Modeling the Propagation of Soft Errors in Programs

Due to the continuous growth of transistors in processors, the cloud computing paradigm and increasingly complex environment, soft errors, which are a category of typical transient errors, have become an urgent challenge in ground-level systems. To handle these errors, making clear the error propagation is the key step. As programs become large, we cannot employ traditional fault injection campaigns to monitor every possible error. This paper proposes a method to study and model the propagation of soft errors within a program. Based on dynamic instructions traced in an error-free program execution, the ACE analysis classifies soft errors in architectural registers into benign ones and non-benign ones. The benign errors are considered to be derated in the propagation. Then we build a crash model and an improved DDG to analyze the propagation of each non-benign error and to predict its consequence (crash or silent data corruption). If the error is considered to cause a crash, the crash latency and the propagation path are also predicted. The method can be used to predict outcomes of programs under soft errors as well as occurrence of correct outputs, silent data corruptions or crashes. Extensive fault-injection experiments are provided to validate the proposed method from multiple perspectives.