup/installation guide

Environment assumptions to use the tool

The tool has been tested on macOS 10.12.4 (Sierra) and Linux Ubuntu 15.04 (Vivid Vervet). A VirtualBox image with Linux Ubuntu 15.04 is available, as explained below.

The base technology that is required to run the tool is as follows:

Although the tool is usable in Eclipse-independent standalone mode, we recommend the use of Eclipse Modeling Tools for editing the models. To work with Eclipse Oxygen (Modelling Package), the following plugins need to be installed:

Getting the tool

  1. Get the tool
    • From GitHub: clone the git repository:
    • From the zipped file: you should have unzipped the file to get this README file already.
  2. Import the project with test cases:
    • From Java perspective: Import > Gradle project
    • Select the project folder subtyping.tests in your git repository.
    • If prompted Overwrite existing Eclipse project descriptors?, choose Overwrite.
    • Use the option Gradle wrapper.
  3. Import the ATL projects using Import > Import Existing Projects into Workspace.

By default, the tool is configured to run the examples for macOS.

To run the tool on Linux, some configuration arguments need to be edited in the project subtyping.tests:

After that explore the test cases under src/test/groovy, which can be run as JUnit test cases. Follow the step-by-step instructions given below.

Linux (VirtualBox)

A VirtualBox image with Linux Ubuntu 15.04 (Vivid Vervet) and Eclipse configured to run the test cases is available for download.

  1. Download and install VirtualBox.
  2. Download the VirtualBox image Ubuntu 15.04_SLE17.ova from Google Drive.
  3. Import the appliance in VirtualBox.
  4. In VirtualBox, right click on image and Start > Normal Start.
  5. Use the following credentials (User/password): ubuntu/reverse.
  6. There is a direct link to Eclipse Oxygen on the desktop. Run Eclipse.
  7. Follow the step-by-step instructions given below.

Step-by-step instructions

In the following subsections, we provide the examples used to illustrate the core contributions in the paper.

Using the tool programmatically

The usual structure of a test exercising the subtyping operation is as follows:

The tool will determine whether (sMMPath,sOcl) denotes a model subtype of the model type denoted by (tMMPath,tOcl). This use case is illustrated in more detail below, in Scenario 1: Structural subtyping (Subtyping, Structural Refinement) and in Scenario 2: Structural subtyping with OCL constraints (Semantic Refinement, Multiple Typing). Note that any of the sets of OCL constraints may be empty (with the empty string "").

If the check fails, there are two main sources of incompatibilities: the model types denoted by the metamodels, and the OCL constraints.

  1. In the first case, the tool points at the source of the problem by showing the classes of the supertype metamodel tMMPath that are not extended by classes of sMMPath in the supertype complement bindingX_supertype_uncovered.ecore. That information is useful to assess the advantage of, for example, prunning the supertype metamodel by computing the effective metamodel w.r.t. a specific model management operation, as illustrated in Scenario 3: DSML evolution (Reuse, Partial Typing).
  2. In the second case, the tool will provide evidence that contradicts the compatibility property of sOcl w.r.t. tOcl in the form of a model conforming to the synthesized extension metamodel bindingX_virtual_mm.ecore, represented in EMF notation (that is in XMI format), that invalidates a constraint in tOcl. Note that the solver needs to be configured appropriately, in the configuration file as explained above, for increasing the likelihood of finding a problem if it exists.

If the check succeeds, the tool guarantees that (sMMPath,sOcl) is a structural refinement of (tMMPath,tOcl). Hence, any EMF model management operation that is defined for (sMMPath,sOcl) can be safely applied to models of (tMMPath,tOcl). Going one step further, the tool also facilitates the reuse of such operation by automatically synthesizing an extension metamodel bindingX_reuse_mm.ecore that can be substituted for tMMPath in the signature of the operation ensuring its application to models conforming to (sMMPath,sOcl) without any further change. This use case is illustrated with two scenarios: Scenario 3: DSML evolution (Reuse, Partial Typing) and Scenario 4: Stepwise Simulation of Deterministic State Machines (Reuse, Dynamic Typing).

The scenarios below illustrate specific examples that have been used to support the results in the paper. These scenarios have been implemented in test cases for ensuring the reproducibility of the results.

Note: for inspecting the generated models using model editors in Eclipse, such as the EMF reflective editor or Exeed, the corresponding metamodel needs to be registered using Epsilon’s Register EPackages facility in advance. This option appears when right-clicking on a metamodel .ecore in Eclipse. The reason for this is that synthesized metamodels reuse information from the original metamodels and the EMF registry needs to be updated to fetch the corresponding metamodel when loading a model.

Scenario 1: Structural subtyping (Subtyping, Structural Refinement)

In this section, we show the expressivity of our structural subtyping operation (without OCL constraints) with respect to model typing 1 2. We compare our approach to model subtyping by considering their example with the following state machine metamodels:

where changes have been highlighted.

The test cases implementing the subtyping checks can be found here and the results are summarized in the following table:

subtypeOf sm1 sm2 sm3  sm4 sm5
sm1 true true false false false
sm2 false true false false false
sm3 true true true false false
sm4 true true false true false
sm5 true true false false true

These results are consistent with those presented in 1 but for the cases where many-bounded references in a supertype are constrained by a lower upper bound in the corresponding subtype. In our tool, those cases are valid.

Scenario 2: Structural subtyping with OCL constraints (Semantic Refinement, Multiple Typing)

In this section, we use the main example of the paper for illustrating how multiple typings can be applied to a metamodel at the classifier level, that is, a class of the subtype metamodel can be typed by more than one class in the supertype metamodel. The generalization of multiple typing at the metamodel level, where several metamodels can be used as supertype of the same subtype metamodel, is supported by providing an extension metamodel for each different pair of metamodel specifications. However, as each such extension metamodel is linked to a different pair of metamodel specifications, with the intention of reusing a model management operation in a given context, we restrict ourselves to an example with one single model management operation.

In the example, we are using the metamodel specifications depicted below for defining graphs (metamodel and OCL constraints) and deterministic state machines (metamodel and OCL constraints), resp. The model types described by both metamodels are structurally similar in that they both describe languages of graphs.

On the one hand, the top metamodel specification characterizes the graph of a function defined over nodes. On the other hand, the bottom metamodel specification characterizes deterministic state machines where transitions can be triggered by an event (indicated in the name attribute of the transition) or are triggerless, e.g. they are completion transitions.

These test cases show how to use the tool to check that the state machine metamodel specification denotes a model subtype of the one denoted by the graph metamodel:

Moreover, if we consider non-deterministic state machines by removing the OCL constraint defining the deterministic condition from the state machine metamodel specification as instructed in this test case (test_singleInheritance_isSubtypeOf_inconsistent) we obtain a non-deterministic state machine that does not satisfy the graph constraint, which is represented in object diagram notation as follows:

The counterexample in generated in the folder temp/model, where temp is the temporary folder specified in the corresponding file. The counterexample is a model conforming to the virtual metamodel and can be casted down to the subtype as explained in the sections below.

These test cases demonstrate the usage of the possible combinations of OCL constraints with metamodels for checking structural subtyping, showing that OCL constraints are optional in metamodel specifications.

Scenario 3: DSML evolution (Reuse, Partial Typing)

In this section, we are going to show how to reuse a model management operation - in this case, a model-to-text transformation with ATL - for a modified version of the state machine metamodel. In this scenario, we discuss how to use the tool to reuse an ATL model transformation defined for a metamodel version 1 for models of a metamodel version 2 when the metamodel version 1 is not exactly a supertype of the metamodel version 2. In addition, we illustrate that the ATL model transformation can be applied even if a model is only partially typed by the metamodel involved in the ATL transformation.

The steps involved in this scenario relating to our tool are fully automated in the test case test_evolution_scenario. The steps that require interaction with ATL need to be carried out manually though.

Initial state machine metamodel (version 1):

We have developed an ATL transformation that serializes a state machine conforming to the previous metamodel into the format proposed by Martin Fowler:

query SM2Text = sm!StateMachine.allInstances()

helper context sm!StateMachine def: compile() : String =
	'events\n' +
	self.edges->iterate(e; acc : String = '' | acc + '  ' + + '\n') +
	'end\n\n' +
	self.nodes->iterate(n; acc : String = '' | acc + n.compile() );
helper context sm!State def: compile() : String =
	'state ' + + '\n' +
	sm!StateMachine.allInstances()->first().edges->iterate(e; acc : String ='' |
		if ( then
			'  ' + + ' => ' + + '\n'
	) +

This operation maps the state machine:



state a
  a->b => b

state b

In an update of our DSL for state machines, a concept Event is added as an explicit class and the concept Observation is removed, producing a new metamodel (version 2):

with the following constraint, ensuring the consistency of event names:

context Transition
inv event_consistency:
not(self.event.oclIsUndefined) implies

For which we can define state machines as follows:

The questions that we address next are:

Our subtyping operation assists us in determining that version 2 of the metamodel together with the OCL constraint is a refinement of version 1 as shown in this test case, which is not due to the removal of the Observation concept. By looking at the generated binding file and at completement of the supertype metamodel (shown below), the modeller has information to find out the source of the problem.

To see if there exists a potential valid refinement for reusing the operation, the user can extract the effective metamodel for the ATL transformation, either:

Our tool provides a facility for prunning a metamodel given the features of interest. The computed effective metamodel is as follows:

The subtyping operation is used again to check whether the effective metamodel is a valid supertype for our metamodel (version 2), which is correct.

The subtyping operation also synthesizes:

The extension metamodel, shown above, can be used to rewrite the signature of the model management operation. In addition, as the subtyping operation had to apply some automatic renamings in order to avoid name clashes, we have to adapt the original model that conforms to version 2 to the extension metamodel as shown in the test case test_evolution_scenario. This operation retypes the objects in the original model according to the renamings inferred by the subtyping operation:

This model can be processed by the ATL transformation, after replacing the original metamodel with the synthesized extension metamodel. Note that the adaptation of the model is only mandatory when the set of class names in version 1 and the set of class names in version 2 are not disjoint.

In case the renamings applied to the subtype metamodel in the extension model make the object type names different from those in the original subtype metamodel, the tool facilitates an adaptation from the extended metamodel to the original metamodel as explained in the following scenario.

Scenario 4: Stepwise Simulation of Deterministic State Machines (Reuse, Dynamic Typing)

In this example, we are considering the reuse of an ATL transformation used in the paper for simulating deterministic state machines stepwise by using an ATL transformation that applies a function graph to a node. The difference with the evolution example is that the simulation example generates new elements in the transformation, which are automatically re-typed in our approach, illustrating how dynamic typing is supported.

The steps in the scenario described below have been implemented in the test case test_simulation_scenario. The resources used in the scenario are reachable from the implementation of the test case.

In these models, we have obliterated the root object Graph.

As discussed in the paper, the ATL transformation could be applicable to deterministic state machines to facilitate their simulation. The state machine metamodel that we are considering is

with the well-formedness constraint

	context Transition
	inv determinism:
	not(Transition.allInstances()->exists(t | 

To reuse the ATL transformation, we use the subtyping operation to obtain the reuse metamodel:

In addition, we can check that the complement of the supertype metamodel only contains datatypes, indicating that all the supertype classifiers are covered by the subtyping relation and, hence, that the metamodel graph is indeed a supertype of the deterministic state machine metamodel specification.

And that the complement of the subtype metamodel contains the references initial, final and subMachines between the classes State and StateMachine, indicating that the inferred subtyping is a partial typing.

To reuse the ATL transformation for state machines, the model representing the state machine is re-typed to the extension metamodel, obtaining a model conforming to the extension metamodel, to which the ATL transformation can be applied. In this scenario, this step is optional as the re-typing does not apply any changes to the model. However, in general, such a re-typing may apply changes as illustrated in Scenario 3: DSML evolution (Reuse, Partial Typing). After applying the ATL transformation, another model conforming to the extension metamodel is obtained where concepts from the supertype may have been created (such as Mark). A second re-typing is necessary in order to ensure that the produced model conforms to the original state machine metamodel. This process is illustrated as follows:

As discussed in the paper, the subtyping relation must be strict in order for the last re-typing to work automatically.


MIT License

Copyright (c) 2017 Artur Boronat

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.



Our tool reuses (adapting and extending) the following third-party tools as libraries:

  1. Jim Steel, Jean-Marc Jézéquel. On model typing. Software and Ssytem Modeling 6, 4 (2007), 401-413.  2

  2. Clement Guy, Benoît Combemale, Steven Derrien, Jim Steel, Jean-Marc Jézequel. On Model Subtyping. ECMFA 2012. 400-415.