The document describes Mendel, a model and set of metrics and rules for understanding class hierarchies. Mendel identifies interesting classes based on size and novelty within a hierarchy. It classifies subclassing behaviors as extending (adding methods) or overriding. Mendel was applied to sample programs like JHotDraw and Azureus to gain insights into their class hierarchies and inheritance usage. Future work includes family analysis and improved visualization.

1. Mendel:
A Model, Metrics, and Rules
to Understand Class Hierarchies
Simon Denier & Yann-Gaël Guéhéneuc
Ptidej Team – Université de Montréal
ICPC 2008 12/06/2008

2. 22 / 28/ 28
Outline
 Motivation and Problem
 Mendel Model
 Interesting Classes
 Subclassing Behaviors
 Conclusion

3. 33 / 28/ 28
Motivation
 Context: maintenance of program without documentation
 Task: identify and understand the role of multiple classes
 Problem 1: where to start?
 Find entry points (main(...))
 Read large classes (by LOC or Number of Methods)
 Problem 2: what about inheritance?

4. 44 / 28/ 28
Problem: Class Hierarchies
 Classes defined incrementally by inheritance
 Class “interesting” in its own hierarchy
 Class profiting from its hierarchy

5. 55 / 28/ 28
Class Hierarchies
JHotDraw
5.2

6. 66 / 28/ 28
Class Hierarchies (2)
Azureus
2.4.0.2

7. 77 / 28/ 28
Outline
 Motivation and Problem
 Mendel Model
 Interesting Classes
 Subclassing Behaviors
 Conclusion

8. 88 / 28/ 28
Mendel Approach
 Goal: insights on class hierarchies
 Interesting classes in their own hierarchy
 Subclassing behaviors: extending (adding) vs overriding
 Characteristics (requirements) of our approach
 Early stage in maintenance
 Intuitive
 Fast

9. 99 / 28/ 28
Mendel: Generic Model
A
foo()
bar()
B
bar()
zorg()
Simple UML
bar()
foo()
zorg()New
Overridden
Inherited
Super
Local
Total
B
Mendel

10. 1010 / 28/ 28
Mendel: Metrics and Rules
 Simple metrics:
 Based on interfaces = sets of method signatures
 Computed on a single class or a hierarchy
 Rules for:
 Selection (threshold)
 Classification
 Comprehension (rule of thumb)

11. 1111 / 28/ 28
Outline
 Motivation and Problem
 Mendel Model
 Interesting Classes
 Subclassing Behaviors
 Conclusion

12. 1212 / 28/ 28
Interesting Classes
 Classes defining a lot of new behaviors
 Large class: overall system
 Budding class: with respect to its hierarchy
 Blooming class: large and budding

13. 1313 / 28/ 28
Large Class
 Class with a large interface
 Large Class: |Local| > threshold( |Local| )
 JHotDraw
1.StandardDrawingView (inherits from JPanel), 55 methods
2.DrawApplication (inherits from JFrame), 53 methods
3.AbstractFigure (inherits from Object), 34 methods

14. 1414 / 28/ 28
Budding Class
 Class with a large interface with respect to its hierarchy
 Metrics:
 Branch Mean Size – bms = |Total| / (DIT + 1)
 Novelty Index – nvi = |Local| / bms(superclass)
 Budding class: nvi > threshold(nvi)
 JHotDraw
1.AbstractFigure, 3.09
2.TextFigure, 1.84
3.PolygonFigure, 1.59

15. 1515 / 28/ 28
Blooming Class
 Synthesis: combining large and budding
 Novelty Score – nvs = |Local| × nvi
 Blooming Class: nvs > threshold(nvs)
 JHotDraw
1.AbstractFigure, 105.09
2.TextFigure, 55.1
3.LineConnection, 49.55

16. 1616 / 28/ 28
JHotDraw Overview with Mendel
 JHotDraw 5.2: 148 classes
 13 large (9%)
 18 budding (12%)
 11 blooming (7%)
 Among 11 blooming classes
 8 in AbstractFigure hierarchy
 + DrawApplication, StandardDrawingView, ConnectionTool

17. 1717 / 28/ 28
JHotDraw Base Classes
 Given documentation: 6 base classes

18. 1818 / 28/ 28
Comparison with Documentation
 5 out of 6 in at least one set
 AbstractFigure in all 3 sets
Budding
Large
Blooming

19. 1919 / 28/ 28
What about StandardDrawing?
 Max rank: 16th
by |Local| order
 Reason: overrider of CompositeFigure (4th
blooming class)
 Conclusion: StandardDrawing to be understood in relation
with CompositeFigure

20. 2020 / 28/ 28
Outline
 Motivation and Problem
 Mendel Model
 Interesting Classes
 Subclassing Behaviors
 Conclusion

21. 2121 / 28/ 28
Subclassing Behaviors
INH INHINH INH
OVR OVROVR
NEW NEW NEW
Overrider Extender Pure ExtenderPure Overrider
|NEW|=0 |OVR|≥|NEW| |NEW|>|OVR| |OVR|=0

22. 2222 / 28/ 28
Extender vs Overrider
 Rules of thumb
 Pure extender
 Adding new methods: no access through superclass
 Used for itself
 Implementation inheritance
 Pure overrider
 Overriding methods
 Used through substitution to superclass
 Interface inheritance

23. 2323 / 28/ 28
JHotDraw in Verso

24. 2424 / 28/ 28
Azureus in Verso

25. 2525 / 28/ 28
Implementation in Java
 Tailored for Java
 BCEL as class files parser
 Performance for Azureus (1681 classes)
 Standard desktop machine (2Ghz processor, 1Go memory)
 9 seconds

26. 2626 / 28/ 28
Outline
 Motivation and Problem
 Mendel Model
 Interesting Classes
 Subclassing Behaviors
 Conclusion

27. 2727 / 28/ 28
Conclusion
 Mendel
 Generic model
 Simple metrics
 Rules
 Results
 Quick insights on interesting classes and concerns
 Large, budding, blooming: different views of interesting classes
 Concerns: keywords in classnames
 Different usage of inheritance in programs
 Some leaned to extending (Log4J, Azureus)
 Others to overriding (JHotDraw, ArgoUML)

28. 2828 / 28/ 28
Future Work
 Family analysis
 Family = class + children
 Common behavior and interface
 Visualisation
 Hierarchy representation
 Visual patterns

29. 2929 / 28/ 28
ArgoUML in Verso

30. 3030 / 28/ 28
Blooming Classes Selectivity
 Log4J: 2 of 206 (1%)
 LogBrokerMonitor (104 methods), Category
 Budding set: Appender hierarchy
 ArgoUML: 18 of 1431 (1%)
 2 very big classes: NSUMLModelFacade, FacadeMDRImpl
 Design choice: Facade design pattern
 Keywords in class names: two products, normal and MDR
 Azureus: 54 of 1681 (3%)
 Keywords in class names: three concerns, download, peer, DHT

31. 3131 / 28/ 28
Behaviors Discrimination
 Two categories
 JHotDraw and ArgoUML: overriding behaviors
 Large proportions of Pure Overriders
 Azureus and Log4J: extending behaviors
 Azureus: 58% of classes are Pure Extenders