Test Effectiveness

White Paper –

Module Testing Techniques – which are the most effective?  – Results of a Retrospective Analysis

 If you are module testing as part of software development then you face a predicament: Which techniques, do you use to design the test cases, and which measures of test coverage do you require to be achieved? Industry guidelines and standards often mandate the required testing and many developers believe they know the appropriate level of testing for their system. However, what is the basis of this knowledge? This paper argues that we are currently basing our testing decisions on unsafe foundations and presents results from an empirical study, which appear to refute common testing wisdom.

A retrospective analysis of an operational system was performed to determine what would have been the effectiveness of the most commonly-used testing techniques if applied to an operational avionics system of approximately 20,000 lines of Ada code, produced to satisfy an ‘essential’ level of criticality.

The test case design techniques covered were equivalence partitioning, boundary value analysis, branch, branch condition, branch condition combination, modified condition decision coverage and random testing; all the techniques were applied as defined in the BCS Software Component Testing Standard (now BS 7925-2 & ISO/IEC/IEEE 29119-4).

The results of the analysis show, for this system, that several widely-held assumptions, some apparently sup-ported by industry standards, do not hold. For instance, testing to achieve 100% branch coverage would have detected fewer faults than using equivalence partitioning. Also, random testing would have been surprisingly effective. Random testing would have achieved similar levels of effectiveness for the same number of test cases as any of the other techniques studied except for boundary value analysis, and with only six random test cases per module would have outperformed branch, branch condition, branch condition combination or modified condition decision coverage (the level required for safety-critical avionics software). In contrast, boundary value analysis fared particularly well in this study, and over 50,000 random test cases per module would have been required to equal its test effectiveness.

Test Effectiveness PaperTest Effectiveness Paper