Efficient Recovery of Statically Redundant Systems

Abstract

The reliability of modular (n-out-of-m)-systems can not only be improved by an increase in the number of replicates, but also by the provision of additional facilities for backward recovery. The latter approach combines static and dynamic redundancy, resulting in two main advantages: Firstly, more fault types are tolerated, even particular propagating faults which cause loss of the majority. Secondly, the expense of joint application of voting and backward recovery can be significantly reduced to be lower than the sum of expenses of separate implementation of the two techniques.

In this paper, we present a concept of an efficient recoverable (n-out-of-m)-system. Its efficiency is achieved by novel solutions in the particular design space induced by the combination of different techniques: • The establishment of recovery points is shifted into idle phases of any of the m redundant processors as far as possible. • The necessary synchronization is implemented by a special agreement protocol which only consists of the messages needed for fault masking anyway, provided that the communication system enables multicast (no reliable multicast required!). • Both relative and absolute tests contribute to fault detection. • Depending on the actual fault situation, the state information for recovery can be obtained from the established recovery points or a faultless replicate, if there is any. Moreover, the replicates may be rolled back independently or in common. These choices enable compromises between fault encapsulation and performance.

A brief evaluation turns out that our concept of recovery in (n-out-of-m)-systems can be more efficient than an increased number of replicates, because during normal operation the overhead of establishing recovery points can be reduced, and in the presence of faults additional fault types can be tolerated.