Program restructuring using clustering techniques

Abstract

Program restructuring is a key method for improving the quality of ill-structured programs, thereby increasing the understandability and reducing the maintenance cost. It is a challenging task and a great deal of research is still ongoing. This paper presents an approach to program restructuring inside of a function based on clustering techniques with cohesion as the major concern. Clustering has been widely used to group related entities together. The approach focuses on automated support for identifying ill-structured or low-cohesive functions and providing heuristic advice in both the development and evolution phases. A new similarity measure is defined and studied intensively specifically from the function perspective. A comparative study on three different hierarchical agglomerative clustering algorithms is also conducted. The best algorithm is applied to restructuring of functions of a real industrial system. The empirical observations show that the heuristic advice provided by the approach can help software designers make better decision of why and how to restructure a program. Specific source code level software metrics are presented to demonstrate the value of the approach.

Introduction

Software evolves over time primarily due to changes in requirements and technologies. As a result, huge amount of effort is spent in maintenance and evolution. Software evolution usually accounts for more than 60% of total software costs (Sommerville, 1996). In today’s highly competitive era, software development is often driven by tight schedules. Hence, software designers often emphasize the functional aspect of a system. Even if a software product is well designed, the code is often modified over time in response to the changing needs of customers. As a consequence, its original structure gradually drifts and quality degrades. Hence, the program becomes difficult to understand. As a result, it is often costly to maintain.

Müller, et al. indicate that 50–90% of software evolution work focuses on program comprehension or understanding (Müller et al., 1995). Program understanding could be at various levels, including architecture, design, and code. At the implementation level, a large or poorly coded function usually involves multiple activities or has low functional cohesion, which makes the program difficult to understand and modify. Program restructuring (Chikofsky and Cross, 1990) or refactoring (Fowler, 1999) can transform these functions to functions that are better organized and easier to understand, without changing their behaviors. The new functions will usually be higher quality and less costly for further evolution. More importantly, a desirable restructuring should achieve high cohesion and low coupling (Briand et al., 1996, Munson, 2003, Pressman, 1997).

Cohesion, as an important measure in restructuring, is to measure how tightly related elements are in a component. The goal of clustering is to group similar or related elements together. It is possible to use clustering analysis to measure the strength of the relationship between elements in a component. Previous articles of software clustering demonstrate research potential in software clustering field (Tzerpos and Holt, 1998) and conclude that clustering methods may be a very good starting point for the remodularization of software (Wiggerts, 1997).

However, existing research on the software clustering field has mainly been concerned with software remodularization at the architecture level and has not been used in program restructuring at the source code level. Source code contains critical information regarding the behavior of a system. The understanding and manipulation of source code is a pressing issue for maintenance and evolution. This paper focuses on source code. Specifically, this paper deals with restructuring of each individual function. One challenge of restructuring at this level is how to meaningfully and effectively group related code segments together inside a large or poorly structured function to form small or cohesive functions, because it is not uncommon that unrelated fragments and functionally cohesive code segments are interleaved in real software products. In addition, the approach should be easy to understand and also effective in practice. Clustering techniques are suitable for this problem, because the objective of clustering is consistent with that of cohesion.

This paper presents an approach to program restructuring using clustering techniques at the function level. It focuses on using automated support for identifying low-cohesive functions and making restructuring decisions, instead of the automated restructuring process. The purpose is to help software designers identify ill-structured functions and provide them with heuristic advices. In detail, this paper discusses how to select entities and how to select attributes that are important to distinguish two different entities from the cohesion perspective. A new resemblance coefficient as a similarity measure is defined. Extensive experiments on the weights of different attributes are conducted. Three hierarchical agglomerative algorithms: single linkage algorithm (SLINK), complete linkage algorithm (CLINK) and weighted pair-group method using arithmetic averages (WPGMA¹), are chosen and an intensive comparative study on them is conducted. These algorithms are highlighted as follows:

• SLINK: also called the nearest neighbor method. It defines the similarity measure between two clusters as the maximum resemblance coefficient among all pair entities in the two clusters.

• CLINK: also called the furthest neighbor method. It defines the similarity measure between two clusters as the minimum resemblance coefficient among all pair entities in the two clusters.

• WPGMA: this defines the similarity measure between two clusters as the simple arithmetic average of resemblance coefficients between two clusters without considering the cluster size.

The algorithm that produces the best result will then be applied to program restructuring of an industrial system.

The structure of the rest of this paper is as follows. Section 2 reviews the related work in both program restructuring and software clustering areas. Section 3 proposes an approach to program restructuring using clustering techniques and discusses the issues involved in the approach. Section 4 provides an extensive study on the similarity measure by weighting attributes differently. Section 5 gives a comparative study of three clustering algorithms: SLINK, CLINK and WPGMA. Section 6 presents a case study of program restructuring using the clustering results on an industrial software system. Empirical observations are also summarized. Section 7 presents the conclusions and future work.