ACDC: An Algorithm for Comprehension Driven Clustering

1
ACDC An Algorithm for Comprehension Driven
Clustering
2
Introduction

• Software clustering techniques are used to
  decompose large software systems into subsystems
• Common clustering criteria include low coupling,
  high cohesion etc.
• ACDC does not attempt to satisfy certain
  criteria.
• ACDC decomposes a software system for the purpose
  of system comprehension
• ACDC discovers clusters that follow commonly
  occurring patterns observed in decompositions of
  large software systems

3
Background

• Two common existing software clustering
  approaches
• knowledge-based approach
• based on domain knowledge
• grouping modules of similar or complementary
  functionality
• structure-based approach
• looks at interactions between entities
  (procedures, variables)
• typical interactions are method calls or
  variable accesses
• clustering becomes partitioning the vertex set of
  a graph
• graph nodes are entities and edges are
  interactions

4
Clustering for Comprehension

• Comprehension driven clustering algorithm
  features
• effective cluster naming
• meaningful cluster names aid in program
  understanding
• bounded cluster cardinality
• creating clusters with a limited number of
  objects
• pattern-driven approach
• using commonly occurring patterns to create
  clusters

5
Subsystem patterns

• Commonly occurring subsystem patterns
• source file pattern
• set of procedures and variables contained in the
  same source file can be grouped together to form
  a cluster
• directory structure pattern
• directories may correspond to subsystems in some
  cases
• body-header pattern
• clustering files that split procedures into two
  different files

6
Subsystem patterns

• leaf collection pattern
• independent set of files serving similar purpose
• usually represent leafs in the system graph
• support library pattern
• set of procedures accessed by the majority of
  subsystems
• central dispatcher pattern
• dual of the support library pattern
• represents nodes with a large out-degree
• common example is the procedure called driver

7
Subsystem patterns

• subgraph dominator pattern
• particular type of subgraph of a system graph
  G(V,E)
• this subgraph must contain a denominator node n0
  and a set of nodes n
• in order to reach nodes n from any of the
  remaining nodes in the graph one must go through
  the denominator node

8
The ACDC Algorithm

• The ACDC algorithm performs clustering in two
  stages
• Stage1 Skeleton construction
• identifying subsystems using pattern-driven
  approach
• Stage 2 Orphan adoption
• maintaining systems decomposition as the system
  evolves

9
Stage1 Skeleton Construction

• ACDC performs following steps in order of
  precedence
• constructing source file clusters
• ACDC clusters resources belonging to the same
  file and uses the name of the file as the cluster
  name
• Body-header conglomeration
• the interface and the implementation of a
  software module from a cluster
• for example, in C, files foo.h and foo.c are
  grouped into cluster foo.ss

10
Stage1 Skeleton Construction

• leaf collection and support library
  identification
• ACDC identifies collections of files according to
  the leaf collection and support library pattern
• no clusters are formed at this point
• ordered and limited subgraph domination
• main step of the algorithm
• considers nodes in ascending out-degree when
  trying to identify nodes following the subgraph
  denominator pattern
• discovered subsystem name is the name of the
  denominator node plus the suffix .ss
• obtained subsystems are organized in a tree-like
  containment hierarchy
• number of nodes in a subgraph is bounded

11
Stage1 Skeleton Construction

• Creation of support.ss subsystem
• final step of skeleton construction process
• left over files previously identified as
  candidates for a support library pattern are
  assigned to this subsystem

12
Stage2 Orphan Adoption

• incremental clustering technique
• assigns all files to some subsystem

13
ACDC Algorithm properties

• The algorithms satisfies three essential
  features
• effective cluster naming
• meaningful names are assigned in each step of the
  algorithm
• bounded cluster cardinality
• number of nodes in each cluster is limited
• pattern-driven approach
• several commonly occurring patterns are used

14
Algorithm validation

• ACDC has been validated on two large systems
• TOBEY
• took only 54 seconds to cluster
• obtained skeleton size was 64.3 the original
  size
• Linux
• took only 84 seconds to cluster
• obtained skeleton size was 51.1 the original
  size

15
Conclusion

• ACDC aids in program comprehension
• uses commonly occurring decomposition patterns
• provides encouraging experimental results