부산대학교 도서관

The technological scaling has made electronic devices more susceptible to radiation-induced faults. To mitigate these faults, researchers have proposed different techniques at the software, hardware, and hybrid levels. Although some of these techniques are very effective, most of them are based on redundancy, which causes non-negligible computational overheads in terms of area, performance, and power consumption. In this paper, we propose FTxAC, which consists in using approximate computing techniques in conjunction with radiation-induced mitigation strategies for the design of fault-tolerant systems with minimal overheads. Given the nature of the approximate computing paradigm, FTxAC is suitable for error resilient applications. To show the applicability of the proposal, we carry out several case studies with the TI MSP430 microcontroller, in which a set of software-based fault tolerance techniques are used jointly with approximate computing methods to reduce overheads. The experimental results are analyzed in terms of fault coverage, accuracy of the results, and overheads of the several levels of redundancy. Results show that, depending on the level of approximation, the evaluated application, and the fault tolerance strategy employed, the performance of the system can be improved, even counteracting completely the implicit overheads of the redundancy.