Simulation of software behavior under hardware faults

In this paper, we introduce a simulation-based software model that permits application specific dependability analysis in the early design stages. The model represents an application program by decomposing it into a graph model consisting of a set of nodes, a set of edges that probabilistically determine the flow from node to node, and a mapping of the nodes to memory. The software model simulates the execution of the program while errors are injected into the program's memory space. The model provides an application dependent parameters such as detection and propagation times and permits evaluation of function and system level error detection and recovery schemes. The paper illustrates, via a case study, the interaction between an application program and two detection schemes. Specifically, Gaussian elimination programs running on a Tandem Integrity S2 system with memory scrubbing are studied. Results obtained from the simulation-based software model are validated with data measured from an actual Tandem Integrity S2 system. Application dependent coverage values obtained with the model are compared with those obtained via traditional schemes that assume uniform or ramp memory access patterns. For our program, some coverage values obtained with the traditional approaches were found to be 100% larger than those obtained with the software model.