CacheFI:基于架构级故障注入的片上缓存容错评估工具

体系架构级缓存容错技术被认为是应对较高的永久位故障率的有效手段，但目前缓存容错机制的体系架构级评估工具较少。针对这个问题，提出CacheFI，即基于Simics的缓存故障注入工具，采用故障生成和注入分离的设计，故障生成是随机分布、模式和时序三个方面的结合，故障注入则考虑了故障可重现性和模块化的需要。在全系统模拟器Simics上，基于15个选自SPEC CPU2000的测试程序，利用CacheFI对Buddy和MAEP等典型的体系架构级缓存容错机制进行评估，展现了其弱点和典型的片上缓存容错机制存在的问题。
Architectural solutions of on-chip Cache fault tolerance are considered as effective means for high persistent failure probabilities. However, less fault injection tools for on-chip Caches are available. Therefore, the CacheFI, a fault injection tool based on the full-system simulator Simics was proposed. A separate mechanism of the fault generation and injection was employed. Fault generation consists of stochastic distribution control, failure patterns and explosive timing. Fault injection was designed to focus on the requirement of repeatability and modularity. According to the experiments based on Simics and 15 benchmarks from SPEC CPU2000, it evaluates typical micro-architectural fault mechanisms, such as Buddy, MAEP(matching access and error pattern) etc. by injecting Cache faults with CacheFI. Consequently, it presents the weakness and issues of these typical mechanisms.