Fault Tolerance based on Time-Staggered Redundancy

Abstract

A new class of fault tolerance techniques is introduced: Time-staggered redundancy is a modification of static redundancy (replication of processes and fault masking). Some of the replicas are executed in parallel, others with an adjustable delay. The latter contribute to n-out-of-m majority voting as usual, and to backward error recovery, too. The delayed processes represent former state information of the process system and therefore can be taken as a recovery point. Staggered execution of process copies enables the concepts of static and dynamic redundancy at a time — without additional checkpointing overhead. As comparison tests and acceptance tests can be applied both, a higher degree of fault tolerance is achieved. Moreover, testing the results of the early processes detects when wrong input data have been processed. In this case improved input data are requested for the late processes. Finally correct output data are chosen among the results of all processes (early and late ones). Time-staggered redundancy should be preferred if multiple faults of different types have to be tolerated, and if time redundancy is limited, but sufficient for delayed process execution. In contrast to periodic or event-driven checkpointing, available time redundancy can be used completely for backward error recovery at any time: The late processes serve as “computing recovery points” with “continuous checkpointing”.