Characterization on the Phase Behavior of Microarchitecture Soft Error Vulnerability

Recent studies on soft errors have focused on dynamic phase-based reliability management, with which system resources can be tuned at runtime according to different reliability characteristics of structures. Architectural vulnerability factor (AVF) is one of the most commonly used estimation metrics of a structure's vulnerability. An effective and fast AVF analysis framework is developed and the AVF values for several key microarchitectures are evaluated. Furthermore, we used basic block profile and performance metrics information are used to characterize the reliability-based phase behavior of several key microarchitectures, and then k-means clustering and regression tree algorithms are applied for reliability-based phase classification. Experimental results show that the combination of performance metrics and regression tree method is a good candidate for soft error vulnerability phase identification.