A watchdog processor based general rollback technique with multiple retries

Abstract

A common assumption in the existing rollback techniques is that transients, the cause of most failures, subside very quickly, implying that a single story retry of the program from the previous rollback point is sufficient. The authors discuss a general rollback strategy with n(n≥2) retries which takes into consideration multiple transient failures as well as transients of long duration. Ways of deriving practical values of n for a given program are also discussed. Furthermore, the authors propose the use of a watchdog processor as an error detection tool to initiate recovery action through rollback, since the watchdog processor offers low error latency. They also discuss the merging of the watchdog processor with rollback recovery technique for enhancing the overall system reliability.