Low-cost Hardware Fault Detection and Diagnosis for Multicore Systems

Continued technology scaling is resulting in systems with billions of devices. Consequently, these devices are are prone to failures from various sources resulting in a growing reliability threat. As this reliability problem is expected to affect the broad computing market, traditional solutions involving high redundancy, or piecemeal solutions targeting specific failure modes will no longer be viable. Recently, researchers have proposed SWAT, a low-cost solution that handles hardware faults by treating the resulting software anomalies. SWAT uses near-zero cost always-on symptom monitors for error detection and incurs higher diagnosis cost in the relatively rare event of a fault. The SWAT approach has been shown to be highly effective for single-threaded applications. It is, however, unclear whether such an approach would remain effective for the widely used multithreaded software on multicore systems. This paper presents mSWAT – SWAT for multicore architectures running multithreaded software, focusing on detection and diagnosis. For detection, we use the symptom-based detectors in SWAT and show that symptom detection results in a very low Silent Data Corruption (SDC) rates for both permanent and transient hardware faults. The challenge in this new setting is in fault diagnosis, owing to the fault propagating out of the faulty core, and causing symptoms from fault-free cores. To this end, we propose a novel permanent fault diagnosis mechanism that identifies the faulty core even under such circumstances. Our diagnosis procedure (1) captures the trace of the in-situ execution that activates the underlying fault, (2) deterministically replays the execution on different cores, and (3) compares the execution traces to isolate the faulty core. Our results show that this diagnosis technique, successfully diagnoses 95% of the detected faults while incurring minimal hardware overheads.