The design of periodically self restoring redundant systems

Abstract

Most existing fault tolerant systems employ some form of dynamic redundancy and can be considered to be incident driven. Their recovery mechanisms are triggered by the detection of a fault. This dissertation investigates an alternative approach to fault tolerant design where the redundant system restores itself periodically to correct errors before they build up to the point of system failure. It is shown that periodically self restoring systems can be designed to be tolerant of both transient (intermittent) and permanent hardware faults. Further, the reliability of such designs is not compromised by fault latency.

The periodically self restoring redundant (PSRR) systems presented in this dissertation employ, in general, N computing units (CU's) operating redundantly in synchronization. The CU's communicate with each other periodically to restore units that may have failed due to transient faults. This restoration is initiated by an interrupt from an external (fault tolerant) clocking circuit. A reliability model for such systems is developed in terms of the number of CU's in the system, their failure rates and the frequency of system restoration. Both transient and permanent faults are considered. The model allows the estimation of system reliability and mean time to failure. A restoration algorithm for implementing the periodic restoration process in PSRR systems is also presented. Finally a design procedure is described that can be used for designing PSRR systems to meet desired reliability specifications.