Modeling recovery time distributions in ultrareliablefault-tolerant systems

摘要

A technique for fitting distributions to empirical recovery timedata that focuses on the components that dominate system reliability isproposed. The technique uses Goldfarb's conjugate gradient descentsearch to minimize the L2 norm of the errorprojected in the Laplace transform domain. A new parametric family ofdistributions is also suggested and is seen to provide uniformly betterpredictions of system reliability than the standard distributions usedfor this purpose, i.e. gamma, Weibull, and log normal. Applications toseveral sets of real recovery time data are provided