Towards Multi-Bit Errors’ Detection by Analyzing Variable Dependency

Abstract

Scaling effects make embedded systems increasingly vulnerable to reliability. Asscaling progresses, soft and hard errors in deep sub-micron ICs and memory system willbe increased. However, conventional error protection techniques can neither detect norcorrect large-scale multi-bit errors without incurring large performance, area, and poweroverheads. Program analysis to check program execution flow is an effective way todetect single and multi-bit errors. Computer program execution can be analyzed on thebasis of the dependence of the variables used in the program. This paper illustrates anovel method to analyze variable dependencies of the instructions used in a programbased on automated generation of dependence graph. Dependence graph depicts theconnectivity among the program variables where variables work as the vertices anddependence between variables perform as edges. The automatic approach of generatingdependence graph also discovers the critical variables of a program and these criticalvariables are used to detect multi-bit errors in an efficient way. Experimental studiesshow that the proposed method outperforms state of the art works in terms of space andtime overheads.