Software systems are often highly structured, consisting of artifacts (types, methods, variables, and packages), and relationships between these artifacts. Domain models, meta models, and software design documentation provide additional artifacts such as roles, associations, use cases, and paragraphs of text. This paper describes, and demonstrates the use of a tool for software structure understanding. The tool consists of a knowledge base containing software artifacts, relationships between artifacts, and rules for generating new relationships. The knowledge base is then explored using formal concept analysis (FCA). Exploration of the software structure via FCA is useful in: (i) understanding the software structure, (ii) the recognition of parts of the software structure violating design principles, (iii) the reorganisation of type and package hierarchies, and (iv) retrieving artifacts of the software development process.
The design of software systems is fundamentally a human activity and therefore necessarily involves human understanding of software structure. This observation coniflcts with the complexity of modern software systems and so there is a need for automated tools which help humans understand software systems. Such tools need to be able to summarise aspects of the software structure so as not to overload the human user with too much information. To do this it seems natural to exploit the explicit generalisation/specialisation in hierarchies and to allow diagrams to range in their level of abstraction from very general to very specific.
The process of software design and implementation often contains many
arbitrary decisions, such as the name of methods or variables or indeed sometimes
the structure of a class hierarchy. As the design proceeds these decisions need to
be reviewed in order to achieve consistent, orthogonal and simple designs. Agile
methods, and extreme programming (XP) in particular, advocate regular
refactoring activities undertaken to regularise and revise the software structure [
Our tool, Conceptual Analysis of Software Structure (CASS), attempts to address these requirements. The architecture of CASS is shown in Figure 1. Source code analysers and profilers produce information which is stored in a knowledge base. A rule based system is then used to extend the knowledge base with new relationships and artifacts. Graph based queries define an aspect of the code to be explored and are used to generate result sets that are visualised using concept lattices. Hypotheses and questions may then be investigated by (i) generating new lattices, perhaps displaying new aspects of the software structure; and (ii) by navigating back to the source artifacts within the software or its documentation. Since each concept lattice is generated from a query graph, a natural renfiement ordering allows general views to be elaborated and made more specific. Thus the user is able to progress from a general view to a more specific view, or vice versa. In addition, the theory of formal concept analysis (FCA) allows two or more aspects of the software structure to be combined coherently in nested line diagrams.
This paper presents and discusses examples of concept lattices extracted during the analysis of a piece of software called OntoRama. OntoRama is an open source project of a reasonable size that has been in development for several years, has had several developers, and has gone through a number of signicfiant design changes. Due to these properties we think that OntoRama is a realistic example of software analysis. OntoRama was choosen because although we were not familiar with the software design we have had some access to the software developers who worked on OntoRama. Thus the examples in this paper are constructed from the point of view of a developer who is new to piece of software and seeking to understand it for the purpose of maintaining or extending it. This is a common situation in the software development industry.
The examples have also been choose to illustrate a range of activities made possible within the CASS framework. The rfist example (Section 4) is an analysis of the call graph between top level packages in OntoRama and gives an indication of how modular the source code is. It shows which packages are dependent on which other packages. The second example shows how an aspect of the call graph can be investigated by unfolding the lattice structure. The third example, still making use of the call graph, focuses attention on two specicfi packages. The forth example shows how FCA can be used to analyse the naming of packages to gain insight into the software structure. In the ffith example, two different aspects of the software structure — one derived from the call graph information, and a second derived from package naming — may be combined together in a nested line diagram.
Before discussing the examples we first review some related work and then explain how concept lattices are produced in response to graph queries. 2
Tilley et al. [
Late phase approaches focus on the maintenance and reengineering of
software. These approaches include; (i) identification of objects [
There are a number of approaches that store and visualise software structure
as graphs. Two well known examples being Rigi [
The novelty of our approach lies in the exploration of program information
stored in a knowledge base using FCA. The techniques presented by other
authors generally focus on a single aspect of software structure and analyse just
that aspect. For example Snelting and Lindig consider preprocessor statements
in C programs, while Snelting and Tip consider the static call graph in a C++
program. Our approach provides a mechanism to bring these different aspects
of software structure together in a single framework and application. By using
a graph based query language we can exploit the implicit structure captured in
class hierarchies, package hierarchies, and call graphs to help the user focus on
particular aspects of the software structure. Our approach differs from tools such
as Rigi and SHriMP in that the data under analysis is organised using concept
lattices. These lattices are algebraic structures and convey information about
the logic underlying the data. The user must be trained to read concept lattices
in the same way that users learn to interpret UML diagrams. Although a full
exposition of the mathematics underlying FCA takes several years of study, an
understanding of the diagrams is usually acquired in about half an hour [
In some instances the query is made up of several graphs. In this case a binary relation is returned from the result set of each graph and a union is taken between these binary relations to form the formal context. This allows two aspects of a software structure to be combined within a single diagram.
The graph in Figure 2 can be expressed in a linear form. The linear form is a list of triples separated by commas. Each triple corresponds to an edge in the graph. Two triples can be merged together if the last node of the first triple is the same as the rfist node of the second triple. The linear form for Figure 2 is: caller in-t g in ontorama, caller calls-t callee, callee in-t m in ontorama.
The objects of the formal context are retrieved using the following query graph: “ g in ontorama.” While the attributes are retrieved using: “ m in ontorama.” Having separate queries for the object and attribute sets is important because otherwise objects and attributes that one might expect, for example in a lattice showing the call structure of the top level packages, can otherwise be omitted from the diagram and thus be a cause of confusion. It is also often the case that these otherwise omitted objects and attributes are a point of interest in the diagram. When they are included, as occurs with explicit query graphs for the object and attribute sets, then they are attached to the top and bottom concepts respectively.
For our knowledge base we use a triple store called Kowari3. Kowari is designed for storing and retrieving RDF statements and scales to hundreds of millions of triples. Both our graph queries and our rules are translated automatically into iTQL which is a query language of Kowari. iTQL is similar at the syntax level to SQL. Our requirements are somewhat simpler than those of RDF since we: (i) don’t make use of namespaces, (ii) don’t distinguish between resources and literals, (iii) we don’t store anonymous nodes within the database, (iv) are rules have no anonymous nodes in their conclusions and so our rule system is decidable and has a finite closure. As a rough guide it takes about 10 seconds to import OntoRama into Kowari and then another 10 or so seconds to calculate the rule closure. Subseqent queries are usually processed in under a second. 4
call graph which is generated by recording the actions of a running program. By taking the transitive closure, if method A calls method B which in turn calls method C then we also record a transitive call from A to C.
The reason, in this instance, for taking the transitive closure is to focus on whether or not the top level packages fall neatly into layers with classes from upper layers making calls to classes from lower layers. If such layers exist in the call graph then we expect to see them in the concept lattice. In contrast, if there are cyclic dependencies then we shall expect to see large intervals between the object and attribute concepts for a package.
OntoRama has cyclic dependencies. To understand why these result in large intervals let us consider an example. Consider the interval formed by the object and attribute concepts for ontorama.ui. The object concept for ontorama.ui is near the bottom of the diagram while the attribute concept for ontorama.ui is around the middle. Objects concepts that occur in the interval both make calls into ontorama.ui and are themselves called into from within ontorama.ui. We know these packages make calls into ontorama.ui because they have ontorama.ui as an attribute. We can also assume that ontorama.ui calls into these packages because the object concepts for these packages are above the object concept for ontorama.ui, and it is usual for each package to make calls into itself.
The top of the concept lattice is a chain product. With the attributes OntoramaConfig, and util forming one chain, the attribute backends forming another, and the conjunction of ontotools and model forming yet a third. These chains can be thought of as dimensions. The model package uses all four dimensions, while importer leaves out util and TestPackage leaves out both OntoramaConfig and util. The attributes at the top of the diagram can be considered low level packages because they are called into by many other packages.
All most all of these low level packages have large intervals indicating that they both make calls into many packages, and are also called into from within many packages. The exception is util which has a relatively small interval. Even so the util package is still involved in cyclic dependencies with both OntoramaConfig and backends.
The concept for TestPackage sits out to the right being mutually exclusive with many of the attributes in the diagram. It calls very few packages and is itself called by very few packages. The TestPackage class initiates tests stored within the other packages. Since TestPackage only makes calls ontotools and model we can assume that many of the packages don’t have testing. Another possibility is that there are other unit tests in the other packages, but they have not been hooked into the TestPackage class.
The lower part of the diagram has to do with ui. The ui package is called by several packages and also makes calls to many packages. We see that the view package is both called by ui and also makes calls into the ui package. We see that view occurs as an attribute fairly low down in the structure indicating that it is a relatively high level package, being called from relatively few other packages, but itself making calls into many other packages. By seeing that the attribute concept for views is on the top plane of the lattice (visually speaking) we can see that it does not make calls into the backends package which is the attribute forming the lower plane of the lattice.
The ui package is somewhat in the middle of the structure both being called by many classes and calling many classes. To see whether the substructure of the ui package reeflcts this we unfold the ui package. This forms the second example. 5
caller in-t g in ontorama, caller calls-t callee, callee in-t m in ontorama.
OR caller in-t g in ontorama, caller calls-t callee, callee in-t m in "ontorama.ui".
The lattice has been drawn as a locally scaled nested line diagram. The attributes of the inner scale are the contents of the ui package while attributes of the outer scale are the same in the first Example.
The rfist interesting aspect is that the inner scale is a very simple lattice being composed of a single chain. By far the majority of the content of ui appear as attributes of the bottom concept, which is instantiated for the rfist time in the object concept for ui. This indicates that this part of the ui module is only used by the ui module itself. The other three parts which are called by other packages are: the class OntoRamaApp, the class ErrorDialog, and the package events. Because OntoRamaApp is attached to the top concept of the inner scale we can infer that everything that calls ui also calls OntoRamaApp. The only package to call ui and not call ErrorDialog is ontotools. While events are accessed (perhaps in the mode of event creation) from the importer and backends packages.
Returning to our initial question of whether or not the organisation of ui reflects the usage of ui by other packages and classes. We can say that usage of ui by other packages is not a primary concern of ui since it is not used by many other classes and packages. Therefore it can be said that its package structure is not organised in that way. Constructing the locally scaled nested line diagram however has provided an organisation over the contents of the ui package that is aligned with its usage by other packages. 6
caller in-t g in "ontorama", caller calls-t callee, callee in-t m in "ontorama.ui.events".
The concept lattice shows that the ui class uses all the events in the package, and also that, the packages importer, views and backends, each have their own events that are not shared with the other packages.
It is curious that of three events, according to their names, are associated with the life cycle of a query, they are each accessed separately from the three different packages. QueryEndEvent is access from importer, QueryStartEvent is accessed from backends, and QueryCancalledEvent is accessed from ui. It ontorama.util ontorama.model
ontorama.views ontorama.views ontorama.conf ontorama.OntoramaConfig ontorama.ui ontorama.backends ontorama.ontotools ontorama.model ontorama.ui.QueryPanel ontorama.ui.ConfigurationManager .o.n.torama.ui.controller ontorama.ui.action ontorama.ui.ProxySettings ontorama.ui.HistoryElement ontorama.ui .o.n.torama.ui..AboutOntoRamaDialog ontorama.ui.OntoRamaApp ontorama.ui.ErrorDialog ontorama.ui.events ontorama.TestPackage ontorama.TestPackage ontorama.importer
ontorama.importer ontorama.ontotools ontorama.backends could be that these events are being generated respectively in each of the three packages. This hypothesis can be verified by submitting the following query 4: caller in-t g in "ontorama", caller calls callee in m, m in-t "ontorama.ui.events", callee has-name "<init>"
This process of generating concept lattices, forming hypothesis and then generation more concept lattices to verify these hypotheses can continue at length as the programmer becomes slowly more familiar with the structure of the software. In the interests of showing a variety of techniques we leave this thread here and move on to the next example.
Packages in Java are identified by path names. The path names contain names separated by full stops and usually are used to place the packages within a tree structure. It is interesting however to consider the names in the path name of a package as attributes in a concept lattice. This has been done for the packages contained in the org.ontorama.model and the result is shown in Figure 6.
The concept lattice exhibits quite a lot of structure. It is the semi-product of an M3 and an M2. The M2 says that the model is divided into graph and tree, while the M3 says that the model is divided into controller, events, and text. The semi-product produces all combinations between these two structures.
One may surmise that the reason such an arrangement of the packages exists within OntoRama is that one of the main developers was familiar with formal concept analysis. Even so there is an irregularity in the diagram. The attribute concept for test has an object in its contingent; the package ontorama.model.test.
It is somewhat curious that both controllers and events could be partioned into graph and tree packages and that there were no controllers or events that are shared between graph and tree. One hypothesis is that such events are also shared perhaps with the ui or view packages and thus are to be found in ontorama.events or ontorama.controllers. Looking back to the lattice in Figure 3 we see that there is indeed neither an ontorama.events nor an ontorama.controller package and thus we can conclude that all events and controllers associated with the model have been assigned to either the graph or the view. 8
Two aspects of the software structure can be combined together within a nested line diagram. In software structure it is often the case that different aspects are correlated and a nested line diagram provides a mechanism to look for, and examine, any such correlations.
Figure 7 combines two different aspects of software structure. It combines the static call graph that we saw in the rfist three examples with the package names that we saw in the previous example. The inner scale organises sub-packages of the model package according to names in their pathnames, while the outer scale organises the packages according to which sub-package and classes in the view class they contain calls into.
Looking inside the top concept of the outer scale we can see that many of the packages don’t contain calls into the view package. We can tell this because the object concept contained in the top concept of the outer scale has no out scale attributes. The packages not making calls in the view are: graph.events and graph.test and tree.test. The graph.controller contains a call to views.textDescription, while tree.controller contains calls to view.tree and view.hyper. This suggests that view.tree and view.hyper are views of the tree while, views.textDescription is a view related to the graph. Interestingly we see that tree.events contains calls to both views.tree and views.textDescription, breaking the separation of the views into tree and graph views. The irregularity is curious and can be further investigated by considering the calls made from tree.events into views.textDescription and the calls made from graph.controller into views.textDescription. 9
In summary, this paper has demonstrated via a number of concrete examples how FCA, when used in conjunction with a knowledge base augmented by a rule based system, can be used to explore software structure from different points of view. We began with a very general lattice summarising the static call graph for the whole program and then elaborated the diagram to look inside the package. Having then become interested in a specicfi package we zoomed in and organised its components based on usage by the top level packages. Changing tack we then explored the names used in the package structure, before combining two aspects of software structure — package naming of components in the model package, and usage by view components. The examples presented here form only a small part of the exploration techniques afforded by our system.
Our approach is primarily aimed at human understanding of software structure with the purpose of refactoring the software to make it simpler and more uniform. Buy understanding the software structure and identifying places where it is irregular or doesn’t meet with expectations the user is able to identify opportunities for refactoring.
This research was supported by a Postdoctoral Fellowship granted by the Australian Research Council, an Australian Postgraduate Award, and the Distributed System Technology Centre. We gratefully acknowledge the contribution of Peter Becker, and Nataliya Roberts. ToscanaJ5 was used for the preparation of concept lattice diagrams.