ACDC An Algorithm for ComprehensionDriven Clustering

1
ACDC An Algorithm for Comprehension-Driven
Clustering
Vassilios Tzerpos R.C. Holt
Presenter Romil Jain
2
Presentation Road Map

• Introduction
• Clustering for Comprehension
• Subsystem Patterns
• The ACDC Algorithm
• Orphan Adoption
• Algorithm Validation
• Conclusions

3
Introduction

• Two main approaches to clustering
• Knowledge-based Use domain knowledge to
  understand source code functionalities and
  cluster files accordingly.
• Structure-based Extract syntactic interactions
  (e.g. function calls, variable reference) and
  cluster the software based on module dependency
  graphs (MDG).
• Most researchers agree on structure-based
  criteria and naming conventions as promising.

4
Introduction(contd)

• Primary goal is software comprehension.
• Various criterions low coupling, high cohesion,
  interface minimization etc.
• Research focuses mostly on increasing performance
  and accuracy of algorithms based on these
  criterions.
• Instead focus should be on creating
  comprehensible clusters.

5
Clustering for Comprehension
http
main
https
6
Clustering for Comprehension

• A clustering algorithm should have the following
  features for easy comprehension
• Effective cluster naming.
• Bounded cluster cardinality Up to 20 files or so
  per cluster.
• Pattern-driven approach Create a decomposition
  based on certain patterns in the software system
  which human beings can easily identify.

7
Subsystem Patterns

• Source File All procedures/variables in same
  file can be grouped together.
• Directory Structure can give clues for
  clustering.
• Body-header E.g. .h and .c files can be put
  together.
• Leaf Collection Modules for similar purposes
  (e.g. set of drivers) can be put together.
• Support Library Modules accessed by majority of
  other modules (e.g. utility/math libraries)

8
Subsystem Patterns (contd)

• Central Dispatcher Pattern (Dual of Support
  Library Pattern) Resources with large out-degree
  (e.g. main.c, engine.java)
• Subgraph Dominator Pattern A set of nodes N
  n0, n1, such that any other node must go
  through n0 to reach ni?N.

9
The ACDC Algorithm

• Performs clustering in 2 stages
• Constructs a skeleton decomposition using
  pattern-driven approach. Gives appropriates names
  to clusters based on the pattern used.
• Completes decomposition using a technique called
  Orphan Adoption, which clusters the files which
  were not clustered in the 1st stage.

10
ACDC - Skeleton Construction

• ACDC performs the following steps in order
• Source file clusters Cluster resources that
  reside in same file together, and consider only
  files for the next steps ().
• Body-header conglomeration Consolidate .h and .c
  files in same cluster ()
• Leaf collection and support library
  identification Just identify, and do not
  cluster, files exhibiting these patterns (For
  support library nodes in-degree 20)

11
ACDC - Skeleton Construction

• Ordered and limited subgraph domination
• Disregard any files with Central Dispatcher
  Pattern (i.e. with out-degree 20) and their
  edges too.
• Find subgraph dominator patterns and create a
  separate subsystem containing dominator node and
  dominated set. ()
• Organize various subsystems in a containment
  hierarchy (maximum cardinality 20).
• Consider the previously disregarded files again
  for subgraph domination w.r.t. these subsystems.

12
ACDC - Skeleton Construction

• Creation of Put any remaining files
  identified previously as support libraries in
  this cluster.

13
ACDC Orphan Adoption

• Many files may still remain un-clustered. These
  are called Orphans.
• Attempt to place each orphan in a subsystem which
  has larger connectivity to it.
• If the leaf files have not been assigned yet, put
  them in

14
Algorithm Validation

• ACDC was run on the following systems
• TOBEY Used by IBM for optimizing compiler
  products. Contains 939 source files and 250,000
  lines of code.
• Linux 955 source files and 750,000 lines of code
• Results

TOBEY is professionally developed, while Linux
is open source.
15
Conclusion

• Clustering should be comprehension driven
• ACDC attempts to do this with the following
• Manageable cluster size (? 20).
• Meaningful subsystem naming.
• Pattern driven clustering (easier to understand)
• Good quality clustering makes it a good candidate
  for engineering projects.

16
Thank You!

• References
• V. Tzerpos and R. C. Holt. ACDC An algorithm for
  comprehension driven clustering. In Proceedings
  of the 7th Working Conference on Reverse
  Engineering, pages 258267, Brisbane, Australia,
  November 2000.