International Journal of Reliability, Quality and Safety Engineering

There are two types of models for predicting software reliability at the end of testing. One is the software reliability growth model (dynamic model) based on a given set of time series data. The other is the software complexity model (static model) based on the development environmental factors which have an influence on the software reliability. As the dynamic model depends on the time factor and the test method used, its prediction accuracy does not necessarily correspond to the data of practical projects. On the other hand, the static model needs the many significant parameters to accurately predict the software reliability. However, it is very difficult to select the main factors that determine the significant parameters out of a great number of factors which affect software reliability. In order to resolve these problems, this paper proposes a model to predict the number of embedded errors in a program at the end of testing phase. This model is based on the testing characteristics such as error detection rate and test case density. The result of an experiment shows that the proposed model is more reliable than the conventional models.

A significant number of software reliability growth models (SRGMs) have been proposed in the literature over the past three decades or so. Most of these models ignored the possibility of imperfect debugging with subsequent introduction of new errors into the software system. In this paper, we present an SRGM which consists of a dual process of debugging followed by introduction of errors in the event of an imperfect debugging. The model also allows for the process to be repeated over several stages. The reliability growth of this model is investigated and is shown to exhibit either an exponential or S-shaped curve. Finally, some reliability measures associated with the SRGM are presented together with illustrative numerical examples.

A new class of software reliability models is proposed, based on practical considerations about the failure process and the influence of debugging on the successive software failure rates. It is assumed that, at each correction, a part of the previous software faults are removed and new faults can be introduced. Several well-known models belong to this class. With some additional assumptions, a simple model is proposed. The maximum likelihood equations for the estimation of its parameters are derived. The model is applied on real data and is compared to usual software reliability models.

This paper develops a plausible software availability model considering two types of failures during the operation phase. The first type is caused by the faults that could not be detected/corrected during the testing phase and the second type caused by those introduced by deviating from the specification during the operation phase. The former and the latter types of software failure-occurrence phenomena are described by a geometrically decreasing and a constant hazard rate, respectively. This model also describes the imperfect debugging environment in which a debugging activity does not always remove a fault perfectly. Taking notice of the cumulative number of faults corrected during the operation phase, we use a Markov process to describe the time-dependent behavior of the software system. Several quantitative measures of software system performance are derived from this model. Finally, numerical examples are presented for illustration of software availability measurement.

This paper treats an allocation problem of discrete testing resources in module testing where software modules are tested independently. The total amount of discrete testing resources is given beforehand. When an allocation is determined in module testing, each software module defines the share per size. Since discrete resources do not allow each software module to get an equal share per size, there exists a dispersion. The aim of this paper is to find an allocation minimizing the variance of this dispersion.

This paper summarizes models for software reliability demonstration testing (SRDT). The models are briefly classified into three types: (1) continuous models, (2) discrete models and (3) models considering damage sizes of software failures. Under the continuous models, the software product of interest is tested for time t and is accepted if the number of software failures in the test does not exceed a prespecified integer s. The values of design variables, t and s are determined based on (i) the concept of a statistical test (a statistical model) and (ii) the Kullback–Leibler information (a K–L model). The K–L model has less parameters to be prespecified than the statistical model. Under the discrete models for SRDT, the software of interest is tested with n input data sets and is accepted if the number of input data sets causing software failures in the test does not exceed a prespecified integer c. A statistical model as well as a K–L model is described for the discrete models. Neither continuous nor discrete models in the above take the damage size of software failures into consideration. The third type of models are continuous and discrete models which consider the cumulative damage size caused by software failures as well as the number of software failures in the test.

The software for a failure diagnosis system can be represented in terms of production rules; the condition part represents a system failure condition and the conclusion part corresponds to a cause of the system failure or an appropriate protective action to be taken. This paper proposes a safety assessment model of the software to evaluate its contribution to the risk caused by the entire failure diagnosis system. The proposed risk criterion considers not only the reliability of hardware components of the failure diagnosis system, but also the reliability characteristics of the system to be monitored. Conventional verification and validation methods of rule-based systems assume that the software reliability can be achieved by maintaining the consistent relation between condition parts and conclusion parts. However, the risk criterion derived in this paper shows that the software for a failure diagnosis system cannot be optimized without considering these environmental factors.

This paper considers the stochastic model of a file recovery mechanism: archive copies with n available generations are created periodically at constant time T. When failures have been detected, the consistent state is restored at the time just before fault occurrence by archive copies. The recovery overhead and its mean time are obtained when the fault latency time has a general distribution and a failure is detected uniformly during the copy interval. An optimal interval T* of archive copies is discussed analytically. Numerical examples are given when the fault latency time has exponential and gamma distributions.