In this paper the behavior of assertion-based error detection mechanisms is characterized under faults injected according to a quite general fault model. Assertions based on the knowledge of the application can be very effective at detecting corruption of critical data caused by hardware faults. The main drawbacks of that approach are identified as being the lack of protection of data outside the section covered by assertions, namely during input and output, and the possible incorrect execution of the assertions. To handle those weak-points the Robust Assertions technique is proposed, whose effectiveness is shown by extensive fault injection experiments. With this technique a system follows a new failure model, that is called Fail-Bounded, where with high probability all results produced are either correct or, if wrong, they are within a certain bound of the correct value, whose exact distance depends on the output assertions used. Any kind of assertions can be considered, from simple likelihood tests to high coverage assertions such as those used in the Algorithm Based Fault Tolerance paradigm. We claim that this failure model is very useful to describe the behavior of many low-cost fault-tolerant systems, that have low hardware and software redundancy, like embedded systems, were cost is a severe restriction, yet full availability is expected.
展开▼
