Characterizing Time-Varying Behavior and Predictability of Cache AVF

With the development of information systems, electronic devices are becoming more and more susceptible to soft errors, especially for the tough environment of drastic electromagnetic interference. Architectural Vulnerability Factor (AVF), which is defined as the fraction of soft errors that result in erroneous outputs, has been introduced to quantify the vulnerability of structures to soft errors. Recent studies have shown that the AVF of several micro-architectures (e.g. issue queue) exhibits significant runtime variations and certain predictability for SPEC2K benchmarks, thus motivating the development of AVF-aware fault tolerant techniques. Through accurate AVF prediction, these techniques provide error protection only at the execution points of high AVF rather than the whole execution lifetime of programs, thereby reducing the overheads of soft error protection schemes without sacrificing much reliability. The native motivation of this paper is to see if cache AVF also exhibits such predictability for the further exploration of AVF-aware cache protection techniques. In this paper, we characterize dynamic vulnerability behavior of level1 data (L1D) cache and analyze the correlation between L1D AVF and performance metrics. Based on the analysis of variance and predictability of L1D AVF, we propose a novel hierarchical classification of SPEC2K benchmarks to provide insights into the varying vulnerability behavior and predictability of cache AVF. We develop a new methodology to select the best-suited predictive method for cache AVF. Our results show that accurate cache AVF predictor is available for SPEC2K benchmarks, and most programs are excellent candidates for AVF-aware fault tolerant schemes.