Perturbation Analysis of Computer Programs

Ermr flow analysis is the study of how errors originate, spread, and propagate during progmm execution based on the three steps of the fault/failure model: execution, infection, and propagation[l6]. These three steps are defined relative to a virtual computer by judiciously selecting the instruction set and data state of this computer the need for infection analysis can be reduced or eliminated in favor of execution and propagation analysis. A key aspect of propagation analysis is injecting errors into the data state and tracing their effect. Perturbation analysis injects errors by directly modifying the data state of an executing program. The resulting code that is executed the tail code is analyzed for its error-flow behavior. Perturbation analysis is a language-independent and eflcient method of chamcterizing the propagation rate of each tail function, the function computed by all tail code originating a t a given location. This paper defines a model for perturbation analysis, and uses the model to explain the performance of analysis techniques (e.g. statement, data flow, and mutation analysis). Perturbation analysis can be used to compare testing stmtegies beyond the tmditional subsumption hiemrchy. Though bmnch testing subsumes statement testing, perturbation experiments indicate that the improvement may be only slight. Furthermore, even though mutation testing and data flow testing are not strictly comparable by subsumption, perturbation experiments show that mutation is superior for a particular set of programs. Perturbation analysis supports other kinds of analysis and testing. It aids safety analysis by characteriz0-7803-3979-7/97/$10.00 01997 IEEE 77 ing the error-flow behavior of progmms under circumstances when the progmm is faced with abnormal data states caused by hardware or software faults. Perturbation analysis may also form the basis for more errorrevealing testing techniques, two of which are described.