Accurate Model for Application Failure due to Transient Faults in Cache

To select an appropriate level of error protection in caches, the impact of various protection schemes on the cache FIT (Failure In Time) rate must be evaluated for a target benchmark suite. However, while many simulation tools exist to evaluate area, power and performance for a set of benchmark programs, there is a dearth of such tools for reliability. This technical report introduces an accurate and versatile cache reliability model called PARMA+ which estimates a cache’s FIT rate in the presence of spatial multi-bit transient faults and for a given benchmark program. PARMA+ is an extension of PARMA [21]. PARMA only dealt with single-bit faults. However, PARMA+ is applicable to single-bit and any spatial multi-bit fault pattern distribution and various error protection schemes for caches including early write-back, bitinterleaving and scrubbing. We first develop the model formally, then we demonstrate its accuracy. We have run reliability simulations with PARMA+ for 13 SPEC2000 benchmark programs and for several distributions of large and small fault patterns and have compared them with accelerated fault injection simulations. PARMA+ has high accuracy and low computational complexity.