CPU-aware, process-level redundancy to tolerate faults in multi-core

This paper proposes: 1) A dynamically scheduled Process-Level Redundancy (PLR) for enhancing reliability of multi-core systems, 2) A comparison between PLR and Thread-Level Redundancy (TLR), and 3) A fault study on the thread selector unit of a modern processor. The proposed technique employs underutilized CPU resources to improve fault tolerance ability of a system. The evaluation on PLR reliability proves that it performs better than Thread-Level Redundancy (TLR) when the reliability of sub modules in a system is higher than almost 0.8. In this technique, a set of redundant processes are created per application process. The number of replicas is then modified dynamically to achieve better performance. The experimental results on some standard benchmarks show that on average, the CPU is utilized less than 20% during the execution time of applications which can be used to provide 100% fault detection and recovery with almost 10% performance overhead using the proposed technique. Also, the fault study proves that among 7000 faults injected into the thread selector module using OpenSPARC simulator, 83.5% of faults are benign faults, and 16.5% of faults lead to system failure which affect either hardware (13.7%), or program outputs (2.8%). These faults can be all detected using this technique.