Analysis for performance and reliability of fault tolerant parallel software

The authors have proposed a method converting parallel programs to construct fault-tolerant parallel software (FTPS) for massively parallel processes without any dedicated facilities. This method comes from a hybrid approach of the primary-backup and the state-machine approaches, and provides an automatic conversion. Introducing fault-tolerance to a parallel system, you should take a trade-off between reliability and performance for granted since some components are saved as redundancy for possible failures. Moreover, fault-tolerance with software also requires adding overheads to original software, which may reduce its performance. Therefore, mere reliability without concern about performance does not reflect improvement of FTPS. This paper proposes a criterion for FTPS, mean workload to failure, to reflect performance as well as reliability, and considers conditions for practical configuration of FTPS systems.