Investigating parallel TMR approaches and thread disposability in Linux

This paper investigates the impact of parallelization and redundancy at thread level in four TMR software implementations and presents a novel use for TMR on applications running on top of Linux OS. In this work, TMR is used not only as protection against silent data corruptions but also hangs. This protection is done by discarding the threads that presented hangs or terminated unexpectedly. The application tolerance to soft errors is evaluated by employing a fault injection module extension based on the OVP simulator (OVPsim-FIM). Results show that ordinary parallel TMRs present more errors than the sequential one. Nevertheless, employing the thread disposability technique, parallel TMR techniques achieve better results than the full sequential one, making their use suitable.