Program fault tolerance based on memory access behavior

Abstract

Fault observability based on the behavior of the memory references is studied. As opposed to traditional studies that view memory as one large entity that must completely work to be considered reliable, this study emphasizes the usage patterns of a particular program's memory. Expressions for the successful execution of a program that take into account the usage of the data are developed. Three variations that depend on whether the program's storage is pre-allocated, dynamically allocated, or constrained in allocation are presented. A theory is proposed to explain the phenomenon that increased workloads lead to increased failure rates, which has been observed in several studies. The model is used to study several program traces, and is shown that increased workloads could cause an increase of the observed failure rates in the range of 27% to 53%.