A recovery model for extended real-time transactions

Abstract

A central problem in the design of fault-tolerant real-time systems is that desirable fault tolerance properties are usually realized by mechanisms that counteract real-time guarantees. A prominent example is the all-or-nothing property (also known as failure atomicity), known from transactions. This property normally is realized by the means of isolation and roll-back recovery. However, isolation often unnecessarily decreases the degree of permitted concurrency and results in missed deadlines. Roll-back recovery implies that work is lost and has to be re-done, again leading to missed deadlines. So-called extended transaction models supersede isolation, but their recovery model induces an increased amount of roll-back recovery. In this paper, we present a fundamentally new recovery model to provide the all-or-nothing property without requiring isolation. Based on an active replication technique, we provide a forward error recovery that avoids unnecessary roll-backs by replacing cancelled primary computations with hot-stand-by alternate computations.