Fault-tolerant systems design-estimating cache contents and usage

Development of the Remote Exploration and Experimentation (REE) Commercial Off The Shelf (COTS) based space-borne supercomputer requires a detailed knowledge of system behavior in the presence of Single Even Upset (SEU) induced faults. When combined with a hardware radiation fault, model and mission environment data in a medium grained system model, experimentally obtained fault behavior data can be used to: predict system reliability, availability and performance; determine optimal fault detection methods and boundaries; and define high Return On Investment (ROI) fault tolerance strategies. The REE project has developed a fault injection suite of tools and a methodology for experimentally determining system behavior statistics in the presence of SEU induced transient faults in application level codes. Where faults cannot be directly injected, analytic means are used in conjunction with experimental data to determine probabilistic system fault response. In many processors, it is not possible to inject faults directly into onboard cache. In this case, a cache contents estimation tool can be used to define probabilistic fault susceptibility which is then combined with direct memory fault injection data to determined fault behavior statistics. In this paper we discuss the structure, function and usage of a PPC-750 cache contents estimator for the REE project.