Fault-tolerant wait-free shared objects

Abstract

A concurrent system consists of processes and shared objects. Previous research focused on the problem of tolerating process failure. We study the complementary problem of tolerating failures. We divide object failures into two broad classes: responsive and non-responsive. With responsive failures, a faulty object responds to every invocation, but responses may be incorrect. With non-responsive failures, a faulty object may also 'hang' without responding. For each class, we consider crash, and arbitrary types of failures. For each type of failure, we are seeking a universal implementation for fault-tolerant wait-free shared objects. We present (deterministic) implementations for all types of responsive failures, including arbitrary failures. In contrast, we show that even the most benign type of non-responsive failures requires the use of randomization. Of special interest is the problem of implementing fault-tolerant objects using only objects of the same type. We present such fault-tolerant self-implementations for many common object types. Graceful degradation is a desirable property of fault-tolerant implementations: the implemented object never fails more severely than the base objects it is derived from, even if all the base objects fail. For several failure models, we show whether this property can be achieved, and, if so, how. In addition to the above possibility/impossibility results, we also consider the resources complexity of fault-tolerant implementations. In many cases, we present lower bounds and give matching algorithms.