Fault Injection Analysis of Automotive Benchmarks using HPC-Based Monitoring

The downscaling of electronic components in terms of voltage and physical dimensions has enabled the development of modern high-performance microprocessors, which are increasingly adopted in safety-critical applications. However, their reduced dimensions have increased the sensitivity to radiation. Radiation-induced soft errors have become a key threat in terms of reliability in safety-critical real-time embedded systems (SACRES). A common technique used to enhance the hardware reliability is N-Modular Redundancy (NMR). With the growing complexity of modern microprocessors, this solution has become unaffordable. The development of lighter alternatives for implementing hardware/software error detection mechanisms has become a crucial area of research. This thesis investigates the effects of fault injections on automotive benchmarks to assess the reliability of the target setup, a crucial step for studying new, lighter robustness solutions. The selected benchmarks are taken from the EEMBC MultiBench™ Multicore Benchmark Suite to strengthen the representativeness of the experimental environment. The strength of this approach lies in the fact that MultiBench combines a wide variety of application-specific workloads with the EEMBC Multi-Instance Test Harness (MITH), which is compatible and portable across most multicore processors and operating systems. The benchmarks are executed in a real-time scenario under FreeRTOS, exploiting its scheduling algorithm to enable task concurrency, on the Xilinx Pynq-Z2 board. With this setup, benchmarks are scheduled concurrently and executed across multiple runs of the system, resulting in different demos. Faults are injected into both memory and registers via host, and their effects are analyzed by leveraging Hardware Performance Counters (HPCs) to monitor micro-architectural events. To maximize the information gathered, each fault is executed multiple times so that its impact can be observed across different monitored events.