Abstract
Error injection is a well accepted method to evaluate hardware error detection mechanisms. An error detection mechanism is effective if it considerably reduces the amount of silently corrupted output of protected applications compared to unprotected applications. For a good representativeness of the error injection, the error model used has to mirror real world errors as accurately as possible. We introduce Error Injection Slicing (EIS) which emulates the symptoms of hardware errors. Furthermore, EIS provides means to debug single injection runs using slicing. With EIS we make the following novel contributions: (1) easy usage through hardware independence, (2) a symptom-based, flexible and comprehensive error model (e.g., not only bit-flips), and (3) debugging support to improve the detection coverage of the evaluated error detection mechanism. We evaluated the usefulness of the injector by analyzing the AN-encoding compiler that applies an AN-code to applications to facilitate hardware error detection.