Evaluating the Inherent Sensitivity of Programming Languages to Soft Errors

Abstract

The choice of programming language depends on a number of factors, such as efficiency and maintainability. With the predicted increase of the soft error rate in coming years due to technological improvements and the popularization of energy-saving techniques, the responsibility of tolerating transient hardware faults is no longer solely on the hardware-side, but is becoming more and more a software problem. Therefore choices taken throughout the software development and compilation process, such as the used programming language, are now a factor that should be taken into account. This paper evaluates the impact of the programming language in the sensitivity and vulnerability to soft errors of an application, through extensive fault injection campaigns that cover 5 benchmarks over 7 programming languages, considering the most common implementation. The results show that both the programming language and workload have a significant correlation with the occurrence of failures due to soft errors. Furthermore, since execution time varied noticeably across programming languages, applications developed in less-efficient languages showed a higher vulnerability time window in which they may be affected by soft errors.