Benchmarking of fault-tolerant systems

This thesis presents a benchmark for evaluating fault tolerance. The benchmark is based on the FTAPE tool, which injects CPU, memory, and disk faults and generates workloads with specifiable amounts of CPU, memory, and disk activity. Two benchmark metrics are produced: (1) a count of the number of catastrophic incidents and (2) the average performance degradation. The catastrophic incident count represents the recovery coverage of the system, while the performance degradation reflects the performance of the system in the presence of faults.The benchmark is fully functional and has been implemented on three Tandem fault-tolerant machines (Prototypes A, B, and C). The benchmarks results show that Prototypes B and C are more fault-tolerant than Prototype A, in that they suffer fewer catastrophic incidents under the same workload conditions and fault injection method. Also, Prototype C suffers less performance degradation in the presence of faults, which might be an important concern for time-critical applications.Fault injection plays an important part in the benchmark because it is the means by which fault-tolerant activity is generated. To ensure a high level of fault activation and error propagation, focused fault injection strategies are used. Two such strategies are presented in this thesis: stress-based injection and path-based injection.U of I OnlyETDs are only available to UIUC Users without author permissio