ACDC: An Algorithm for ComprehensionDriven Clustering

1
ACDC An Algorithm forComprehension-Driven
Clustering

• Vassilios Tzerpos
• R.C. Holt

2
Highest cohesion clustering
SS1
SS6
SS9
SS5
SS12
SS2
SS8
SS4
SS11
SS7
SS10
SS3
3
Contents of SS11
4
Essential comprehension features

• Effective cluster naming
• Bounded cluster cardinality
• Familiarity
• Comprehension as pattern recognition
• Certain subsystem patterns emerge often in manual
  decompositions of software systems

5
Source file pattern
File1
File2
Var2
Var1
Var3
6
Directory structure pattern
Dir1
Dir2
File1
File4
File5
File2
File6
File8
File3
File7
File9
7
Body-header pattern
bob.c
alice.c
bob.h
alice.h
8
Leaf collection pattern
sin.c
tan.c
cos.c
9
Support library pattern
busy.c
weary.c
tired.c
10
Central dispatcher pattern
dispatcher.c
11
Subgraph dominator pattern
dominator.c
a.c
c.c
b.c
z.c
g.c
f.c
e.c
d.c
12
The ACDC algorithm

• Two stages
• Using a pattern-driven approach, a skeleton of
  the final decomposition is created. Subsystems
  are named appropriately.
• The decomposition is completed by applying an
  extended version of the Orphan Adoption algorithm

13
Skeleton construction

• Source file clusters
• Body-header conglomeration
• Leaf collection and support library
  identification
• Ordered and limited subgraph domination
• Creation of support.ss

14
Orphan Adoption

• Incremental clustering technique
• Orphan a newly introduced resource to a software
  system
• Orphans are adopted by the subsystem that
  interacts mostly with them
• Assuming that a substantial skeleton has been
  constructed in the first stage, the same
  technique can be applied here

15
ACDC properties

• Subsystems have familiar or intuitive names
• The cardinality of the subsystems is bounded
• The final decomposition is nested and unbalanced
• Limited use of the directory pattern
• Magic numbers not important

16
Algorithm validation

• We experimented with two different software
  systems, TOBEY and Linux.
• We measured the following
• Performance
• Stability
• Skeleton size
• Quality

54 sec 84 sec
81.3 69.4
64.3 51.1
64.2 55.7
17
Conclusions

• Clustering approaches should focus on
  comprehension
• Pattern-driven approach appears to perform
  satisfactorily
• Impact of ACDCs features on comprehension
  remains to be determined