Estimating the Soft Error Vulnerability of Register Files via Interprocedural Data Flow Analysis

Abstract

Subsequently to the wall of performance and power consumption, the dependability of computing, caused by soft errors, has become a growing design concern. Since Register Files (RFs) are accessed very frequently and cannot be well protected, soft errors occurred in them is one of the top reasons for affecting the reliability of programs. To access the soft errors vulnerability of RFs, this paper presents a static estimating method via interprocedural data flow analysis. Adopting a previous method, the vulnerability of a register is firstly decomposed into intrinsic and conditional basic block vulnerabilities. Under the prerequisite of context sensitivity, we focus on the computation the post conditions of basic blocks, which can be viewed as the living probability of the target register in the future usage. Finally, the program reliability can be calculated quantitatively under the occurrence of soft errors in RFs. Experimental results from the MiBench benchmarks indicate that our method is more accurate, and compatible with the AVF methods. We also reveal that the reliability of a program has a connection with its structure, such as the RVF factors, which suggests adopting the application specified protected mechanisms for tolerating soft errors occurred in RFs.