Electronic Thesis and Dissertation Repository

Degree

Doctor of Philosophy

Program

Computer Science

Supervisor

Dr. Nazim H. Madhavji

2nd Supervisor

Dr. Mechelle S. Gittens

Joint Supervisor

Abstract

Corrective maintenance, which rectifies field faults, consumes 30-60% time of software maintenance. Literature indicates that 50% to 90% of the field failures are rediscoveries of previous faults, and that 20% of the code is responsible for 80% of the faults. Despite this, identification of the location of the field failures in system code remains challenging and consumes substantial (30-40%) time of corrective maintenance. Prior fault discovery techniques for field traces require many pass-fail traces, discover only crashing failures, or identify faulty coarse grain code such as files as the source of faults. This thesis (which is in the integrated article format) first describes a novel technique (F007) that focuses on identifying finer grain faulty code (faulty functions) from only the failing traces of deployed software. F007 works by training the decision trees on the function-call level failed traces of previous faults of a program. When a new failed trace arrives, F007 then predicts a ranked list of faulty functions based on the probability of fault proneness obtained via the decision trees. Second, this thesis describes a novel strategy, F007-plus, that trains F007 on the failed traces of mutants (artificial faults) and previous faults. F007-plus facilitates F007 in discovering new faulty functions that could not be discovered because they were not faulty in the traces of previously known actual faults. F007 (including F007-plus) was evaluated on the Siemens suite, Space program, four UNIX utilities, and a large commercial application of size approximately 20 millions LOC. F007 (including the use of F007-plus) was able to identify faulty functions in approximately 90% of the failed traces by reviewing approximately less than 10% of the code (i.e., by reviewing only the first few functions in the ranked list). These results, in fact, lead to an emerging theory that a faulty function can be identified by using prior traces of at least one fault in that function. Thus, F007 and F007-plus can correctly identify faulty functions in the failed traces of the majority (80%-90%) of the field failures by using the knowledge of faults in a small percentage (20%) of functions.


Share

COinS