Computing Accurate AVFs using ACE Analysis on Performance Models: A Rebuttal

Abstract

ACE (architecturally correct execution) analysis computes AVFs (architectural vulnerability factors) of hardware structures. AVF expresses the fraction of radiation-induced transient faults that result in user-visible errors. Architects usually perform this analysis on a high-level performance model to quickly compute per-structure AVFs. If, however, low-level details of a microarchitecture are not modeled appropriately, then their effects may not be reflected in the per-structure AVFs. In this paper we refute Wang, et al.'s (2007) claim that this detail is difficult to model and imposes a practical threshold on ACE analysis that forces its estimates to have a high error margin. We show that carefully choosing a small amount of additional detail can result in a much tighter AVF bound than Wang, et al. were able to achieve in their refined ACE analysis. Even the inclusion of small details, such as read/write pointers and appropriate inter-structure dependencies, can increase the accuracy of the AVF computation by 40% or more. We argue that this is no different than modeling the IPC (instructions per cycle) of a microprocessor pipeline. A less detailed performance model will provide less accurate IPCs. AVFs are no different.