| Abstract | It is well-known that the dependability that can be achieved by a fault-tolerant systemis particularly sensitive to both the asymptotic value of coverage and to the time distribution of
coverage. However, most previous work on coverage evaluation by statistical processing of the
results of fault-injection experiments has only been concerned with estimating asymptotic
coverage. In this paper, we tackle the problem of estimating the parameters of models that also
account for coverage latency. After discussing some data sets resulting from fault-injection
experiments on practical systems, we propose a series of coverage latency models that might be
considered to account for the observed phenomena in a system dependability evaluation. We
consider both exponential and non-exponential models, and assess their pertinence by means of a
sensitivity study. We confirm previous results that latency can have an extremely important effect
on the achievable dependability. We also show that the shape of the latency distribution has only a
minor impact in the practical case of systems with high asymptotic coverage. A simple action
model based on an exponential latency distribution is therefore proposed. We show how worst-case
confidence limits can be obtained for the parameters of this action model and study the effects of
data truncation that are unavoidable in any practical measurements of latency. We conclude by a
critical assessment of the proposed estimation technique and a demonstration of its application to
practical data sets.
|
